# Changelog
Source: https://docs.abundly.ai/changelog/index

New features, improvements, and fixes in Abundly

New releases and improvements to the Abundly platform.

<Update label="August 7, 2026" description="Claude Fable 5, enterprise plan display fix, What's New per-user persistence, dark mode status badge fix">
  ## ⚠️ Changed Behavior

  * **For workspace admins: Claude Fable 5 is now on by default** — Fable 5 previously had to be turned on explicitly before workspaces could use it. It now follows the same rules as every other model in the catalog, so it appears for any workspace that hasn't disabled it. [Choosing a model](https://docs.abundly.ai/features/model-selection)

  ## New Models

  * **Claude Fable 5 available in the model picker** — Anthropic's most capable model can now be selected for any agent, without needing it enabled first. It's the top of the Claude range, above Opus, for the hardest problems. [Choosing a model](https://docs.abundly.ai/features/model-selection)
  * **New "Claude Fable Best" alias** — Pick this instead of a specific Fable version and your agent automatically follows the latest version we recommend, the same way the Sonnet, Opus, and Haiku aliases work. [Choosing a model](https://docs.abundly.ai/features/model-selection)

  ## Improvements

  * **For workspace admins: Enterprise plans no longer listed twice** — Workspaces on an enterprise plan now see only the plan they are subscribed to on the billing page, instead of one entry per underlying price.
  * **"What's New" dismissal now syncs across devices** — Dismissing or reading the "What's New" popup on one device now keeps it dismissed everywhere you're signed in, instead of reappearing per device.

  ## Bug Fixes

  * **Service status badge readable in dark mode** — The status badge in the Help & support menu now follows your light or dark theme.
</Update>

<Update label="August 6, 2026" description="GitHub Enterprise, bulk agent management actions">
  ## GitHub Enterprise Support

  The GitHub capability now works with GitHub Enterprise — both Enterprise Cloud with data residency and self-hosted Enterprise Server. Workspace admins can enter the instance's host name in the new **GitHub host** field when configuring the GitHub capability; leave it empty to keep using github.com.

  Agents in GitHub Enterprise workspaces can browse repositories, review pull requests, and manage issues on their company's GitHub instance just like they already could on github.com. [GitHub integration](https://docs.abundly.ai/integrations/github)

  ## Bulk Agent Management Actions

  Two new bulk actions in the Agent Management view make it faster to work with many agents at once:

  * **Trigger** — Select any number of agents, enter a message, and fire all of them at once. Useful for broadcasting a task or message to multiple agents simultaneously.
  * **Send notice** — Workspace admins can select multiple agents and send an owner notice to all of their owners in one action, without visiting each agent individually.

  ## Improvements

  * **Removed misleading model badge** — The model badge in the agent selector and management table has been removed. It showed a single model but agents use different models per purpose (chat, scheduled tasks, email, etc.), making it actively misleading.
</Update>

<Update label="August 5, 2026" description="Agent owner notices, model picker improvements">
  ## Agent Owner Notices

  Agents can now send notices to their owners when they are blocked, need a decision, or want to propose a change they cannot make themselves — for example, a token that needs refreshing, a question only a human can answer, or suggested rewording for their own instructions.

  The new always-available `notify_owners` tool takes a `title`, a `message` (up to 8 000 characters), and an optional `severity` (`info`, `warning`, or `error`). Each notice is delivered to the agent's owners through the notification system, escalating to team and workspace admins for ownerless agents.

  Owners can view all notices in full on a new **Owner Notices** tab on the agent's Activity page. Each notice includes a button that opens a chat seeded with the notice content, so the owner can act on it immediately.

  ## Model Picker Improvements

  * **Distinct chat-default option** — Chats that follow the agent's model default now show a distinct *(chat default) Model Name* entry in the picker. Selecting the plain model name pins the chat to that model explicitly — fixing a case where this was impossible on mobile.
  * **Alias target tooltip** — Hovering the picker button for an alias selection now shows which model the alias currently points to, so the moving-pointer nature of the alias is visible at a glance.

  ## Bug Fixes

  * **Model picker on new-chat page** — The picker was showing the alias's resolved target instead of the alias itself before the first message was sent. It now shows the alias consistently from the start.
</Update>

<Update label="August 4, 2026" description="Per-chat model selection, multi-account Gmail">
  ## Per-Chat Model Selection

  Every new portal chat now keeps the model it was started with. If you change the agent's default model later, existing conversations continue with the model they were created under — no surprises.

  A model picker in the chat top bar (desktop and mobile) lets you switch the model for that chat only.

  ## Multi-Account Gmail

  One agent can now be connected to multiple Gmail accounts — useful for assistants managing both a personal and a work inbox, for example. When two or more accounts are linked, Gmail tools gain a required `account` parameter to specify which inbox to use; single-account setups are unchanged.

  ## Improvements

  * **Platform-default alias stamping** — Agents with no model configured now track the platform-default alias itself, so they follow platform retargets automatically. The settings page also makes it clearer when a model is inherited rather than explicitly set.
  * **Mobile model selector** — The chat model selector on mobile is now functional and correctly styled.
</Update>

<Update label="July 30, 2026" description="Gemini 3.6 Flash, Gemini 3.5 Flash Lite, value dashboard link fix">
  ## New Models

  * **Gemini 3.6 Flash** (`gemini-3.6-flash`) — Google's new default workhorse model is now available. Same input price as Gemini 3.5 Flash ($1.50/M tokens) with cheaper output ($7.50/M vs \$9/M), placing it in cost tier 2. 1M context window with thinking support.
  * **Gemini 3.5 Flash Lite** (`gemini-3.5-flash-lite`) — The fastest and most affordable model in the 3.5 generation ($0.30/$2.50 per 1M tokens, cost tier 2). 1M context window with thinking support.

  ## Bug Fixes

  * **Value dashboard "Learn more" link fixed** — The link in the value dashboard now correctly points to the docs page (previously led to a 404).
</Update>

<Update label="July 28, 2026" description="Unified delete confirmation dialog">
  ## Improvements

  * **Consistent delete confirmation** — The delete confirmation dialog is now identical whether you delete from the document explorer or from the open document side panel. Previously the two surfaces showed different-looking dialogs for the same action.
</Update>

<Update label="July 28, 2026" description="Folder deletion safeguard, LLM credential fix, model picker improvements">
  ## Folder Deletion Safeguard

  Deleting a document folder that contains documents or subfolders now requires explicit confirmation. The confirmation dialog shows what will be deleted, helping prevent accidental data loss.

  ## Bug Fixes

  * **LLM API keys no longer leak across models** — A custom LLM configuration's API key was incorrectly applied to all models from the same vendor. The key now applies only to its specific model, and the settings UI no longer auto-copies it to the shared vendor key on page load.
  * **Custom models now visible in the model picker** — Catalog models pinned to a specific API key via a per-model override now appear correctly in the model picker.

  ## Improvements

  * **LLM settings header shows the editing scope** — The LLM section header in the configuration panel now reads "LLMs (global)", "LLMs (workspace)", or "LLMs (agent)" to make it clear which level you are editing.
</Update>

<Update label="July 27, 2026" description="Claude Opus 5, Grok 4.5, thinking toggle fixes">
  ## New Models

  * **Claude Opus 5** (`claude-opus-5`) — Anthropic's latest frontier model is now available on the platform. Like Opus 4.8, it uses adaptive thinking with the `effort` parameter.
  * **Grok 4.5** (`grok-4.5`) — xAI's Grok 4.5 is now available as a selectable model (standard context: $2/M input, $6/M output).

  ## Improvements

  * **Thinking toggle now correct for GPT mini/nano** — `gpt-5-mini`, `gpt-5-nano`, `gpt-5.4-mini`, and `gpt-5.4-nano` now correctly show the thinking toggle as available in the model picker.
  * **Inherited model thinking gating** — When no model is explicitly selected, the thinking toggle now correctly reflects whether the inherited/default model supports thinking, matching behavior for explicitly-selected models.
</Update>

<Update label="July 24, 2026" description="Value dashboard, per-interaction actual value, Gemini image model fix">
  ## Value Dashboard

  A new **Value dashboard** is now available in the Management section of the portal. It gives creators a cross-agent overview of triggers & messages, credits used, and actual vs. expected value per agent — with card and table views, trend sparklines, and links to value capture settings for each agent.

  The dashboard is visible when workspace value capture is enabled.

  ## Per-interaction Actual Value

  Agents can now be configured with a **per-interaction rate** for value capture. Instead of entering a manual value, set an amount per trigger or message and the dashboard will compute actual value automatically from usage counts. New value capture entries default to per-interaction mode.

  ## Bug Fixes

  * **Gemini image model fix** — The `gemini-3.1-flash-image` model alias has been updated (the previous `-preview` suffix caused the model to not be found).
</Update>

<Update label="July 23, 2026" description="create_agent team assignment fix, workspace default model fix">
  ## Bug Fixes

  * **`create_agent` tool: agents now assigned to creator's team** — Agents created via the `create_agent` tool are now automatically placed in the same team as the creating agent. Previously, created agents were always ungrouped regardless of which team the creator belonged to.
  * **Workspace default model now saves correctly** — Changing the default model in Workspace Settings now persists after a page reload. The setting was being applied optimistically in the UI but not saved to the backend.
</Update>

<Update label="July 21, 2026" description="Eval improvements, Claude Sonnet 5 fix">
  ## Improvements

  * **Eval editor: "Run plan and security" gated on system prompt** — The *Run plan and security* switch is now disabled (and automatically forced off) when the agent system prompt is not included in an eval (`agentContext = none`). An explanatory tooltip explains why. This prevents creating invalid eval configurations.
  * **Eval poll timeout extended to 30 minutes** — Long-running evals no longer time out after 5 minutes. The UI also shows a clearer status message if polling does time out, replacing the previous contradictory "Eval Run Complete / Polling timed out" state.

  ## Bug Fixes

  * **Claude Sonnet 5: sampling parameters fixed** — Fixed an issue where passing custom `temperature`, `top_p`, or `top_k` values caused API errors with Claude Sonnet 5. The model now uses default sampling parameters, matching Claude Opus 4.7+ behavior.
</Update>

<Update label="July 20, 2026" description="Anthropic quota reliability fix">
  ## Bug Fixes

  * **Anthropic daily/monthly quota: fail-fast instead of retrying** — Agent runs that hit Anthropic's daily or monthly token quota now fail immediately with a clear error instead of burning through \~10 retries over 30 minutes. Short-term rate limit 429s (TPM/RPM) remain transient and continue retrying as before.
</Update>

<Update label="July 10, 2026" description="GPT-5.6 models, plain-English HTTP status, Gemini Flash Lite fix">
  ## New Models: GPT-5.6 Sol, Terra, Luna

  OpenAI's three new GPT-5.6 models are now available in the model picker:

  * **GPT-5.6 Sol** — Frontier tier. The `gpt-5.6` alias points here, matching GPT-5.5 pricing.
  * **GPT-5.6 Terra** — Mid-tier, balancing intelligence and cost.
  * **GPT-5.6 Luna** — Cost-sensitive tier.

  All three support reasoning tokens and the 1.05M context window.

  ## Improvements

  * **Clearer HTTP request progress** — Agent progress messages for HTTP requests now show plain-English outcomes (`success`, `redirected`, `request was rejected`, `server error`) instead of raw status codes, making the activity log easier to read.

  ## Bug Fixes

  * **Gemini Flash Lite** — Agents configured with the deprecated `gemini-3.1-flash-lite-preview` model (shut down by Google on 2026-05-25) have been automatically migrated to the GA model ID `gemini-3.1-flash-lite`. Affected agents will now run correctly again.
</Update>

<Update label="July 9, 2026" description="Remove mic from doc/instructions editors, Stripe billing reliability fix">
  ## UI Improvements

  * **Microphone button removed from document and instructions editors** — Voice-to-text input has been removed from the document editor and agent instructions editor. The mic remains available in the chat composer and message edit form.

  ## Bug Fixes

  * **Stripe billing reliability** — Fixed an issue where duplicate Stripe subscriptions could be created, blocking workspaces with a `SubscriptionNotActive` error. The billing plan card now correctly shows "Cancels on" (instead of "Renews on") when a subscription is set to cancel at period end.
</Update>

<Update label="July 7, 2026" description="Workspace user action log, new Google Calendar tool, Fortnox file upload tool">
  ## Workspace User Action Log & Audit Logs

  Workspace admins now have a full audit trail of configuration changes in their workspace. The new **User action log** tab on the Usage page shows an append-only log of changes to agents, teams, secrets, API capabilities, and workspace settings — with CRUD-colored badges and expandable detail rows.

  Sensitive values (such as instructions and prompts) are redacted to hash tokens so admins can confirm that content changed without seeing what it was. Workspace admins and team admins see views scoped to what they have access to.

  A CSV/JSONL export is available (up to 50,000 rows).

  ## New Google Calendar Tool: Get Event Details

  The Google Calendar capability gains a new `get_google_calendar_event` tool for fetching the full details of a single event by ID: description, organizer, creator, attendee response statuses, htmlLink, and timestamps. The existing list tool returns a compact view; this tool gives you everything.

  ## New Fortnox Tool: Upload Files

  A new `fortnox_upload_file` tool lets agents upload PDFs and images directly to Fortnox Inbox or Archive without routing raw bytes through the LLM. Uploaded files can optionally be attached to vouchers or supplier invoices in the same call. Requires the `inbox`, `archive`, or `connectfile` OAuth scope depending on the operation.
</Update>

<Update label="July 6, 2026" description="Proactive context summary button, new Gmail draft tool, shareable document panel links">
  ## Proactive "New Chat with Summary" Button

  A button now appears next to the context usage indicator once your conversation reaches 80% of the context limit. Clicking it summarizes the conversation and starts a fresh chat — the same flow as the context limit dialog, but available proactively before you hit the wall.

  ## New Gmail Tool: Create Draft

  The Gmail capability gains a new tool — `create_gmail_draft` — for composing standalone new emails. Previously you could only draft replies and forwards anchored to an existing email. Now you can start a fresh email (with to/cc/bcc/subject/body), review it in Gmail, and send.

  ## Shareable Document Panel Links

  Maximizing or going fullscreen in the documents panel now syncs the state to the URL (`?view=maximized` / `?view=fullscreen`). The state survives a page refresh, and sharing the URL opens the document in the same view.

  ## Gmail Improvements

  * **Reply-To handling** — Replies now correctly use the original email's `Reply-To` header (rather than `From`) when one is present, matching standard email client behavior.
  * **Display name quoting** — Names with special characters (e.g. `Engborg, Nikolaj`) are now correctly quoted in email headers, preventing malformed headers that some servers reject.

  ## Bug Fixes

  * Fixed an error toast appearing on the admin Credit Usage page when loading workspace statistics.
</Update>

<Update label="July 3, 2026" description="Collaborator workspace role">
  ## New Role: Collaborator

  A new **Collaborator** workspace role sits between Creator and Guest in the workspace hierarchy.

  Collaborators can:

  * Chat with agents and interact with shared ones
  * Edit agents they've been given access to
  * View and select existing secrets and API capabilities when configuring agents they edit

  Collaborators cannot:

  * Create, clone, or import agents
  * Create or manage workspace secrets or API capabilities

  Role hierarchy: **Admin > Creator > Collaborator > Guest**

  The existing
</Update>

<Update label="July 2, 2026" description="SharePoint test connection fix, activity log event type improvements">
  ## Bug Fixes

  * **SharePoint app-only test connection** — Test connection for SharePoint app-only credentials now works correctly with `Sites.Selected` permissions. Previously it always returned 401 because it probed an endpoint requiring tenant-wide access. The fix uses a live token round-trip as the test, and also ensures the failure flag clears reliably when credentials are corrected.

  * **Activity log event type filtering** — `read_activity_log` and `read_agent_activity_log` now list all valid `eventType` values (e.g. `WebhookCallEvent` was previously missing from the docs). Passing an unknown event type now returns an error with the valid options instead of silently returning zero rows.
</Update>

<Update label="July 1, 2026" description="Claude Sonnet 5, faster Slack acknowledgement">
  ## New Model: Claude Sonnet 5

  Claude Sonnet 5 is now available as a selectable model. You can choose it from the model selector — it has a 1M token context window and supports adaptive thinking. The default `claude-sonnet` alias remains on `claude-sonnet-4-6` for now; Sonnet 5 is available as an opt-in while we validate it.

  ## Improvements

  * **Faster Slack 👀 acknowledgement** — Slack agents now show the eyes reaction immediately when a message arrives, instead of waiting for the background worker to pick it up. You'll see confirmation right away.
</Update>

<Update label="June 30, 2026" description="Gmail attachments, folder visibility cascade, admin-blocked capability discovery">
  ## Read Gmail Attachments

  Agents with Gmail access can now read files attached to emails. Two new tools make this possible:

  * **`list_gmail_attachments`** — lists all attachments on a Gmail message (filename, MIME type, size).
  * **`download_gmail_attachment`** — downloads an attachment and saves it as an agent document, ready to use with `get_document` (text/PDFs) or `get_media_content` (images).

  Previously the Gmail capability was text-only — attachments were invisible to agents.

  ## Folder Visibility Cascade

  You can now hide entire document folders from your agent. When you set a folder to hidden, the visibility cascades to all documents and sub-folders inside it — they are all hidden at read-time without changing individual document settings. The file explorer shows a greyed-out indicator on documents that are hidden via an ancestor folder.

  ## Improvements

  * **Admin-blocked capability discovery** — Capabilities that a workspace admin has blocked are now shown as greyed "Not available" cards in the Built-in tab instead of disappearing entirely. You can click them to see why they're unavailable. The Built-in tab sorts unavailable capabilities below available ones, and category badges show an available/total count when some are blocked.
</Update>

<Update label="June 29, 2026" description="Save to Google Drive, PDF export fixes and improvements, Slack app manifest fix">
  ## Save Chat Documents to Google Drive

  You can now save any chat document directly to Google Drive. A new "Save to Google Drive" option in the chat document menu creates a native Google Doc in your own My Drive and opens it in a new tab. If you haven't connected Google yet, you'll be guided through a quick OAuth flow.

  ## PDF Export Improvements

  Several improvements to document PDF export:

  * **HTML, SVG, and React documents** now export correctly as PDFs — previously these produced a PDF full of raw source code instead of the rendered page. They now use browser print for a proper visual export.
  * **Images** are now included in PDF exports of image-type documents.
  * **Large data collections** (over 1,000 rows) no longer show the Download as PDF option, preventing misleadingly partial exports. Use Export Data for full datasets.
  * **Emoji** in documents are now handled gracefully in PDFs instead of showing as missing glyphs.

  ## Bug Fixes

  * **Slack app connection** no longer fails for agents with long descriptions. Slack enforces a 140-character limit on app descriptions — the manifest fields are now properly truncated so connecting any agent to Slack works regardless of description length.
</Update>

<Update label="June 26, 2026" description="WhatsApp capability, multi-team secrets, scoped workspace API, ISO 27001">
  ## WhatsApp Capability

  Agents can now communicate with users directly over WhatsApp. Using Meta's Cloud API, you can give any agent a WhatsApp number and have it hold real 1:1 conversations with contacts — with full chat continuity per sender, image support, and mandatory sender whitelisting to control who can reach your agents.

  ## Read-only Workspace API

  A new read-only REST API at `/workspaceapi` lets you query workspace and agent data programmatically, authenticated with existing workspace API keys (`wk_`). Keys now support selectable scopes to limit what each key can access.

  ## Assign Secrets and API Capabilities to Multiple Teams

  Secrets and HTTP API capabilities can now be assigned to multiple teams at once. Use the multi-select team picker in the Secrets and API Capabilities editors to share access across teams without duplicating credentials.

  ## Improvements

  * **ISO 27001 certified** — The security docs now reflect our ISO 27001 certification.
  * **Account deletion now auto-cancels subscriptions** — Deleting your account automatically cancels all active Stripe subscriptions, so you're never charged after deletion.
  * **Chat top bar is now width-aware** — The action bar in chat smoothly switches between icon-only and full-label layout based on its own width, improving the experience in narrow or document-panel layouts.
</Update>

<Update label="June 24, 2026" description="Mistral Medium 3.5, chat delete permissions fix">
  ## New Model: Mistral Medium 3.5

  Mistral Medium 3.5 (128B, 256k context) is now available via Scaleway Generative APIs. It supports vision and configurable reasoning, making it a strong mid-tier model option.

  ## Bug Fixes

  * **Chat delete options now respect access level** — The delete chat options (single and bulk) were incorrectly shown to users with Use-level access. They are now correctly hidden for users without Edit permissions.
</Update>

<Update label="June 23, 2026" description="MCP server icons across more portal surfaces">
  ## Improvements

  * **MCP server icons** are now displayed consistently across the portal — on MCP tab cards, the enabled integrations tab, the settings drawer, and in the "Add a server" library. Icons are matched to known server templates and shown with a small add badge in the library view.
</Update>

<Update label="June 22, 2026" description="Team guest permission fix">
  ## Bug Fixes

  * **Team guests can no longer create agents in teams** — A bug allowed workspace members with guest-level team access to create new agents in those teams. This has been fixed; the restriction now applies at every layer (agent creation, snapshot import, and clone). Only team members and above can create agents in a team.
</Update>

<Update label="June 19, 2026" description="edit_file_in_vm false-success fix">
  ## Bug Fixes

  * **`edit_file_in_vm` no longer silently fails on non-matching edits** — The tool previously reported success even when `oldString` wasn't found in the file, leaving the file unchanged with no indication that the edit didn't take effect. It now returns `success: false` with a clear message when there are zero matches, so agents can detect and recover from missed edits.
</Update>

<Update label="June 18, 2026" description="Advanced MCP OAuth config, disabled model indicator in chat, LLM reliability improvements">
  ## Advanced OAuth Configuration for MCP Connections

  The MCP connection auth dialog's advanced section now includes additional fields for servers that auto-discovery can't fully configure:

  * **Custom token endpoint URL** — For servers like NetSuite where the token endpoint lives on a different host than the authorize endpoint.
  * **Explicit scopes** — Set a scope string that is honored on both the auto-discovery and custom-auth paths.
  * **Client ID / Client Secret** are now always visible in the advanced section, not only when the server signals it requires custom credentials.

  ## Disabled Model Indicator in Chat

  When an agent is configured with a model that has been disabled, the chat now makes this immediately visible: the model name appears in red with a "(disabled)" badge in the action bar, and the message composer is blocked with a clear warning — instead of sending into a guaranteed error. Agents with disabled models now also fail fast at startup instead of silently substituting a different model.

  ## Improvements

  * **LLM reliability** — Transient provider outages (5xx/429/overloaded) and Claude streaming socket drops are now retried automatically, reducing failed scheduled runs and long agent turns caused by temporary network blips.
</Update>

<Update label="June 17, 2026" description="Webhook HMAC signing, PDF downloads, voice mode fix, admin→owner rename, bug fixes">
  ## Webhook HMAC Signature Verification

  Inbound agent webhooks can now be **secured with HMAC signature verification**. Configure a signing secret on a webhook definition and the platform will verify the signature header before dispatching — requests with invalid or missing signatures are rejected with 401. A new dedicated **Webhook signing secret** type keeps signing secrets clearly separate from API keys. Covers the common HMAC-over-raw-body pattern used by Jira, GitHub, Linear, Shopify, and similar services.

  ## Download Documents as PDF

  Documents can now be downloaded as **PDF** from the document panel menu and the file-explorer row menu, sitting right next to the existing "Download as docx" option.

  ## Improvements

  * **Agent Management filter terminology** — The "Admin" labels in the workspace Agent Management filter panel have been renamed to "Owner", completing the terminology migration on that page.
  * **`code_execution` tool-call view** — The chat now shows bold `Output:` and `Return value:` labels with execution time in the header for `code_execution` tool calls.

  ## Bug Fixes

  * **Voice mode restored** — Fixed HTTP 400 errors from the OpenAI Realtime API caused by an unsupported `session.temperature` parameter that broke every voice session.
  * **Workspace admin team role** — Admins who add themselves to a team can now change their own role (previously shown as a static "MEMBER" badge with no dropdown).
  * **Chat title generation errors** — Fixed a regression causing repeated `LlmJsonParseError` errors during chat title generation.
  * **Stale agent load error** — Navigating away from an inaccessible agent now correctly clears the error state, preventing a persistent error screen on subsequent pages.
  * **Gemini 3 thinking parameter** — Disabling thinking on Gemini 3 models now uses the native `thinkingLevel: 'none'` instead of the deprecated `thinkingBudget: 0`.
</Update>

<Update label="June 16, 2026" description="Guest access by email, Scaleway Qwen 3.6 35B, Gemini Vertex AI credentials, Continue after tool-call limit, file explorer performance">
  ## Guest Access — Invite by Email

  People can now be given access to a single agent **without becoming full workspace members**. From an agent's User Access settings, admins can invite anyone by email. If they're already a workspace member the grant is applied immediately; otherwise a guest invitation is sent.

  Guests can log in but only see the agents they've been explicitly granted access to — workspace, team, secrets, and API pages are off-limits. Guests **don't consume seats** and are hidden from non-admin member and team lists.

  ## Scaleway EU Inference — Qwen 3.6 35B

  A new EU-hosted model is now available: **Qwen 3.6 35B Preview** via Scaleway's serverless inference. Scaleway is registered as a subprocessor from 2026-06-16.

  ## Gemini via Vertex AI / Project ID

  Gemini can now authenticate against **Vertex AI** without an API key. A `projectId` field in the vendor credentials lets Gemini use Application Default Credentials (ADC) — ideal for GCP-hosted deployments where the runtime service account already has the required `aiplatform.user` role.

  ## Continue After Tool-Call Limit

  When an agent exhausts the per-turn tool-call loop limit (100 iterations), portal chat now shows a **warning card with a Continue button** instead of a generic error. Clicking Continue resumes the task without truncating history, so long agentic workflows can pick up cleanly.

  ## Improvements

  * **Copy API Capability** — Each API Capability card now has a Copy button that opens the creation modal pre-filled with all fields from the source (name suffixed with "(copy)", secrets stripped for safety).
  * **File explorer performance** — Document listing queries no longer load full text bodies, fixing "Failed to load folder" errors in folders with many content-rich documents. Character counts are still shown via a stored `textContentLength` field.
  * **Instruction-cleanup skill** — New `instruction-cleanup` platform skill guides a behavior-preserving tidy-up of an agent's full instruction suite (main instructions + referenced documents) — without changing what the agent actually does.
  * **`upload_file_to_vm` with `documentId`** — Agents can now transfer existing agent or chat documents directly into a VM without loading the content into LLM context first.
</Update>

<Update label="June 15, 2026" description="get_models tool, document deep-links & Copy link, output tokens raised to 48k, per-surface API & MCP allowlists">
  ## `get_models` Tool

  Agents now have a **`get_models` tool** to enumerate all available models at runtime. The `<models>` section in the system prompt has also been streamlined to show only model aliases — keeping prompts leaner without losing discoverability.

  ## Document Deep-Links & Copy Link

  **Deep-linking to documents** is back: `?view=maximized` and `?view=fullscreen` query parameters on the Documents page now correctly open the document in the chosen view. A new **Copy link** button on the Documents page makes it easy to share direct links to any document or app.

  ## Higher Output Token Limit

  Max output tokens raised to **48k** across all LLM providers — Claude, Gemini, GPT, and Grok. Agents working on long-form content, large code files, or detailed reports can now produce significantly more in a single response.

  ## Per-Surface API & MCP Allowlists

  **API Capabilities and MCP servers** are now properly scoped per surface. Widget conversations only see the capabilities configured for them — API Capabilities and MCP servers no longer leak into widget contexts where they shouldn't be available.

  ## Improvements

  * **Better `query_document_data` errors** — Unsupported filter operators now return a helpful error listing all valid operators, instead of just naming the bad one.
</Update>

<Update label="June 12, 2026" description="Auto-transcription & document descriptions, Claude Fable 5, MCP server templates, document action menus unification, secrets & API fixes">
  ## Auto-Transcription & Document Descriptions

  Uploaded binary files (PDF, Word, Excel, PowerPoint, etc.) are now **automatically transcribed** in the background — no more needing to explicitly ask the agent to read them. The agent's system prompt now correctly reflects the file's contents from the moment it is uploaded. A **description is also auto-generated** after transcription for easier discovery. Auto-transcription now works across all document scopes: agent, workspace, and team.

  ## Claude Fable 5

  **Claude Fable 5** (`claude-fable-5`) is now available as a model option. Anthropic's latest top-tier model, featuring adaptive thinking support and a 1 million token context window.

  ## MCP Server Templates

  MCP server connectivity is now driven by a **code-defined template library** rather than admin-curated capabilities. Connect to any server from the template gallery or add one by URL — both produce standard user MCP servers on your agent.

  ## Improvements

  * **Unified document actions** — All document actions (Chat about this, Rename, Move, Publish, Export/Import data, Download, Delete) are now consistently available with the same ordering in both the file explorer row menu and the document panel menu.
  * **Friendly no-access page** — Opening a link to a private agent you don't have permission to view now shows a clear, friendly "no access" message instead of a generic error.
  * **Dropbox MCP** — The Dropbox MCP server template now correctly guides you to set up your own Dropbox App credentials, since Dropbox requires a custom OAuth app.

  ## Bug Fixes

  * **Field-based secrets not saving** — Fixed a silent bug where Snowflake, Cloud Storage, Salesforce, and Meta Ads secrets were silently not being saved when added or edited from the workspace Secrets page.
  * **Document API insert validation** — `POST /data/insert` now returns clear 400 errors with a helpful hint when the wrong field name is used (`rows` instead of `items`). The public docs have also been corrected.
  * **`insert_document_data` with null items** — Fixed a crash when `insert_document_data` was called with null or missing `items`.
</Update>

<Update label="June 11, 2026" description="Responsive member lists, agent creation fix, transcription credit tracking, download file extension fix">
  ## Improvements

  * **Responsive members & invitations lists** — The workspace members and pending-invitations tables now adapt to narrow viewports, rendering as stacked cards on mobile and as a grid on wider screens.

  ## Bug Fixes

  * **Agent creation with long names** — Creating an agent whose name (combined with the workspace name) produced a handle longer than 64 characters could fail outright. The handle collision auto-suffix (`…2`, `…3`) now works correctly for long names, and concurrent create requests are handled more robustly.
  * **Credits for document transcription** — Using the portal "Transcribe" button or the `transcribe_agent_document` tool now correctly deducts credits, matching the behaviour of all other transcription paths.
  * **Download text content file extension** — Downloading a document's text content now gives the file the correct extension (e.g. `.txt`, `.js`, `.css`) based on its MIME type, instead of always using `.md` or no extension.
</Update>

<Update label="June 10, 2026" description="Endpoint Management capability, workspace tags for teams & members, SVG sandbox rendering, security fixes">
  ## Endpoint Management Capability

  Agents can now programmatically manage their own exposed surfaces via a new **Endpoint Management** capability. Five tools let agents list, add, update, and remove HTTP API endpoints, webhooks, MCP server tools, and document API entries — and toggle surface-level settings like CORS and activity logging.

  ## Workspace Tags for Teams & Members

  Tags are now available on **teams and workspace members**, not just agents. A unified workspace tag vocabulary covers all three entity types, and filtering by tags works across the agents page, teams page, members list, and usage views.

  A new **Workspace Tag Manager** capability (admin-only) gives agents four tools to list and update tags on agents, teams, and members in bulk.

  ## Bug Fixes

  * **SVG documents** — SVGs are now rendered in a sandboxed cross-origin iframe, matching how HTML documents are handled. This prevents embedded `<style>` blocks from affecting the app layout and closes a latent XSS vector.
  * **Email approval cards** — Email HTML previews in send-email approval cards are now sanitized with DOMPurify, preventing malicious or overly broad `<style>` rules from restyling the app.
  * **Stale document panel** — Clicking "New chat" no longer causes a document that was auto-opened in the previous chat to reopen in the new chat.
</Update>

<Update label="June 9, 2026" description="Member adoption analytics, Document API, Workspace API keys, file explorer drag-and-drop">
  ## Member Adoption Analytics

  Workspace and team admins now have a **Member adoption** tab in Usage — with period-scoped KPIs, a daily active-member trend chart, team filter, and a paginated member table. The Members list also shows each member's last login and last activity timestamps.

  ## Document API

  Agents can now expose specific documents via **built-in REST endpoints** authenticated by API keys. Supports reading, querying, inserting, updating, and deleting data — no custom HTTP endpoint scripts required. Ideal for partner integrations that need direct document access.

  ## Workspace API Keys

  New **workspace-scoped API keys** (`wk_` prefix) authenticate against any agent in the workspace that has opted in. Managed from Workspace → API Keys.

  ## File Explorer Drag & Drop

  **Drag and drop** is now available in the file explorer on desktop: move documents and folders between scopes by dragging. You can also drop files from your OS to upload them directly into any scope.

  ## Improvements

  * **Image thumbnails** — Image files now show real thumbnails in the file explorer and search results.
  * **Agent Settings layout** — Agent Main Settings has a cleaner, more consistent design with an improved notifications panel.
  * **SMS/MMS pricing** — Outbound SMS now consumes 5 credits per message, 10 per MMS, billed per delivered recipient.
  * **Skill-files + Decompression** — `DecompressionStream` and `CompressionStream` are now available in `code_execution`. New `skill-files` agent skill for installing `.skill` ZIP packages.

  ## Bug Fixes

  * **Email sign-in** — Fixed auth failures for users who originally signed up via Google OAuth with mixed-case emails and later tried the email magic link.
</Update>

<Update label="June 8, 2026" description="Export/clone with chat history, mobile home screen (PWA), gpt-realtime-2 for phone calls, workspace notifications, live voice stability fix">
  ## Export/Clone with Chat History

  Agents can now be exported or cloned with their conversation history. When enabled, the snapshot includes messages and any chat-scoped uploaded files. Default off (history can be large or sensitive); imports restore chats automatically.

  ## Mobile Home Screen (PWA)

  Abundly can now be added to your iOS or Android home screen via "Add to Home Screen". Once installed, it launches without browser chrome and maintains a more persistent login — especially useful on iOS where Safari limits session duration.

  ## gpt-realtime-2 for Phone Calls

  A new *Latest* voice model option (gpt-realtime-2) is available in Make/Receive Phone Call capability advanced audio settings. The existing *Stable (gpt-realtime-1.5)* remains the default — switch to Latest to get the newest model.

  ## Improvements

  * **Workspace name in notifications** — Notification titles now include your workspace name (e.g. "failed in workspace "Acme""), so in-app notifications and emails share consistent, clear context.
  * **Agent usage Users KPI** — The Users count in agent usage analytics now correctly shows portal members only, excluding Slack participants, email senders, and bots.

  ## Bug Fixes

  * **Live voice stability** — Browser-based live voice sessions are now persisted to the database, fixing intermittent "Voice session not found" errors on multi-instance Cloud Run deployments.
</Update>

<Update label="June 5, 2026" description="MongoDB capability, multi-file apps for all scopes, live delegate_task activity, bi-weekly triggers, Slack DM threading">
  ## MongoDB Capability

  **MongoDB** is now available to all customers — discoverable in the portal and ready to connect without admin intervention. Add it from the Capabilities page, provide your connection string, and agents can query, insert, update, and delete documents directly. A dedicated docs page covers setup, permissions, and example prompts.

  ## Multi-File Apps: All Document Scopes

  Multi-file apps (the Play button in the document explorer) now work for **workspace, team, and chat** document scopes — not just agent documents. Build and share interactive apps from any document library.

  ## Live Sub-Agent Activity in Chat

  When an agent uses `delegate_task`, its work is now visible in a **collapsible accordion** in chat — tool calls, text output, and progress appear in real-time as the sub-agent runs. The timeline persists and reloads correctly after a page refresh.

  ## Bi-Weekly Trigger Schedules

  Recurring triggers set to "every other week" or "bi-weekly" now use proper nth-weekday cron syntax (e.g. `WED#1,WED#3` for every other Wednesday). Previously the scheduler fell back to a weekly-with-gate workaround that fired every week.

  ## Improvements

  * **Slack DM threading** — `send_slack_direct_message` now accepts a `threadTs` parameter so messages and file attachments land in the correct DM thread.
  * **Attachment viewer** — Attachments now open in a slide-in DocumentDrawer overlay (supports maximize/fullscreen). Activity log email attachments are also openable.
  * **Attack detection notifications** — Notifications now deduplicate within a 1-hour window per agent and group multiple attacks into a single notification with clear separation.

  ## Bug Fixes

  * **Chat auto-scroll** — Auto-scroll no longer permanently disables during heavy streaming. Layout shifts (thinking blocks expanding, tool results loading) no longer falsely register as user scroll.
  * **Workspace model policy for delegate\_task** — `delegate_task` now enforces live workspace model restrictions. Workspace-disabled models are properly rejected.
  * **Agent log display** — `delegate_task` sub-agent activity is correctly shown in the phase timeline. `code_execution` error output uses readable colors and script return values render properly.
</Update>

<Update label="June 4, 2026" description="Google Chat capability, agent usage analytics, document search, create/upload buttons, team People & Agents redesign">
  ## Google Chat

  New **Google Chat** capability: connect an agent to Google Chat as a classic HTTP Chat app. People can @mention it in spaces or DM it, and it responds with full context and conversation history. Each conversation maps to a persistent chat that appears in the portal sidebar.

  ## Agent Usage Analytics

  Workspace and team admins now have an **Agent usage** tab on the portal Usage page. Track total interactions, most active agents and channels, unique people reached, per-member interaction counts, and trend comparisons against the prior period.

  ## Document Search

  A free-text **search field** is now available on agent docs, workspace documents, and team documents pages. Search by name, description, or document type — results appear as a flat list with folder-path breadcrumbs, scoped to the current section.

  ## Create & Upload Document Buttons

  **Create document** and **Upload document** buttons now appear at the top of all document pages (agent docs, workspace Documents tab, team Documents tab), making it quicker to add new content without navigating folder menus.

  ## Redesigned Team People & Agents Tab

  The team **People & Agents** tab is now a unified card grid with search and All/People/Agents filters. Pending invites and support access appear as cards alongside regular members.

  ## Improvements

  * **Better import/clone error messages** — Agent import and clone failures now surface the specific reason instead of a generic message.
  * **OAuth2 JSON token bodies** — HTTP API capabilities using OAuth2 client credentials now honor a `Content-Type: application/json` override for token endpoints that require JSON bodies.

  ## Bug Fixes

  * **MCP secret preview** — API token previews for MCP capabilities were showing garbled base64 characters instead of the last 4 digits of the token.
</Update>

<Update label="June 3, 2026" description="FortNox invoice billing, Google Drive service account auth, agent docs bulk actions, 'Full Access' rename, ordered list fix">
  ## FortNox Invoice Billing

  Workspaces not using Stripe can now top up credits via FortNox invoice directly from the billing page. After confirming the purchase, an invoice is issued and credits are added to your workspace. Admin configuration options for FortNox OAuth, per-workspace payment terms, and a purchase history overview are also included.

  ## Google Drive: Service Account Authentication

  The Google Drive capability now supports **service account** authentication as an alternative to personal OAuth. A service account connects to Drive with the exact access it has been granted via file/folder sharing — no personal token or saved file allowlist required. Shared Drive support is included. Setup steps and an IT team request template are available directly in the capability drawer.

  ## Agent Documents: Unified Explorer & Bulk Actions

  The agent documents page now uses the same tree-based file explorer as workspace and team documents, showing workspace, team, and agent scopes in a unified view. Multi-select is available on all full-page document surfaces with bulk **move**, **download**, and **delete** actions.

  ## 'Owner' Renamed to 'Full Access'

  The top agent access level is now labelled **Full Access** everywhere in the UI (dropdowns, helper text, toast messages). This avoids confusion with the separate *Agent Owner* role. The underlying permission values are unchanged.

  ## Bug Fixes

  * **Ordered lists** — Numbered lists in chat messages and documents now render with numbers instead of bullet points.
  * **Stale secret references** — When a secret or HTTP API capability is deleted, stale bindings are now cleaned up immediately.
  * **Chat input overlap** — The sticky chat input bar no longer renders behind an expanded message edit textarea near the bottom of a long chat.
</Update>

<Update label="June 2, 2026" description="Document explorer bulk download/delete, workspace/team document publishing, admin-agents write access policy, agent management search fix">
  ## Document Explorer: Bulk Download / Delete

  Added **Download all** and **Delete all** actions to the document explorer across all scopes — workspace, team, agent, and chat. These actions appear in section headers and folder rows, making it easy to bulk-export or clean up entire document sets.

  ## Workspace & Team Document Publishing

  Workspace and team documents can now be published (made publicly accessible via link). The document explorer and document settings expose the same publish flow already available for agent documents. A note has been added to the documents feature page documenting this capability.

  ## "Admin Agents" Write Access for Shared Libraries

  A new **admin agents** access policy is available for workspace and team document libraries. It sits between "no agents" and "all agents" — only admin-scoped agents may create or edit documents in the shared library. Useful when you want AI assistants to manage shared content without opening writes to every agent in the workspace.

  ## Improvements

  * **Agent management search** — The search box on the Agent Management page is now immediately responsive, even on large agent lists. Previously the input could feel laggy because value updates ran inside `startTransition`.
</Update>

<Update label="June 1, 2026" description="Agents can create workspace/team documents, unified move_document tool, 'No tags' agent filter, inline credit limit modal, user settings tabs, code block copy toolbar, deprecated models hidden">
  ## Agents Can Create Documents in Workspace & Team Libraries

  Agents with edit access can now create new documents and folders in workspace and team libraries — not just edit existing ones. The `add_document` and `add_folder` tools now accept `workspace` and `team` scopes, with the same permission checks that govern editing.

  ## New `move_document` Tool

  A single `move_document(documentId, targetScope?, targetFolderId?)` tool replaces the previous `move_to_folder` and `promote_document_to_agent_scope` tools. Moving a document into a folder and moving it to another library scope are now one unified operation.

  ## Inline Credit Limit Modal

  When an agent hits its daily credit limit, the warning bar's "Edit limits" link now opens an inline modal so you can adjust the limit without navigating away from your chat.

  ## User Settings: Profile, Preferences & Tokens Tabs

  The profile/settings page has been reorganized into three focused tabs: **Profile** (avatar, name, email), **Preferences** (theme, language, notifications), and **Personal Tokens** — no behavior change, just a cleaner layout.

  ## Code Block Copy Toolbar

  A consistent copy button now appears on code blocks throughout the UI — markdown, widget settings, admin LLM log renderers, and activity log script sections all use the same compact toolbar.

  ## Deprecated Models Hidden from Pickers

  Deprecated model versions no longer appear in the eval model picker, grader picker, or agent settings model picker. Only supported models are shown.

  ## "No Tags" Filter in Agent Management

  A new "No tags" filter option is available on the Agent Management page. Toggle it on to find agents with zero tags — useful for bulk-moving untagged agents to junk or assigning them tags. The filter is mutually exclusive with specific tag filters and works across both the Agents and Agent Limits tabs.

  ## Improvements

  * **Scheduled triggers always sync** — Instruction edits now always re-sync scheduled triggers, regardless of subscription status, credit balance, or capability state. Previously, triggers could silently fall out of sync if limits were hit during an edit.
  * **Workspace name in notification emails** — Alert emails now include the workspace name in both the subject line and body, so recipients managing multiple workspaces can immediately tell which one triggered the alert.
  * **Shell: /workspace and /team mounts documented** — The shell capability's `/workspace` and `/team` paths are now described in the tool docs and skill guide. These paths were already supported but undocumented.
  * **"Move to team" in agent gear menu** — Workspace and team admins can now move agents using the per-agent gear icon menu, not just the Agent Management bulk-action bar.

  ## Bug Fixes

  * **Members view: collapsed team badges** — Team badges in the Workspace Members table are now capped at 2 visible per row, with a `+N more` toggle, preventing very tall rows for users with many team memberships.
  * **Workspace dashboard: active member count** — The "active in last 14 days" member count on the workspace dashboard was always 0. The activity tracking now correctly counts member activity.
</Update>

<Update label="May 29, 2026" description="Google Drive visual PDF extraction, Claude Opus 4.8, chat widget unified pipeline & hardened prompts, document explorer improvements">
  ## Google Drive: Visual PDF Extraction

  A new `read_google_drive_pdf_visually` tool is now available for the Google Drive capability (both full access and single-file variants). Agents can run vision-based PDF page extraction directly on Drive files by file ID — no need to import the file first. Uses the same visual extraction pipeline as `read_pdf_pages_visually`.

  ## Claude Opus 4.8

  Claude Opus 4.8 is now available in the model picker. It supports adaptive thinking and a 1M token context window.

  ## Chat Widget: Unified Pipeline

  Chat Widget conversations are now stored as real platform chats — visible in the portal chat list alongside Slack and other channels, with full message history and admin tooling. Internal reasoning (`thinking` content) is no longer streamed to end users in widget sessions.

  ## Chat Widget: Hardened System Prompt

  Widget system prompts now only describe the capabilities that are actually enabled in the widget configuration — the same whitelist used to filter the widget toolbox now also controls which sections appear in the system prompt. A dedicated widget prompt template adds behavioral guidance for the compact viewport, media embedding rules, and blocks unsupported portal XML tags (such as `<capability_card>`).

  ## Document Explorer Improvements

  * **Context-based actions** — "Mention in chat" always appears in the chat documents side panel; "Chat about this" always appears in agent document views. Actions now depend on which UI surface you're in.
  * **Metadata in kebab menus** — File explorer row menus now show a read-only info block (description, last edited time, file size) above the action items. The same metadata is also shown in the document drawer.
  * **Scope-correct badges** — Composer attachment chips and in-message document links now display the document's actual scope.
</Update>

<Update label="May 28, 2026" description="Agent management kebab menu fix">
  ## Bug Fixes

  * *Agent Management table layout* — Fixed the ⋮ action menu on agent rows breaking onto a new line for workspaces that don't use Agent Value Capture. The row cell count now correctly matches the active grid template whether the Value column is shown or not.
</Update>

<Update label="May 28, 2026" description="Chat widget CDN improvements, file explorer row actions & Run app button, unified move modal, per-scope document write access, Salesforce binary downloads, code_execution fetch fix, agent value capture">
  ## Chat Widget Improvements

  The chat widget embed now uses a versioned CDN loader — a stable `widget.js` that pulls the latest versioned bundle, with short TTL for the loader and immutable cache headers for bundles. Internal reasoning (`thinking` content) is no longer streamed to end users in widget sessions. Comprehensive public docs are available at `/features/chat-widget` with a live showcase.

  ## File Explorer: Row Actions & Run App

  Every file and folder row in the unified file explorer now has a kebab (⋮) menu with Rename, Move, and Delete actions — no need to open the document first. App folders and entry files get an inline *Run app* button directly in the row, making it easy to launch React apps without navigating away.

  ## Unified Move Modal

  Moving documents across scopes and folders is now handled by a single **Move** flow: pick the destination scope (workspace, team, agent, chat…) then pick a folder. All move types — within-scope re-parenting, cross-scope file and folder moves — go through one consistent experience.

  ## Per-Scope Document Write Access

  Workspace and team admins can now control who can write documents in each scope. Toggle policies for humans and agents independently — for example, allow all agents in a team to write to team documents, or restrict writes to admins only. The shell's `/workspace/` and `/team/` mounts also respect these settings, so agents with write access can modify shared documents through shell scripts.

  ## Agent Value Capture

  Replace the old free-text value impact field with structured, categorized value entries and an operational criticality level. Workspace admins configure the category catalog and workspace-wide requirements; team admins can tighten (but not loosen) policy; agent owners fill in values on their agent's Main Settings page.

  ## Salesforce Binary File Downloads

  The `salesforce_api_request` tool can now save binary responses (PDFs, attachments) as agent documents via `saveResponseAsDocument`, matching how Notion, GitHub, HubSpot, and other vendor API tools already work. The `http_request` tool also gains a new `responseDocumentMimeType` parameter for explicitly setting the MIME type of saved binary files.

  ## Bug Fixes

  * **code\_execution fetch()**: `fetch()` inside `code_execution` scripts now correctly handles chat-scoped and trigger-scoped document URLs. Previously, accessing a file attached in chat from inside a script would fail with "access denied" — even though the agent had read access everywhere else.
  * **API Capabilities and Secrets sorted alphabetically**: Both lists in workspace settings are now sorted alphabetically by name, consistent with other portal lists like agents, teams, and members.
</Update>

<Update label="May 27, 2026" description="Meta Ads capability, voice prototype improvements, document explorer folders, shared drive fix, chat stability fix">
  ## Meta Ads Capability

  Agents can now connect to Meta's Marketing API for Facebook and Instagram paid advertising. Add a Meta system user access token as a workspace secret — tokens refresh automatically. The `meta_ads_api_request` tool provides access to campaigns, ad sets, ads, creatives, and performance insights.

  ## Voice Prototype: Persistent Transcripts, Turn Modes & More

  The live voice prototype (`/portal/.../voice`) has received a major batch of improvements:

  * **Persistent transcripts** — Voice sessions now sync all transcripts, generated images, and tool calls back into the bound chat, so sessions are naturally resumable as text conversations after they end.
  * **Text↔voice continuity** — Starting a voice session from inside an existing text chat carries the prior conversation context into the session.
  * **Turn modes** — Choose between hands-free, push-to-talk (PTT), and toggle modes. Interrupt controls and a mute toggle are also available.
  * **Cost explorer** — A live credit usage breakdown for the current voice session.
  * **Capability settings** — Full capability configuration is now accessible directly from within the voice session UI.

  ## Document Explorer: Folders and Unified Library

  The document explorer now supports nested folders across workspace, team, and chat document libraries. Kebab menus in every section and folder offer New Folder, New Document, and Upload actions. Agents can browse and search across workspace, team, and chat libraries with full folder awareness.

  ## Use-Level Access for Chat Documents

  Users with Use access (not just Edit) can now create, rename, upload, and delete chat-scoped documents in the side panel. Agent library documents still require Edit access.

  ## Bug Fixes

  * **Google Drive shared drives** — `view_google_drive_file` now correctly handles files on shared/team drives (previously returned 404).
  * **Chat: streamed text preserved on tool call** — When the model streamed text and then issued a tool call in the same message, the streamed text would briefly disappear. Text blocks are now merged rather than replaced, so the content stays visible.
</Update>

<Update label="May 26, 2026" description="Login domain restrictions, agent credit reset history, limits tab redesign, chat and document fixes">
  ## Login Domain Restrictions

  Workspace admins can now configure email domain restrictions self-serve from portal Workspace Settings → Login restrictions. Set which email domains are allowed to join your workspace — no platform admin involvement required.

  ## Agent Credit Reset History

  When a workspace admin manually resets an agent's daily credit counter, that action is now logged and surfaced in the UI. Account warnings show cumulative credits used since the last reset, limit cards show a "↻ N today" badge, and a Reset History panel shows who reset and when.

  ## Agent Management: Limits Tab Redesign

  The Limits tab in Agent Management has been reworked for clarity:

  * Agents are grouped by team with the same filters and search as the Agents tab
  * Compact strip cards show workspace defaults and team monthly limits inline
  * A docked side panel handles per-agent daily limit configuration (responsive on mobile)

  ## Bug Fixes

  * **Document drawer transitions** — Fixed render-lifecycle bugs causing state-lag glitches when switching between documents in the workspace/team Documents view.
  * **`read_pdf_pages_visually` with chat-uploaded PDFs** — The tool now correctly handles flat-layout document URLs used by chat uploads and generated files, which were previously silently rejected.
  * **Chat stability** — The final phase of the message-ID migration lands: the client now reconciles every socket event by stable message ID, eliminating edge-case glitches from positional merge logic.
</Update>

<Update label="May 25, 2026" description="Figma capability, published document viewer redesign, SharePoint certificate fix, chat stability improvements">
  ## Figma Capability

  Agents can now connect to Figma. Add a Figma Personal Access Token as a workspace secret and your agents get access to a `figma_api_request` tool — browse projects, read design files, render frames as images, and post comments, all via Figma's REST API.

  ## Published Document Viewer Redesign

  The public `/shared/document/` viewer has been redesigned around content-first principles:

  * **Mobile-friendly layout** — the page is now comfortable on small screens, not just desktop.
  * **Compact chrome** — publisher/source-agent badges are toned down so your document takes centre stage.
  * **"Copy to my agent"** — visitors can now clone a published document directly into one of their own agents from the viewer.

  ## Bug Fixes

  * **SharePoint app-only certificate** — The PEM private key field for SharePoint app-only (certificate) auth is now a multiline textarea. Previously, pasting a PEM key into the single-line password input silently stripped newlines, causing cryptic OpenSSL errors on the Microsoft side.
  * **Chat stability** — Chat messages now carry a stable ID end-to-end. This eliminates a class of subtle glitches (messages appearing cut, full chat not rendering, delayed responses) caused by positional merge logic between the socket stream and database-fetched history.
</Update>

<Update label="May 22, 2026" description="SharePoint app-only auth, My Chats filter, trigger notifications, eval improvements, thinking in activity logs, bug fixes">
  ## SharePoint App-Only Authentication

  You can now authenticate SharePoint using app-only (client credentials) mode — ideal for workspaces with Conditional Access policies that expire delegated tokens daily. The agent runs as its own Microsoft Entra identity, with no per-user token required. Configure it from the SharePoint capability settings.

  ## "My Chats" Filter

  A new filter in both the sidebar chat list and the full chats page lets you quickly see only chats where you've sent at least one message — useful for finding your own active conversations in busy workspaces.

  ## Trigger Notifications

  Agents now notify you when a trigger is blocked or fails. Three new notification categories: *Trigger Failed*, *Trigger Stopped by Plan*, and *Trigger Stopped by Security* — so you're always informed when something doesn't run as expected.

  ## Agent Optimization Skill

  Agents can now load an `agent-optimization` skill when asked how to reduce credit usage. It walks through diagnostic steps and concrete techniques ordered by typical impact.

  ## Eval Improvements

  * *Eval results in export/import* — Agent snapshots now include eval run results, with a separate opt-in toggle to include or exclude run history.
  * *Agent Context setting* — A new toggle in eval Execution Settings lets you run evals without the agent's identity (system prompt stripped), useful for benchmarking raw model behavior independently of agent configuration.

  ## Improvements

  * *Thinking in activity logs* — Trigger run timelines now show the model's reasoning and assistant text interleaved with tool calls, across all four LLM adapters (Claude, GPT, Grok, Gemini).
  * *Plan and security phases reinstated* — The plan → security → execute pipeline is back for screened triggers (email, SMS, script-escalations), while trusted triggers keep the simplified single-turn pipeline.
  * *`delegate_task` no longer requires `capabilities`* — Pure-reasoning sub-agent delegations with no extra capabilities are now valid.
  * *Image generation: source images no longer re-described* — Agents using the image generation tool with source images no longer waste tokens describing them in the prompt.

  ## Bug Fixes

  * *OpenAI Responses API fix* — A 400 error caused by replaying `status` fields as input on subsequent turns is fixed. Was hitting 32% of all errors in affected chats.
  * *Starred agents fix* — Clicking through starred agents no longer causes them to silently vanish from the Starred section until a page reload.
  * *Inline secret editor* — Fixed wrong default scope on existing secrets and missing protection for rich (user+key) access setups.
  * *Tool approval 404* — Missing tool approvals now return HTTP 404 instead of HTTP 500.
</Update>

<Update label="May 21, 2026" description="Workspace LLM model policy, email notifications default on, OAuth2 Basic auth, Slack formatting improvements, Update Instructions security fix">
  ## Workspace LLM Model Policy

  Workspace admins can now control how new Abundly catalog LLM models are treated. Choose between:

  * *Opt-out (default)* — New models are available automatically; disable the ones you don't want from the Model selection tab.
  * *Opt-in* — New models stay off until you explicitly enable them.

  When switching to opt-in mode, your currently enabled models are preserved — only models added to the catalog later require an explicit enable.

  ## Email Notifications Now On by Default

  Email notifications are now enabled by default for all users. If you previously hadn't opted in, you'll start receiving email notifications for agent activity. A clear opt-out link is included in every notification email footer, and you can manage preferences from your account settings.

  ## OAuth2 Client Credentials: Basic Auth & Extra Headers

  API Capability configs using OAuth2 client credentials now support:

  * *HTTP Basic auth* on the token endpoint (required by providers like Spotify and Fortnox)
  * *Extra request headers* on the token POST (e.g. Fortnox `TenantId`)

  Existing setups continue to work without changes.

  ## Improvements

  * *Slack formatting hints* — Agents using `post_slack_message` and `send_slack_direct_message` in non-chat contexts now receive inline formatting hints about Slack's mrkdwn syntax, reducing incorrect markdown output (e.g. `**bold**` instead of `*bold*`).
  * *Update Instructions security tightened* — The `chatOnly` setting on the Update Instructions capability now strictly blocks Slack, MS Teams chats, and autonomous triggers — only Abundly portal chat can update instructions when this setting is on.
</Update>

<Update label="May 20, 2026" description="Gemini 3.5 Flash, remove per-user access overrides, SVG preview improvements, personal token fixes, notification fixes">
  ## Gemini 3.5 Flash

  **Gemini 3.5 Flash** is now available in the model picker. Released May 19, it supports thinking mode and a 1M token context window.

  Alongside this, agents now correctly self-report their active model — if a per-context model override is set (e.g. different models for chat vs. scheduled triggers), the system prompt now reflects the actual model in use.

  ## Remove Per-User Access Overrides

  A red trash button now appears next to each per-user access override row in agent settings. Clicking it removes the override and reverts the user to the workspace default access level. A confirmation modal prevents accidental removal, with a last-owner guard to prevent locking yourself out.

  "Nothing (No Access)" has also been repurposed as an explicit deny — it now keeps the user in the list with explicit no-access rather than silently removing them.

  ## Bug Fixes

  * **Personal Tokens reconnect flow** — Reconnecting a personal token (e.g. Google Drive) from the Personal Tokens tab now returns you to the correct tab instead of the default Profile tab.
  * **Large SVG previews** — Tall/portrait SVGs now render at full drawer width with vertical scrolling, rather than being heavily letterboxed.
  * **SharePoint/Outlook personal tokens** — Removing a SharePoint, Outlook, or Drive Full personal token from agents was silently failing. Fixed.
  * **Notification recipients** — Support-access staff and removed users are no longer included in notification recipient lists.
</Update>

<Update label="May 19, 2026" description="Salesforce capability, Snowflake multi-field credentials, pending-approvals indicator, deferred doc transcription, Slack remote chat fixes, navigation improvements, bug fixes">
  ## Salesforce Capability

  You can now connect Salesforce to your agents. Configure an External Client App in your Salesforce org and enter the Login URL, Consumer Key, and Consumer Secret in the Salesforce workspace capability to get started.

  ## Snowflake Multi-Field Credentials

  Snowflake credentials are now configured as a multi-field workspace secret (account identifier, login name, and PAT), matching the Salesforce pattern. Existing agents continue to work without any changes.

  ## Pending-Approvals Indicator

  A warning triangle now appears next to the **Activity** tab and in the sidebar whenever an agent has pending tool approvals waiting for review. Never miss an approval again.

  ## Deferred Transcription on Agent Doc Uploads

  When you drag a PDF, image, audio, or video file into the agent docs view, it is no longer transcribed immediately on upload. Transcription happens only when the agent actually reads the file — saving you time and credits on files the agent may never need.

  ## Improvements

  * **Serialize Slack chat turns across instances** — When an agent receives multiple Slack messages in quick succession, turns are now serialized across Cloud Run instances and bursts are coalesced, preventing duplicate or interleaved replies.
  * **Monthly credit limit in Settings tab** — The "Set monthly credit limit" action has moved from the kebab menu to the Settings tab as an inline form, making it much easier to find.
  * **Member dropdown alphabetical sort in User Access** — The "Add rule for a specific user" dropdown in agent Settings → User access now lists workspace members alphabetically.

  ## Bug Fixes

  * **Slack file posts now appear in portal remote chat** — Files uploaded by an agent via `post_slack_message` in a Slack remote-chat thread now correctly appear in the Abundly portal.
  * **SVG text rendering** — SVG documents converted to PNG (e.g. via `post_slack_message` with `svgContent`) no longer render text as tofu boxes — fonts are now included in the Docker image.
  * **SVG previews no longer clipped** — SVG documents opened in the document drawer now display fully without the lower portion being cut off.
  * **Broken image chips fixed** — Chat document chips for inline-only artifacts (SVGs with no file URL) no longer show a broken-image placeholder.
  * **Command+click to open in new tab** — \~25-30 navigation controls across the app now support Command+click (Mac) / Ctrl+click (Windows) to open in a new tab.
  * **Slack posts to private channels** — Agents posting to a private Slack channel with `chat:write` scope no longer fail with `missing_scope` even when `channels:join` is absent.
  * **Notification and tool-approval fixes** — Multiple fixes to the notification/tool-approval flow; admins now receive a notification when a scheduled trigger's eval fails.
  * **Tab-hash navigation** — Clicking a notification that links to a specific tab (e.g. `#approvals`) now correctly activates that tab even when you're already on the page.
</Update>

<Update label="May 18, 2026" description="Notification settings per agent/team/workspace, rerun eval results, attack detection notifications, extended thinking for Claude 4.x models, light-mode code rendering fix">
  ## Notification Settings per Agent, Team, and Workspace

  Workspace and team admins can now configure notification categories at every level of the hierarchy. Open the **Notifications** card in agent settings, team settings, or workspace settings to control which notification types are enabled — for example, disable credit alerts on a specific agent while keeping them on for others.

  ## Rerun Eval Results

  Eval results now have a **Rerun** button. From the result detail view, rerun that specific eval/model combination. From a run summary, rerun the full eval lineup. Nested `code_execution` sub-calls are also now correctly counted in tool-call validation assertions.

  ## Attack Detection Notifications

  A new **Attack Detected** notification category has been added. When the platform's prompt-injection / attack detection flags an incoming trigger, you'll now receive an in-app notification. MCP, HTTP API, and Widget error notifications are also now properly wired up. The `McpHttpApiFailed` category has been split into separate **MCP Failed** and **HTTP API Failed** toggles for finer control.

  ## Bug Fixes

  * **Extended thinking on Claude 4.x models** — Claude Opus 4.5, Sonnet 4.5, Haiku 4.5, and Sonnet 4.0 were incorrectly marked as not supporting extended thinking. This is now fixed — you can enable extended thinking on these models as expected.
  * **Inline code rendering in light mode** — Inline `` `code` `` spans no longer render with a dark background in light mode; they now correctly adapt to the active theme.
</Update>

<Update label="May 15, 2026" description="In-app notifications, notification preferences, async approval alerts, capability settings cleanup, chat start message control, better export/import, low-credit threshold scaling, bug fixes">
  ## In-App Notification System

  Abundly now has a built-in notification centre. A bell badge on your user avatar shows unread notifications, and clicking opens a panel with your recent alerts — credit debits, scheduled trigger events, and more.

  ## Notification Preferences

  You can now control exactly which notifications you receive. Head to your user profile and open the new **Notifications** card to opt out of specific categories or enable email delivery for the ones that matter most.

  ## Async Tool Approval Alerts

  When an agent queues a tool call for human approval during an automated run (scheduled trigger, webhook, email, SMS, etc.), the people who can act on it now receive an in-app notification — and optionally an email — so approvals don't get missed.

  ## Improvements

  * **Agents can manage Chat Start Message settings** — Agents can now update their own `chatStartMode` and `chatStartMessage` via `update_agent_metadata`, letting them control how new chats begin programmatically.
  * **Capability settings cleanup** — The approval model and UI for all six communication capabilities (Send/Receive Email, SMS, Phone Call) are now consistent and simplified.
  * **Better export/import fidelity** — Document IDs embedded anywhere in an agent (instructions, descriptions, scheduled trigger messages, and more) are now correctly remapped when importing into a new environment.
  * **Low-credit alerts now scale with your plan** — Credit warning thresholds at agent, team, and workspace level now trigger at 20% of the configured allowance instead of fixed absolute amounts, making them relevant regardless of plan size.

  ## Bug Fixes

  * **Slack group DMs** — When two Abundly agents share a Slack group DM, they can now see each other's messages and collaborate properly.
  * **Capability drawer** — The capability side panel no longer auto-opens for users who don't have Edit access on the agent.
  * **Credit refill timing** — The credit refill calculation now correctly aligns with the GCP scheduler at 00:00 UTC, fixing rare edge cases where refill status could appear off.
</Update>

<Update label="May 13, 2026" description="Team Analytics & Explorer capabilities, admin restriction tiers, Slack reply threading, collapsible Teams sections, model vendor grouping, chat document chip polish, bug fixes">
  ## Team Analytics & Team Explorer

  Team admins now have two new capabilities to explore and analyze their team's data:

  * **Team Analytics** — usage metrics, member activity, and agent stats scoped to a single team, mirroring the workspace analytics view
  * **Team Explorer** — read-only access for team admins to browse their team's agent documents and data

  ## Capability Access Control: Admin Restriction Tiers

  The per-capability **Admin-only** toggle has been upgraded to a three-tier picker:

  * **None** — any workspace member can enable the capability
  * **Admin** — requires workspace admin
  * **Team admin** — requires team admin or above

  This gives workspace admins finer-grained control over which capabilities are available to regular members vs. admins.

  ## Slack: Replies to Agent-Initiated Threads Now Trigger the Agent

  When an agent with Slack chat mode posts a new top-level message in a channel (e.g. a daily release summary), replies to that thread now trigger the agent — without needing to @mention it.

  ## Improvements

  * **Teams page** — People and Agents sections are now collapsible, making it easier to navigate teams with many members or agents
  * **Workspace model selection** — Models are now grouped by vendor with bulk-toggle and a **Latest** badge on alias models
  * **Chat document chips** — Unified document card design across all chat surfaces, with auto-open in drawer on post

  ## Bug Fixes

  * Cloning or importing an agent with uploaded files (flat-layout GCS paths) no longer silently loses those files
  * Deleted workspaces no longer appear in the portal workspace picker or sidebar
</Update>

<Update label="May 12, 2026" description="Feedback modal before account deletion, portal navigation refresh, activity log nesting, trial restart on graduation, agent deletion fix">
  ## Feedback Before Account Deletion

  Before deleting an account, users now see a brief optional feedback modal — choose from common reasons (Too expensive, Missing features, Hard to use, Found a better alternative, etc.) and add free-text detail. The feedback is voluntary and the deletion flow continues either way.

  ## Portal Navigation Refresh

  The portal sidebar has been restructured to introduce a dedicated **Management** section, separating admin and everyday user views. Teams and Workspace surfaces now use the platform's canonical hierarchy color palette for clearer visual organization.

  ## Improvements

  * **Activity log: nested sub-agent calls** — Sub-agent tool calls now correctly appear nested under their `delegate_task` parent in the activity log, instead of showing flat.
  * **Trial restart on workspace graduation** — When graduating into an existing workspace, the destination workspace's trial is now refreshed (or freshly created) instead of being silently inherited.

  ## Bug Fixes

  * **Agent deletion** — Fixed a validation error that prevented users from deleting agents in certain cases.
</Update>

<Update label="May 11, 2026" description="Grok 4.3 model, A2A graph fix, mobile chat banner fix, iframe callTool improvements">
  ## Grok 4.3

  xAI's Grok 4.3 is now available as a selectable model. It features a 1M-token context window, adjustable reasoning, multimodal support, and thinking mode. Long-context billing tiers are applied automatically based on input token count.

  ## Improvements

  * **iframe callTool consistency** — Sandboxed React/iframe apps now route document helpers (`abundly.loadAgentDocument`, etc.) through the same toolbox path as `callTool`, giving consistent access control and a unified interface.

  ## Bug Fixes

  * **Private agents missing from A2A graph** — Private workspace-discoverable agents now correctly appear in the agent-to-agent graph, including SHOW ALL mode and the Add Connection picker.
  * **Mobile chat token-limit banner** — Fixed the token-limit warning banner taking over the full mobile viewport; it now stacks vertically on small screens.
</Update>

<Update label="May 8, 2026" description="Browser automation & Typst PDF in VM, GPT-5.5 model, generic Notion API, multi-image OpenAI gen, interrupted chat error handling, script improvements, bug fixes">
  ## Virtual Machine: Browser Automation & Typst PDF

  VM image v0.8 adds full browser automation capabilities and Typst PDF generation:

  * **Playwright browser automation** — Bundled Chromium (for npm/pip `playwright` scripts) and Google Chrome stable (for `playwright-cli`), with new `vm-browser-automation` skill covering stateful sessions, SPA waits, cookie-banner handling, and stealth escape hatches.
  * **Typst PDF generation** — The Typst binary is now available in the VM, enabling markup-to-PDF compilation with A4 defaults, math, and tables. A new `vm-typst` skill is included.

  ## GPT-5.5

  OpenAI's GPT-5.5 is now available as a selectable model. It has a 1.05M token context window and thinking support.

  ## Notion: Single Generic API Tool

  The 13 dedicated Notion tools (`get_notion_page`, `query_notion_database`, `append_block_children`, etc.) have been replaced by a single **`notion_api_request`** tool — a pre-authenticated generic REST tool, matching the pattern already used for GitHub and HubSpot. `upload_file_to_notion` is kept as-is for multi-step file uploads.

  ## Multiple Source Images for OpenAI Image Generation

  The `generate_image` tool (OpenAI) now accepts an array of source image URLs (up to 16), matching the Gemini interface. This enables multi-image use cases like remixes, compositing, and style transfer.

  ## Interrupted Chat Responses Handled Gracefully

  Errors, credit limit hits, and user-aborted responses are now surfaced clearly, persisted to the chat record, and billed correctly. Remote chats on Slack and Teams now post a user-facing error message when a response is interrupted, rather than silently dropping it.

  ## Script Improvements

  * **Unified `args` parameter** — Scripts now receive a single `args` object with shape `{ body, triggerType, headers, meta }`. Escalation is explicitly opt-in via `{ escalate: true, message }`. Updated in scripts documentation.
  * **`code_execution` `input` parameter** — The `args` tool parameter has been renamed to `input` to avoid confusion with the sandbox's internal `args`. The type is now a proper union so LLMs serialize structured values correctly.
  * **Improved script output in activity logs** — Script execution results now show both console output and return values, with better formatting.

  ## Bug Fixes

  * **Move agents between teams** — Workspace and team admins can now move agents between teams without hitting an access-denied error.
  * **VM inline images** — `download_file_from_vm` now returns the correct document URL, fixing broken inline image embeds.
  * **Google Drive picker sign-in** — Users not already signed into Google can now complete the OAuth flow and use the Drive file picker.
  * **Agent search input** — Fixed a race condition that caused keystrokes to be dropped when typing in the agent search/filter field.
</Update>

<Update label="May 7, 2026" description="MS Teams remote chat, Outlook/Teams/SharePoint for all, billing promotion codes, starred agents, MCP error feedback, eval model alias fix">
  ## Microsoft Teams as a Remote Chat Surface

  Microsoft Teams is now a full remote-chat surface on par with Slack. Channel threads, group chats, and 1:1 personal chats all map to persistent Abundly chats with proper threading, @mention support, and a Teams-tailored system prompt. In-thread continuation via replies is also supported.

  ## Outlook, Teams & SharePoint Available to All

  The Outlook, Microsoft Teams, and SharePoint capabilities are now visible and available to all customers — no admin-side enabling step required.

  ## Billing: Promotion Codes and Discounts

  Workspace admins can now apply Stripe promotion codes directly from the Plans/billing page without needing to switch to a different plan first. Subscription update flows also support promotion codes.

  ## Starred Agents at the Top of the Agent Selector

  Starred agents now appear in a dedicated section at the top of the agent selector dropdown, making it faster to access your favorites in workspaces with many agents.

  ## Improved MCP Setup Error Feedback

  MCP server connection errors now surface short, actionable messages — distinguishing timeouts, auth failures, and network errors — with a 15-second request timeout to avoid long hangs. The UI also better handles MCP servers that transition from authenticated to unauthenticated.

  ## Improvements

  * **Invitation emails** now show who sent the invite, reducing confusion for recipients.
  * **Evals tab** — The model selector modal is wider so alias names aren't truncated, and the UI has been generally tidied up.

  ## Bug Fixes

  * **Eval model aliases** (e.g. "Claude Sonnet Latest") are now correctly recognized as enabled models. Fixed a related issue where eval run banners could get stuck when all models were skipped.
</Update>

<Update label="May 6, 2026" description="Usage tab activity stats, faster triggers, document URL fixes, invite login fix, agent management UI">
  ## Workspace Activity Stats on Usage Tab

  The Usage tab now shows two new cards — **Members** and **Agents** — alongside the existing balance and average use cards. Each card displays the total count and how many were active in the last 14 days.

  ## Faster and Cheaper Triggers

  Trigger execution has been streamlined from three LLM turns (plan → security review → execute) to a single turn, reducing execution time and credit usage per trigger run. Attack detection is preserved as a separate pre-execution step.

  ## Bug Fixes

  * **Document URL resolution** — Fixed a bug where multiple tools (image generation, code execution, email attachments, PDF visual reads, and more) failed to resolve document URLs for files in the new flat GCS storage layout.
  * **Transcription failures** — Fixed transcription failures for documents stored under the new flat GCS path.
  * **Invite with login restrictions** — Fixed a bug where accepting an invite via the wrong login method (e.g. email when the workspace requires Google sign-in) would permanently consume the invite and lock the user out.
  * **Agent management column pinning** — The row actions menu in the Agents table now stays visible when the table overflows horizontally on narrow viewports.
</Update>

<Update label="May 5, 2026" description="Workspace & team document libraries, model selection, Send Email improvements, activity log nesting, buy credits quantity, Drive picker, image gen default">
  ## Workspace and Team Document Libraries

  Agents can now access **workspace-scoped** and **team-scoped** document libraries in addition to their own agent library. A unified document explorer lets workspace and team admins manage shared documents from new Documents tabs, and the chat side panel shows all libraries an agent has access to.

  ## Workspace Model Selection

  Workspace admins get a new **Model selection** tab under Workspace Management. Use it to turn off specific catalog LLM models for the entire workspace — helpful for controlling costs or standardizing model usage. Platform-wide disabled models remain locked and cannot be re-enabled.

  ## Send Email: Safer Defaults and Clearer UI

  The Send Email capability has been redesigned with better defaults out of the box: the owner is always on the allowed list, recipients outside the list require approval, and the agent is notified when decisions are made. The old modal is replaced by inline settings in the capability drawer with a plain-English summary of the configured behavior.

  ## Activity Log: Nested callTool Calls

  Tool calls made from within `code_execution` (via `callTool`) now appear **nested under the parent call** in the activity log. This makes it much easier to follow what happened inside a script-based execution.

  ## Improvements

  * **Buy Credits quantity** — You can now select how many credit packs to purchase (1–99) directly on the Buy Credits card before going to Stripe. The total credits and price update live.
  * **Google Drive file picker** — The left-hand navigation pane (My Drive, Shared with me, Recent, Starred) is now shown again, and filtering has been improved for easier file discovery.
  * **Default image generation changed to GPT** — New agents now default to OpenAI's GPT Image for image generation instead of Gemini. Agents that have explicitly chosen Gemini are unaffected.

  ## Bug Fixes

  * **Agent crashes on missing credentials** — Fixed a crash that killed entire agent runs when a configured capability referenced a deleted or corrupted credential.
  * **OAuth state corruption** — Personal-token OAuth flows (Google, Microsoft, Autodesk) now correctly encode state as base64url JSON, preventing corruption with non-ASCII names. After OAuth completes, integration UIs also correctly reload updated credential information.
</Update>

<Update label="May 4, 2026" description="Agents page redesign, in-platform contact support, activity log improvements, chat rename fix">
  ## Redesigned Agents Page

  The portal Agents page has been completely reworked to surface the agents you actually use. It now features a **Starred** section for favorites, a **Recently Used** section (which you can hide via preferences), and a flat A–Z / Recent-first toggle for everything else. Search and tag filters persist in the URL, and disabled agents visually recede to reduce clutter.

  ## In-Platform Contact Support

  A new **Contact support** flow is now available from the **Help & support** menu in the sidebar. A 2-step modal lets you describe your issue (category, short summary, details), review before sending, and submit directly — no need to track down a support email address.

  ## Activity Log Improvements

  * Nested tool calls now show **parent-tool links**, making it easy to follow the call chain within a single run.
  * Script-based scheduled runs now appear as dedicated entries in the activity log.
  * A new **Activity** docs page explains the activity log in detail.

  ## Improvements

  * **Reset filters** — A "Reset" button now appears in Agent Management when filters are active, letting you clear them in one click.

  ## Bug Fixes

  * **Chat rename** — Users with Use-level access can now rename their own chats. Previously this incorrectly required Edit access.
  * **Document folder validation** — The platform now validates that a document's parent folder belongs to the same agent when creating or moving documents.
</Update>

<Update label="April 30, 2026" description="askAgent in all contexts, API wildcard patterns, generated images as documents, help menu redesign, one-off script triggers">
  ## askAgent() Works Everywhere

  The `askAgent()` function in bypass scripts now works in **all contexts** — chat, triggers, and webhooks — not just triggers. It uses `delegate_task` with full context under the hood, and requires the Delegate Task capability to be enabled.

  ## Wildcard API URL Patterns

  API Capability URL patterns now support wildcards **anywhere** in the URL, not just at the end. For example, `*.shopify.com/admin/api/*` or `https://api.example.com/v*/users` now match correctly.

  ## Generated Images as Documents

  Images created with `generate_image` now automatically appear as chat-scoped documents. This means agents can promote them to the document library, reference them by ID, and manage them like any other document.

  ## Help Menu Redesign

  The floating help bubble has been replaced with a **Help section in the user menu**. It includes links to documentation, changelog, feedback, service status, and support — all in one place.

  ## One-Off Script Triggers

  One-off scheduled triggers now support `scriptDocumentId`, allowing them to run a bypass script directly instead of the full agent pipeline — matching what was already available for recurring triggers.

  ## Improvements

  * **Basic auth hint** — The API capability form now clarifies that Basic auth secrets should be stored as `username:password` (the platform handles encoding).
  * **Capabilities tab cleanup** — Workspace admin capabilities are now shown at the top for admins; duplicate MCP server list removed.
  * **Pricing page** — The trial button now shows "25,000 credits included".

  ## Bug Fixes

  * **Agent move from management page** — Fixed the bulk/single agent move action that was failing with a 400 error.
  * **Generated image documents** — Fixed an edge case where the document wasn't created for agents without certain capability settings.
  * **Agent import/clone file copy** — Document files are now properly copied in cloud storage when importing or cloning an agent.
</Update>

<Update label="April 29, 2026" description="Help bubble, delegate_task fullContext, VM downloads as documents, Owner rename, workspace analytics tables, folder delete fix">
  ## In-Product Help Bubble

  A new floating **Help** button now appears on all portal pages, giving you quick access to documentation, changelog, feedback links, and a Getting Started guide. You can show or hide it from your user preferences.

  ## delegate\_task: Full Context Mode

  The `delegate_task` tool now supports a `fullContext: true` option, letting an agent delegate tasks to itself with its full system prompt and tool set — but a clean context window. Useful for long-running conversations where accumulated history makes each turn expensive.

  ## VM Downloads Appear in Documents

  Files downloaded via `download_file_from_vm` now appear in the documents panel (chat, trigger, or agent library). Use the optional `scope` argument to control where they're stored.

  ## "Admin" Renamed to "Owner"

  The top agent access level is now called **Owner** everywhere in the UI — in agent settings, user access, agent management, and agent cards. This makes it clearer that workspace-level admin and agent-level top access are separate concepts.

  ## Workspace Analytics: More Tables

  The Workspace Analytics capability gains two new tables:

  * `agent_user_access` — per-user access grants on agents
  * `team_memberships` — replaces the previous lossy single-team column

  The `agents` table now also exposes `defaultAccessLevel`, `linkAccessLevel`, and `adminOnly`.

  ## New Model Support

  Claude **Opus 4.6** and **Opus 4.7** are now available as model choices for `read_pdf_pages_visually`. **Claude Sonnet 4.6** is also now available as a model choice and has been set as the default Sonnet model.

  ## Improvements

  * **Admin-only capabilities** are now shown as disabled cards on non-admin-only agents, so workspace admins can discover and enable them without first toggling the agent's admin-only setting.

  ## Bug Fixes

  * **Folder deletion** now cascades correctly — deleting a folder removes all its contents. Previously, children were orphaned and kept appearing in agent LLM contexts. Non-empty folders now show a confirmation modal before deletion.
  * **Agent Management on narrow screens** — The Agent column no longer collapses when the window is narrow; less-important columns hide responsively instead.
</Update>

<Update label="April 28, 2026" description="Virtual Machine for all, workspace analytics, admin-only agents, evals improvements, capability panel fixes">
  ## Virtual Machine for Everyone

  The **Virtual Machine** capability is now publicly available — no need to contact support to enable it. It lets your agent run code, create Office documents (Word, Excel, PowerPoint), and generate PDFs. Find it in the Built-in capabilities tab in your agent settings.

  ## Workspace Analytics

  A new **Workspace Analytics** capability gives workspace admins SQL-based access to aggregated, metadata-only snapshots of their workspace's agents, chats, triggers, capabilities, usage, credits, and evals — all powered by a built-in DuckDB instance. Requires an admin-only agent (see below).

  ## Admin-Only Agents

  Workspace admins can now restrict an agent to admin-only access, preventing non-admin users from using it. This is required for sensitive capabilities like Workspace Analytics and is ideal for internal admin tooling.

  ## Credit Details Defaults

  Set a **workspace-wide default** and per-team override for the Credit Details setting. New agents automatically inherit the setting via team → workspace → off. A new bulk action in Agent Management lets you update all existing agents at once.

  ## Improvements

  * **Capability panel toggle** — Enable or disable a capability directly from the settings panel without going back to the capability card. OAuth from the chat panel now restores the panel on return and auto-links the new token.
  * **Evals tab** — Model Settings is now a collapsible card with a summary line showing configured models. Eval-run and grading credits are now shown separately in result views. New "Chat with agent" button to get started with evals.
  * **Admins column** — The Admins column is back in the agent management table, visible at a glance without expanding each row.

  ## Bug Fixes

  * **Scheduled agent crash** — Fixed a crash that occurred when a scheduled agent referenced a deleted or missing tool credential.
</Update>

<Update label="April 27, 2026" description="New SharePoint tools, eval model filtering, thinking mode & context meter fixes">
  ## New SharePoint Tools

  Four new tools are now available for SharePoint integrations: **move/rename items**, **copy items** (cross-site capable with automatic retry), **delete items** (to recycle bin — recoverable), and **create folders**. These expand the SharePoint toolkit to cover comprehensive file and folder management.

  ## Eval Improvements

  Admin-disabled LLM models are now properly excluded from eval runs. Disabled models no longer appear in the eval model add-picker, and already-selected disabled models are shown as orange warning tags with a tooltip explaining they will be skipped during the run.

  ## Thinking Mode & Context Meter Fixes

  * The thinking mode toggle is now hidden for models that don't support it, keeping the interface clean.
  * Corrected Opus 4.7's thinking support flag — thinking mode is now correctly configurable for that model.
  * Fixed the context fullness meter, which was previously underestimating usage by \~3×.
  * The context weight indicator now opens as an interactive popover with more detail.
</Update>

<Update label="April 24, 2026" description="Save messages log, file downloads in agent apps, gpt-image-2 upgrade, Autodesk fixes, typing lag fix">
  ## Save Messages (Email & SMS Message Log)

  Workspace operators can now enable an optional **message log** for email and SMS interactions. When enabled, incoming and outgoing messages are saved and viewable from a dedicated monitoring view — with pagination and tool save hints for operators building message-processing workflows.

  ## File Downloads in Agent-Built Apps

  Agent-built React apps can now generate and trigger **client-side file downloads** (Excel, Word, ZIP, and more). The frame-src CSP now includes `blob:` and `data:` URIs, and the react-documents skill includes a worked example and a CDN library table covering popular file formats.

  ## Upgraded Image Generation

  OpenAI image generation now uses **gpt-image-2**, offering improved image quality.

  ## Autodesk Integration

  Fixed several issues with the Autodesk integration:

  * OAuth now works correctly for all users (not just the authorizing user)
  * Token refresh no longer requires re-authentication
  * Forge Viewer correctly selects 2-legged or 3-legged tokens based on the use case

  ## Improvements

  * **Soft-delete for admin resources** — Deleting users, workspaces, and agent configs now uses a soft-delete approach with a recovery grace period, consistent with other platform resources.
  * **Agent details in Usage Reports CSV** — Usage report exports can optionally include a per-agent breakdown.
  * **RSS feeds** — Permanently broken RSS feeds are no longer retried indefinitely.

  ## Bug Fixes

  * **Typing lag on new chat** — Fixed a debounce issue causing typing lag when opening a new chat.
  * **Auto-resize edit message textarea** — The edit-message input now grows to fit the content.
  * **Last activity in CSV export** — Fixed the Last activity column in workspace and team management CSV exports.
  * **Agent document ownership check** — Document deletion now correctly verifies the requesting agent owns the document.
</Update>

<Update label="April 23, 2026" description="Built-in docs search tool, Trello event filters, admin oversight for private agents, code rendering improvements, soft-delete with recovery, dark-mode icons">
  ## Built-in Documentation Search

  Agents now have a built-in `search_abundly_docs` tool that searches the Abundly documentation site. When you ask your agent a question about the platform, it can look up the real documentation instead of relying on its training data.

  ## Trello Per-Agent Event Filters

  The Trello integration now supports **per-agent event filtering**. Instead of receiving all board events, you can configure each agent to listen for specific action types (card created, moved, commented, etc.), optionally scoped to specific lists. A new generic `TrelloActionEvent` replaces the previous fixed set of event types, enabling broader coverage of Trello activity.

  ## Admin Oversight for Private Agents

  Workspace admins can now view key details of private agents directly from the agent management page — including description, instructions, agent-to-agent connections, and exposure hooks (MCP server, HTTP API, webhooks, widget). This gives admins visibility into private agents without needing explicit access.

  ## Code Rendering Improvements

  Script documents (JavaScript, TypeScript, Python, etc.) now render with **proper syntax highlighting** in view mode, including toggleable line numbers and a copy button. All code blocks across the app use a consistent dark theme.

  ## Soft-Delete with Recovery Period

  Deleting secrets, HTTP API configurations, agent documents, and agent chats now uses a **soft-delete** approach — items are marked for removal but remain recoverable during a grace period before permanent deletion.

  ## Improvements

  * **Dark-mode vendor icons** — Model selectors now show proper light-fill icons for GPT and Grok in dark mode.
  * **Chat hero alignment** — The new-chat hero text and elements now align correctly with the chat input width.
</Update>

<Update label="April 22, 2026" description="In-chat A2A connections, agent discoverability, batch eval runs, chat start modes, HTTP API CORS, model provider icons, LLM vendor credentials">
  ## In-Chat Agent-to-Agent Connections

  You can now set up and tear down agent-to-agent connections **directly from a chat conversation**. Ask your agent something like "connect me to the Research agent" and it emits a connection card with a toggle switch — flipping it adds or removes the target from the agent's allowed agents list and toggles the Call Other Agent capability as needed. No need to navigate to Settings → Agent Communication.

  ## Agent Discoverability Settings

  A new per-agent **discoverability** setting controls which other agents can find and contact your agent — independently of human user visibility:

  * **None** — no other agent can discover or reach this agent
  * **Team** — only agents in the same team can contact this agent
  * **Workspace** — all agents in the workspace can discover this agent

  ## Batch Eval Runs

  You can now **select a subset of evals** and run them as a single batch from the eval list. Previously you could only run one eval at a time or all at once — now you can pick exactly the evals you need.

  ## Chat Start Mode

  The `generateStartMessage` setting is replaced by an explicit tri-state **Chat Start Mode**:

  * **Silent** (default) — no opening message; the user types first
  * **Fixed** — the agent sends a seeded opening message
  * **Auto** — the agent generates a context-aware opening message

  Brand-new agents now get an onboarding nudge suggesting a use-case discovery skill if no job has been defined yet.

  ## HTTP API: Per-Agent CORS Settings

  Agents with an HTTP API endpoint can now configure which **browser origins** are allowed to make requests. The HTTP API settings UI includes a "Allow requests from Abundly apps" checkbox and a freeform input for custom origins.

  ## Model Provider Icons & Cost Tiers

  Model selectors throughout the portal now show **provider logos** (Claude, GPT, Gemini, Grok) and **cost tier indicators** (dollar-sign icons) — making it easy to compare models at a glance by provider and relative cost.

  ## LLM Vendor Credentials

  Admins now configure LLM credentials **once per vendor** (Anthropic, OpenAI, Gemini, xAI) rather than row-by-row per model. All catalog models from that vendor become available automatically. Pricing changes ship via deploys instead of live database edits.

  ## Improvements

  * **Slack channel name in agent context** — Agents now see the human-readable channel name (e.g. `#dev`) in their Slack context alongside the channel ID, avoiding unnecessary API calls to look it up.
  * **"User access"** — The agent access settings tab is renamed from "Team access" to "User access" for clarity.
</Update>

<Update label="April 21, 2026" description="Human-friendly shell view, eval import overwrite, code execution timeout increase, export fix">
  ## Human-Friendly Shell Activity View

  The Shell capability now has a **Human-friendly mode** toggle in its settings. When enabled, raw terminal commands in the live thought bubble and tool-call headers are replaced with plain-language descriptions and emoji illustrations — for example, `cat report.md | grep revenue` becomes "📖 Reading report.md and 🔍 looking for 'revenue'". Toggle it off to see the raw commands as before.

  ## Eval Import: Overwrite Existing Evals

  The eval import modal now has an **"Overwrite existing evals (matched by name)"** checkbox. It only appears when the import has name conflicts, and lets you update existing evals in place rather than skipping them. The preview distinguishes between evals that will be created, overwritten, or skipped, and the import summary reports counts for each.

  ## Improvements

  * **Code execution timeout** — The default timeout for code execution has been raised from 5 seconds to **30 minutes**. Agents doing long-running work (delegating to sub-agents, polling APIs) no longer need to manually set a `timeoutSeconds` value to avoid interruption.
  * **`github_api_request`: more token-efficient file editing** — Base64 file content is now stripped by default from GitHub API responses, making file read/edit workflows significantly more token-efficient. Raw bytes can be requested by setting `includeBase64: true`, ideally from inside `code_execution`.

  ## Bug Fixes

  * **Export Agent from Settings** — The Export Agent modal's Export button was always disabled when opened from agent Settings. Now works correctly.
  * **HTTP API endpoint validation** — Creating an HTTP API endpoint with an empty path now correctly shows an inline error instead of silently failing.
</Update>

<Update label="April 20, 2026" description="Google Drive token reconnect, Drive file improvements">
  ## Google Drive: Reconnect Without Losing Files

  When a personal Google Drive token expires or is revoked, you can now **reconnect it in place** directly from the Drive capability card, the Drive Files panel, or the Personal Tokens manager. Previously, the only recovery path was to delete the token and recreate it — wiping every agent's curated file list in the process.

  The reconnect flow re-runs Google OAuth consent for the existing token while keeping the same token ID, so all agents that reference it inherit the refreshed credentials automatically. A same-account check prevents accidentally switching to a different Google account.

  ## Improvements

  * **Drive file listing** — Files are now fetched individually instead of via a bulk list, removing the previous 100-file limit and enabling per-file accessibility status. File names and metadata are cached so they remain visible even after a token loses access.
  * **Clearer Drive error messages** — The portal now distinguishes between a token-level auth failure (prompts you to reconnect) and a per-file access issue (shows which files are inaccessible), rather than showing a generic error.
</Update>

<Update label="April 17, 2026" description="Google Drive full access, Virtual Machine capability rename, agent management improvements, billing UI enhancements">
  ## Google Drive Full Access

  Agents can now connect to **Google Drive with full access** using personal tokens. After connecting via OAuth, you can optionally restrict access to specific drives or folders directly in the portal — giving you fine-grained control over what files the agent can read and write.

  ## Improvements

  * **"Virtual Machine" capability** — The capability previously called "Sandbox" has been renamed to **Virtual Machine** throughout the portal, making it clearer that it provides a full cloud VM environment.
  * **Agent management: filter overlay** — The filter panel on the agent management page is now a right-side overlay drawer. Multi-select filters (capabilities, teams, tags, models) now use searchable checklists for easier navigation.
  * **API capabilities and MCP servers in agent management** — Agent cards in the management view now show API capability and MCP server badges in the expandable section, with searchable filters for both in the filter overlay.
  * **Billing: credit allowance reference line** — The credit balance history chart now shows a dashed line at your monthly credit allowance level, making it easy to see how your balance relates to your plan.
  * **Approval tab improvements** — The approvals list now has filter buttons (All / Pending / Approved / Rejected) with server-side pagination for better performance with large approval lists.

  ## Bug Fixes

  * **Trial expiry with remaining credits** — The account-status banner now correctly appears for workspaces whose trial period has expired but still have credits remaining.
</Update>

<Update label="April 16, 2026" description="http_request save binary responses as documents, agent avatar thumbnails, credit widget fixes, billing plan fixes">
  ## http\_request: Save Binary Responses as Documents

  The `http_request` tool now supports a `saveResponseAsDocument` option. When enabled, agents can download binary files — PDFs, images, ZIPs, and more — and save them directly as agent documents, instead of attempting to serialise binary data as JSON (which previously corrupted the output).

  ## Improvements

  * **Faster agent avatar loading** — Agent avatars now use compact WebP thumbnails in sidebar and list views, significantly reducing bandwidth and improving load times.

  ## Bug Fixes

  * **Credit usage widget** — The sidebar credit bar now updates immediately when a chat error occurs or when you switch back to a browser tab, preventing stale or incorrect credit counts after background work.
  * **Credit widget for non-admin users** — Users with Use or Edit access to an agent (but not a workspace admin) now see correct credit usage in the sidebar.
  * **Billing page** — Customers whose subscription plan has been archived in Stripe now continue to see their current plan correctly on the billing Plans tab, with the correct "Current subscription" badge.
</Update>

<Update label="April 15, 2026" description="Voice call cost tracking, comprehensive agent management filters, support access management">
  ## Voice Call Cost Tracking

  AI voice calls now show detailed credit usage in agent log entries. Each call's log entry includes the duration, provider, and a full cost breakdown covering telephony minutes and OpenAI Realtime token usage (audio and text, input and output).

  ## Comprehensive Agent Management Filters

  The agent management page now has a full filter bar with seven dimensions: status, privacy, capabilities (with text search), credits used today (min/max range), last changed, last usage, and admin user. Active filters are visually highlighted, making it easy to find specific agents in large workspaces.

  ## Improvements

  * **Support user access** — Workspace admins can now add support users as admins on private agents directly from the agent overview row. Revoking a support user's workspace access automatically removes them from all agent access lists. A Support badge is now shown next to support members in the agent user access settings table.
</Update>

<Update label="April 14, 2026" description="Cloud storage capability, Slack trigger mode, sandbox streaming uploads, chat shimmer UI, Office file sandbox suggestions, remove disabled MCP servers, voice recording preserved on credit limit">
  ## Cloud Storage Capability

  Agents now support a new **Cloud Storage** capability with S3-compatible API tools. Connect to AWS S3, Google Cloud Storage (HMAC), Cloudflare R2, MinIO, Backblaze B2, or any other S3-compatible provider. Five tools are available: list objects, get object metadata, get object content, put object, and delete object — with size guardrails and pagination support.

  ## Slack Trigger Mode

  The Slack integration now supports two modes, selectable in the capability settings:

  * **Chat mode** (default) — the existing behaviour, where Slack messages arrive as agent chat messages
  * **Trigger mode** — raw Slack Events API deliveries are passed directly to the agent as a `SlackTriggerEvent`, enabling custom handling of any bot event (channel joins, reactions, etc.) via agent instructions or scripts

  ## Sandbox: No More File Size Limit

  Sandbox file uploads and downloads now stream data directly between cloud storage and the sandbox, removing the previous 50 MB file size limit. Files of any size are now supported (bounded only by sandbox disk space).

  ## Chat Feedback and Loading Indicators

  The chat interface now feels more responsive:

  * **Progress messages** appear when the agent hands off to a tool, so you know work is happening
  * **Shimmer loading states** on thinking blocks, tool calls, and thought bubbles give smooth visual feedback during processing

  ## Improvements

  * **Office file suggestions** — Agents with the Sandbox capability disabled now correctly suggest enabling it when you ask to create Word, Excel, or PowerPoint files (`.docx`, `.xlsx`, `.pptx`), rather than suggesting unrelated alternatives
  * **Remove disabled MCP servers** — You can now remove a disabled MCP server from an agent without having to re-enable it first

  ## Bug Fixes

  * **Rename Agent** — The Save button in the Rename Agent modal is now correctly enabled when accessed from the Agent Settings page
  * **Voice recording preserved on credit limit** — Fixed a bug where hitting a credit limit (team, daily, or account blocked) while recording audio via microphone would silently destroy the in-progress recording or transcription; the input area now stays mounted during an active recording flow
</Update>

<Update label="April 13, 2026" description="Value/impact field for agents, chat filter improvements, Slack paragraph break fix, document names in tool call UI, http_request document body">
  ## Value / Impact Field for Agents

  You can now add an optional **Value / impact** description to each agent, editable under Agent Settings → Basic Information. This description appears as a column in the agent management overview — with a tooltip for long values and a subtle "Not set" hint when empty. The field is preserved when cloning or exporting agents.

  ## Chat Filter Improvements

  The chat sidebar filter has been updated for clarity:

  * **"All chats"** now truly shows all chats, including remote (e.g. Slack) chats
  * The filter previously called "Remote chats" is now **"Hide remote chats"** — acting as a toggle to exclude transport-bound chats when desired
  * The "Hide remote chats" option is hidden when an agent has no remote chats
  * The filter now resets to "All chats" when you switch between agents, preventing confusing empty states

  ## Improvements

  * **Slack responses** — When an agent produces multiple text segments in a single turn (before and after a tool call), these are now separated by paragraph breaks in Slack instead of running together on one line.
  * **Tool call UI** — Document names are now shown for array document ID parameters (e.g. in `delegate_task`), matching the annotation already shown for single `documentId` fields.
  * **`http_request` tool** — A new `dataDocumentId` parameter lets you send an agent document file as the raw request body, useful for forwarding files to external APIs.
</Update>

<Update label="April 10, 2026" description="SharePoint fixes and improvements, agent document transcription, capabilities deep links, document panel auto-reload">
  ## SharePoint Fixes and Improvements

  Several improvements to the SharePoint integration:

  * **Correct site access enforcement** — Site-level access restrictions now correctly match compound site IDs, fixing cases where restrictions could silently fail to apply.
  * **Case-insensitive library matching** — Document library names like "Documents", "documents", and "Shared Documents" are now treated equivalently.
  * **Tenant URL in tool descriptions** — The agent now knows your SharePoint base URL pattern after connecting via OAuth, reducing guesswork.
  * **Clearer setup guidance** — SharePoint and Outlook connection instructions have been updated to correctly direct you to the personal token flow and Entra admin consent, preventing confusing suggestions about Azure AD app registrations.

  ## Agent Document Transcription

  When file-to-text conversion is configured on your workspace, agents now have access to a `transcribe_agent_document` tool that converts stored documents (PDFs, Office files, images) to text — matching the "Transcribe to text" action available in the portal UI.

  ## Capabilities Deep Links

  The Capabilities page now supports direct links to specific capability panels via a `?open=` URL parameter. OAuth return flows and MCP setup now bring you directly back to the right panel after completing external auth steps, instead of landing on the general capabilities overview.

  ## Document Panel Auto-Reload

  The document side panel now automatically reloads after an agent finishes editing or transcribing a document, so you always see the latest content without a manual refresh.
</Update>

<Update label="April 9, 2026" description="Agent descriptions on overview cards, portal performance improvements, Slack webhook fixes">
  ## Agent Descriptions on Overview Cards

  Agent cards on the overview page now display the agent's description directly in the card — no more opening each agent to understand what it does. When an agent has no tags, the description appears in the tag area, truncated with a tooltip for the full text on hover.

  ## Improvements

  * **Portal performance** — The agent list, chat list, HTTP APIs, and secrets pages now load significantly faster. Large workspaces that previously saw 3–4 second load times should see sub-100ms responses.
  * **Slack webhook verification** — Custom webhook endpoints now automatically handle Slack-style URL verification handshakes, making it easier to connect third-party services that use this verification pattern.

  ## Bug Fixes

  * **Slack capability enforcement** — Fixed a regression where Slack webhooks were processed even when the Slack capability was disabled on an agent.
</Update>

<Update label="April 8, 2026" description="Slack remote chat unification, chat filter menu, iframe tool calling, HTTP auth improvements">
  ## Slack Remote Chat Unification

  Slack is now fully unified as a remote chat channel. All Slack webhook paths flow into the same persistent-chat pipeline as the Abundly portal, so agents behave consistently whether users talk to them in Abundly or Slack. Conversations, history, and context are shared across both surfaces.

  ## Chat Filter Menu

  The sidebar chat filter has been upgraded from a simple toggle to a dropdown pill with three options: **All chats**, **Starred chats**, and **Remote chats** — each with a live count badge. This makes it easy to browse only Slack/remote chats separately from portal chats.

  ## Iframe Tool Calling

  Sandboxed HTML and React iframe documents can now call allowlisted agent tools directly using `abundly.callTool(name, params)` — the same interface as `callTool` in code execution blocks. This enables richer, more interactive embedded UIs that can read data, query collections, and invoke approved agent capabilities without leaving the iframe. A warning is shown in the publish modal for documents using this feature, since tool calls are disabled on public share links.

  ## Improvements

  * **HTTP API secret injection** — Secret handling is now split into three clear tiers: manual placement, auto-injection (bearer, API key, basic auth, custom header, query param), and managed auth (OAuth 2.0, GCP service account, JWT). The API auth UI groups these tiers clearly so it's easy to pick the right approach.
  * **Scheduled trigger error clearing** — Error messages on scheduled trigger runs are now automatically cleared when the next run completes successfully, so stale error states no longer linger in the UI.
</Update>

<Update label="April 7, 2026" description="Slack as a chat channel, unread indicators, self-service invoices, context fullness fix">
  ## Slack as a Chat Channel

  Slack is now a first-class chat channel for agents. Incoming Slack messages are routed into persistent agent chats — the same abstraction as the portal — so conversations, history, and behavior are unified whether the human is chatting in Abundly or Slack. Thread handling and ordering keep busy Slack threads usable.

  ## Unread Indicators for Background Chats

  When you have multiple chats open, non-active chats now show a small red dot when they receive new messages or activity — so you never miss what's happening in the background. The sidebar toggle icon shows an aggregate dot when any background chat is unread, even when the list is collapsed. Unread state persists across page reloads and syncs across tabs.

  ## Self-Service Invoices for Credit Purchases

  Credit top-up payments now automatically generate a proper Stripe invoice (including VAT/tax). A new **Invoice** button on the Payments tab lets you download invoices directly, without contacting support.

  ## Improvements

  * **Accurate context fullness indicator** — The conversation fullness ring now starts at \~0% on a fresh chat (previously \~20%), since platform overhead is excluded from the calculation. The ring grows as your actual conversation content accumulates.
  * **Voice input protection** — The send button and Enter key are now disabled while a voice recording or transcription is in progress, preventing accidental message sends that would discard the in-progress audio.

  ## Bug Fixes

  * **Billing page access** — Invoice-billed workspaces (non-Stripe) could reach the billing page via a leaked link in the user menu or by navigating directly to the URL. Both entry points are now blocked with a redirect to the workspace home.
</Update>

<Update label="April 2, 2026" description="Granular SharePoint access, Claude Adaptive Thinking, voice call error handling, activity log fix">
  ## Granular SharePoint Access

  SharePoint personal tokens now support **site-level and folder-level access restrictions**. After completing the OAuth flow, a new settings page lets you choose exactly which SharePoint sites — and optionally specific folders within those sites — an agent can access. A recursive folder tree picker makes it easy to drill down to any depth. All SharePoint tools enforce these restrictions at runtime, so agents can only see what you've explicitly allowed.

  ## Claude Adaptive Thinking

  Thinking for Claude 4.6 models (Opus 4.6 and Sonnet 4.6) has been upgraded to the new **Adaptive Thinking** API. This replaces the previous fixed-budget extended thinking mode and resolves token exhaustion issues that could cause thinking-enabled requests to fail. Thinking-enabled agents on Claude 4.6 should see improved reliability.

  ## Improvements

  * **Voice call error handling** — When a voice connection fails, callers now hear a friendly spoken message instead of dead silence. Swedish is supported with a localized fallback message.
  * **API capability panel stability** — Fixed an issue where opening "Add secret" from the in-chat API capability flow would reset the panel and drop users back to the empty "configure new API" form after creating a secret.

  ## Bug Fixes

  * **Activity log tool call counts** — Fixed activity log tools reporting double the actual number of tool calls, which caused agents reading their own activity logs to incorrectly conclude their executions were redundant or inefficient.
</Update>

<Update label="April 1, 2026" description="Slack bot-to-bot communication, Telnyx SMS provider, local timezone scheduled triggers, sidebar chat improvements">
  ## Slack Bot-to-Bot Communication

  Agents can now **participate in conversations with other Slack bots**. Previously, all bot-originated Slack events were silently dropped. The new system uses smart safeguards — a self-loop guard blocks the workspace's own app ID, and a runaway streak detector pauses triggers when 6+ consecutive bot messages appear within a 60-minute window. @-mentions in outgoing messages are automatically resolved to Slack user IDs so bots can tag each other effectively.

  ## Telnyx SMS Provider

  A second SMS provider — **Telnyx** — is now available alongside Twilio. Workspace admins can choose the provider at agent, workspace, or global level. Inbound SMS/MMS from Telnyx is verified with Ed25519 signatures and deduplicated automatically.

  ## Local Timezone for Scheduled Triggers

  Scheduled trigger cron expressions are now **interpreted in the agent's local timezone** (defaulting to Europe/Stockholm) instead of UTC. This means when you write "9 AM" in a cron schedule, it fires at 9 AM local time — with DST transitions handled automatically. Existing triggers have been migrated.

  ## Sidebar Chat Improvements

  The chat sidebar now shows **all your chats** instead of limiting to 10. A new toggle switches between "all chats" and "starred chats" views with a count badge. The context menu has been unified with the full chats page, adding rename, star/unstar, and delete options.

  ## Improvements

  * **Shell tool UI** — Shell commands now render in a distinctive terminal-style panel with a `$` prompt and monospace font, with live typewriter animation during execution
  * **Environment name in browser tab** — Non-production environments now show "Abundly Test" or "Abundly Stage" in the browser tab title
  * **Gemini streaming reliability** — Transient errors during Gemini streaming (429/503) are now retried automatically with safeguards against duplicate output

  ## Bug Fixes

  * **API capability secret binding** — Fixed the secret binding dropdown not appearing for managed auth methods (GCP Service Account, JWT Bearer, OAuth 2.0 Client Credentials) on capability cards
</Update>

<Update label="March 31, 2026" description="Agent name in browser tab, restricted API capabilities visibility, URL-synced pagination">
  ## Agent Name in Browser Tab

  Browser tabs now display **your agent's name** in the title — formatted as `AgentName | Abundly` — making it easy to identify which agent you're working with when multiple tabs are open. The agent name appears first so it stays visible even when tabs are narrow.

  ## Restricted API Capabilities Visibility

  API capabilities that are restricted to specific agents or teams are now **shown as greyed out** with a "Restricted" badge, instead of being hidden entirely. This gives you visibility into what capabilities exist even when they aren't available to the current agent.

  ## Improvements

  * **URL-synced pagination** — Page numbers across admin and portal pages are now synced to the URL, making them bookmarkable and preserved on refresh
</Update>

<Update label="March 30, 2026" description="Multi-field secrets, user account deletion, virtual shell, generic HubSpot & Trello API tools, nested code execution, subagent skills access">
  ## Multi-Field Secrets

  Secrets can now store **structured key-value pairs** with individual fields marked as sensitive or non-sensitive. Agents reference specific fields using a dedicated syntax, with sensitive values decrypted only when needed. The secrets page has been redesigned with a card-based layout, expandable details, and inline create/edit support from chat panels — including per-field rotation controls for updating individual values.

  ## User Account Deletion

  You can now **delete your own account** from your profile page. Deletion uses a 30-day soft-delete grace period — during this time your login is blocked, but you can contact support to reinstate your account. After 30 days, all your data is permanently removed.

  ## Virtual Shell Capability

  Agents can now use a **virtual shell** to run bash commands, execute Python or Node.js scripts, and manage files in a secure sandbox environment — useful for data processing, automation scripts, and technical workflows.

  ## Generic HubSpot & Trello API Tools

  The HubSpot and Trello integrations now each use a **single generic API tool** (similar to the GitHub integration) instead of multiple hardcoded tools. This gives agents access to the full HubSpot and Trello APIs through natural language, without being limited to specific pre-built operations.

  ## Improvements

  * **Nested code execution** — Agent scripts can now call other scripts, enabling multi-file project organization for complex automation
  * **Skills library for sub-agents** — Delegated sub-agents now have access to the skills library, improving their ability to handle specialized tasks
  * **Team admin usage reports** — Team admins can now view usage reports scoped to agents in their admin groups
  * **Code execution abort** — The stop button and timeouts now properly abort running code execution scripts
  * **Error boundary clarity** — The error recovery button now clearly describes what it does
  * **Image caching & Firefox fixes** — Improved image cache headers and fixed drag-and-drop overlay behavior in Firefox

  ## Bug Fixes

  * **Scheduled trigger editing** — Fixed an issue where agents could fail to edit documents when running from scheduled triggers
  * **Agent file migration** — Fixed files not being updated when moving an agent between workspaces
  * **Voice call stability** — Fixed the silence watchdog incorrectly firing during long agent responses
</Update>

<Update label="March 27, 2026" description="Send email from Outlook, voice recording fixes, modal UX improvements">
  ## Send Email from Outlook

  Agents with the Outlook integration can now **send emails directly** — not just create drafts. A new per-agent toggle controls whether the agent can send on your behalf, giving you full control over when to allow direct sending versus draft-only mode.

  ## Improvements

  * **Voice recording filenames** — Voice recordings now use the correct file extension based on your browser's audio format, fixing inconsistencies where uploaded and downloaded files could have mismatched extensions
  * **Modal UX** — The rename chat dialog now pre-fills and selects the existing name for quick editing, and all smaller dialogs are now vertically centered for a cleaner look

  ## Bug Fixes

  * **Usage page loading** — Fixed an issue where the Usage page could briefly flash incomplete content before all data had loaded
</Update>

<Update label="March 26, 2026" description="Leave team & workspace, Trello webhook reliability, relaxed email link validation">
  ## Leave Team & Workspace

  Members can now **leave a team or workspace** directly from the Members page. New "Leave team" and "Leave workspace" buttons with confirmation dialogs make it easy to remove yourself when you no longer need access. The last remaining admin is prevented from leaving to ensure teams and workspaces always have at least one admin.

  ## Improvements

  * **Trello webhook reliability** — Per-agent webhook URLs now correctly respond to HEAD probes from Trello (and similar services), fixing webhook registration failures. API credentials are also automatically trimmed of whitespace to prevent authentication errors
  * **Email link flexibility** — Agents can now include clickable links to any URL in outgoing emails, including links that require authentication. Recipients simply click through and log in as needed. Embedded resources like images and backgrounds are still validated
</Update>

<Update label="March 25, 2026" description="Sandbox capability for code execution, API Capabilities with OAuth support, platform links in emails, PDF tool enhancements">
  ## Sandbox Capability

  Agents can now create **isolated sandboxes** to run bash commands, Python scripts, and Node.js code in a secure environment. Tools include creating and destroying sandboxes, executing commands, and transferring files (upload, download, read, write, edit). Sandboxes automatically stop after inactivity.

  ## API Capabilities with OAuth Support

  HTTP APIs have been redesigned as **API Capabilities** with named authentication slots and per-agent secret binding. The same API configuration can now be shared across agents with different credentials. A new **OAuth 2.0 Client Credentials** flow supports providers like Shopify and HubSpot, with automatic token caching and refresh.

  ## Platform Links in Emails

  Agents can now include **clickable links to the Abundly platform** in outgoing emails — for example, linking a recipient to an agent chat page. The agent asks for confirmation before sending platform links, while embedded resources (images, backgrounds) remain blocked since email clients can't authenticate.

  ## Improvements

  * **PDF page info** — `readPdfPagesVisually` now returns which pages were skipped and the total page count, so agents can handle out-of-range pages gracefully
  * **Clearable search fields** — Search inputs in the customer filter and admin users page now include a clear button
  * **Star icon visibility** — Starred chat icons are now easier to see across both light and dark mode
  * **Sub-agent visual QA** — Delegated sub-agents can now use image inspection tools, improving visual quality checks in multi-agent workflows

  ## Bug Fixes

  * **Email attachment button** — Fixed the "Open" button on email attachments being invisible due to a color issue
  * **Voice webhook URLs** — Fixed voice call webhooks sometimes using an internal URL instead of the public agent URL
</Update>

<Update label="March 24, 2026" description="Full GitHub API tool, generic Slack API tool, agent chats page, tool call char counts, chat documents redesign, usage report improvements">
  ## Full GitHub API Access

  The GitHub integration now uses a **single generic tool** that covers both REST and GraphQL endpoints, replacing the previous set of 11 dedicated tools. Agents can access the full GitHub API — issues, labels, releases, workflow runs, search, org-level queries, and more — using natural language, without being limited to specific pre-built operations.

  ## Generic Slack API Tool

  A new **`slack_api_request` tool** lets agents call any Slack Web API method not already covered by dedicated tools — for example `pins.add`, `usergroups.list`, or `bookmarks.add`. Includes actionable error guidance for common failures like missing OAuth scopes.

  ## Agent Chats Page

  A new **Chats page** lists all conversations for an agent sorted by last activity. Features include starred chats at the top, search with sort options, multi-select with bulk delete, and per-chat actions (rename, star/unstar, delete). The sidebar now shows a "See all" link and caps the recent chats list at 10.

  ## Improvements

  * **Tool call character counts** — Collapsed tool calls now display total character count (input + output) with a subtle color-coded heat map, making it easy to spot expensive calls at a glance
  * **Chat documents panel** — The side panel now uses card-based entries with image thumbnails, matching the visual style of the documents and chats list pages
  * **Usage reports** — Trend queries now align to UTC day/hour boundaries for accurate period data; export defaults to the last month with group-by-team support and improved UI
  * **MCP admin docs column** — A new inline-editable Docs column in the MCP capabilities admin table lets admins set documentation links directly

  ## Bug Fixes

  * **Basic auth secrets** — HTTP API secrets using `basic_auth` injection are now automatically Base64-encoded at resolution time, so users can store plain `email:token` values without manual encoding
</Update>

<Update label="March 23, 2026" description="Mobile-friendly capabilities, Autodesk model tree & DWG conversion, iframe fullscreen Escape key, invitation flow fix">
  ## Mobile-Friendly Capability Cards

  On mobile devices, capability and MCP settings now **expand inline within the card** instead of opening a side panel. A chevron button toggles the settings section, keeping everything accessible on smaller screens. Desktop behavior is unchanged.

  ## Autodesk Capability Improvements

  Several enhancements to the Autodesk capability:

  * **Model tree tool** — A new `get_autodesk_model_tree` tool returns the element hierarchy of translated models (names, IDs, nesting) with optional depth truncation, complementing the existing properties tool for drill-down workflows
  * **DWG conversion formats** — `convert_autodesk_file` now supports DWG → PDF and DWG → DWF output in addition to DXF
  * **Flexible OSS uploads** — `upload_to_autodesk_oss` now accepts optional bucket and retention policy parameters
  * **Forge Viewer skill** — A new skill captures knowledge about Three.js overlays, Revit property conventions, and coordinate systems for agents working with Autodesk models

  ## Iframe Fullscreen Escape Key

  Pressing Escape while interacting with content inside an iframe now correctly **exits fullscreen mode**. Previously, cross-origin iframes trapped the keyboard event so the Escape key had no effect. Pointer lock for 3D navigation is also now supported in sandboxed iframes.

  ## Eval Validation Rules Include Skill Tools

  The eval tool picker for validation rules now includes **skill tools** like `get_skills`, matching what's actually available to agents at runtime.

  ## Bug Fixes

  * **Invitation flow** — New users with pending invitations no longer see a distracting "Create a new workspace" button
  * **Widget system prompt** — Whitespace-only system prompts now correctly fall back to the agent's default message
  * **Autodesk token refresh** — Fixed OAuth token auto-refresh and added `offline_access` scope for proper refresh token issuance
</Update>

<Update label="March 20, 2026" description="Autodesk capability, HTTP API auto-injection auth, email attachment saving, relative scheduled trigger times, React sandbox fixes">
  ## Autodesk Capability

  A new **Autodesk** capability lets agents interact with Autodesk services — upload files to OSS, convert DWG to DXF, translate models, extract properties, generate interactive 3D viewers, and browse ACC projects and folders. Supports both app-level (2-legged) and user-level (3-legged OAuth) authentication.

  ## HTTP API Auto-Injection Auth

  HTTP API secrets now support **automatic credential injection** with two new methods: GCP Service Account (OAuth2 token exchange) and JWT Bearer (RS256/ES256 signed tokens). Credentials are resolved and injected transparently — agents never see the raw secrets. Tokens are cached and refreshed automatically.

  ## Save Email Attachments

  Agents can now **save incoming email attachments** as permanent documents. All email attachments are available as documents that agents can read on demand and promote to their document library — matching how chat uploads work.

  ## Relative Time for Scheduled Triggers

  The one-off scheduled trigger tool now accepts **relative times** like `'now'`, `'+5m'`, `'+2h'`, and `'+1d'`. Times are resolved server-side when the tool call executes, so "schedule something now" works reliably even after a long chain of prior steps.

  ## Bug Fixes

  * **React sandbox globals** — Lucide icon exports named `Infinity` and `Map` no longer shadow JavaScript globals in React documents. A new `Lucide` namespace (e.g. `<Lucide.Camera />`) is also available
  * **Activity log details** — Expanding a log entry now correctly loads the full event data, fixing blank fields for email body and webhook payloads
  * **Voice call temperature** — Custom temperature settings are now correctly applied to voice calls
</Update>

<Update label="March 19, 2026" description="Configurable agent voice with ElevenLabs, multi-file app folders, token limit handoff summaries, binary file access in scripts, recursive folder ZIP downloads">
  ## Configurable Agent Voice with ElevenLabs

  Your agent's voice — used by the Speak button and Walk & Talk mode — is now **fully customizable**. A new voice picker in Agent Settings lets you choose from both OpenAI and ElevenLabs voices in a unified dropdown, with instant audio previews so you can hear each voice before selecting it.

  ## Multi-File App Folders

  Agent document folders can now act as **runnable apps**. Create a folder with an `index.html` (or `.jsx`/`.tsx`) entry file and the agent renders it as an interactive application — complete with multi-file asset support, a reload button, and fullscreen mode. A new **3D Visualizations skill** provides Three.js guidance for building 3D scenes.

  ## Token Limit Handoff Summary

  When a conversation approaches its context limit, an **inline warning bar** now appears above the chat input. Click "Start new chat with summary" and an LLM generates a structured handoff recap — capturing your goal, progress, pending tasks, and key decisions — so you can continue seamlessly in a fresh conversation.

  ## Binary File Access in Scripts

  The `code_execution` sandbox now includes **`fetch()` and `Buffer`** globals, letting agent scripts download their own document files (images, audio, PDFs) and perform binary conversions. This makes it possible to send files to external APIs directly from scripts.

  ## Recursive Folder ZIP Downloads

  Downloading a folder now **includes all nested subfolders** with the full directory hierarchy preserved in the ZIP file. Sibling folders with the same name are automatically deduplicated.

  ## Improvements

  * **Telnyx SIP voice enhancements** — Parallel tool execution, adaptive silence watchdog, per-language speaking instructions, and session pre-building for lower latency
  * **HTML document API access** — HTML documents in iframes can now use the same document API functions (query, insert, update, delete) previously limited to React apps
  * **File download extensions** — Downloading text-based documents like CSVs no longer appends incorrect file extensions
  * **Move documents from folders** — The "Move to folder" option now works correctly for documents already inside a folder

  ## Bug Fixes

  * **Immediate scheduled triggers** — One-off triggers scheduled for "now" or seconds in the future are no longer silently dropped
  * **Script trigger sync** — Converting a script-based trigger via "Sync from instructions" now correctly switches to the agent pipeline
</Update>

<Update label="March 18, 2026" description="Chat widget header color customization">
  ## Chat Widget Header Customization

  The embeddable chat widget now supports **custom header colors**. Use the new `headerColor` and `headerTextColor` embed config options to style the chat panel header to match your brand — for example, set a yellow header with dark text to blend seamlessly into your site design.
</Update>

<Update label="March 18, 2026" description="Configurable voice audio settings, Gemini image fallback, Redis-backed voice state, HTTP API admin access">
  ## Configurable Voice Audio Settings

  Phone call capabilities now include a full **audio settings panel** with control over speaking speed, noise reduction, turn detection mode (semantic vs silence-based), VAD responsiveness, silence duration thresholds, and transcription keywords. Advanced settings cover response temperature, transcription model, and VAD activation threshold.

  ## Gemini Image Generation Reliability

  Image generation now uses **automatic dual-model fallback** — when the primary model hits rate limits, it seamlessly switches to an alternative model with a separate quota pool. Both pools must be exhausted before you see an error. Surprise Me requests are also batched for better throughput.

  ## Improvements

  * **HTTP API admin access** — The HTTP API form modal now includes an admin access selector, so you can grant or restrict management access when creating or editing an API configuration
  * **Voice call stability** — Voice call state is now shared across server instances, fixing rare issues where calls could go silent when webhook events were routed to different servers
  * **Voice call tool tracking** — Tool calls made during phone conversations are now recorded, and a silence watchdog nudges the agent if it goes quiet without acting

  ## Bug Fixes

  * **Agent log page crash** — Fixed the activity log returning errors when entries contained large email payloads
  * **Missing phone call events** — Fixed phone calls not appearing in the event log when the post-call summary was empty
  * **Voice tool errors** — Fixed voice sessions hanging when a tool call failed — the error now flows back so the agent can inform the caller
</Update>

<Update label="March 17, 2026" description="Telnyx SIP voice calls, save chat documents permanently, progress indicators during tool streaming, widget mobile UX, higher output token limits">
  ## Telnyx SIP Voice Provider

  A new **Telnyx SIP voice provider** is now available alongside Twilio. Audio flows directly between the caller and OpenAI over SIP — no server-side audio proxying — for lower latency and simpler infrastructure. Supports inbound calls, outbound calls with two-leg bridging, and agent-initiated call transfers. Voice provider is now determined automatically by the phone number configuration, with support for workspace-level default phone numbers.

  ## Save Chat Documents as Agent Documents

  Documents created during a chat session can now be **saved permanently** to the agent's document library. A new "Save as agent document" option appears in the document panel dropdown for chat-scoped documents — one click promotes them from temporary to permanent.

  ## Progress Indicators During Tool Streaming

  When the agent is composing a large tool call — such as writing a full document or drafting an email — a **contextual progress bubble** now appears in the chat instead of just a spinner. This gives you immediate visual feedback that work is happening.

  ## Improvements

  * **Widget mobile experience** — The embeddable chat widget now properly handles mobile virtual keyboards, prevents background page scrolling, and avoids iOS auto-zoom on text input
  * **Higher output limits** — Agents running via triggers (email, scheduled tasks, webhooks) can now produce significantly larger outputs, fixing rare silent truncation issues on complex tasks
</Update>

<Update label="March 16, 2026" description="Agent export/import, chat-scoped documents replace artifacts, eval sandboxing, widget style customization, phone call instructions from documents">
  ## Agent Export & Import

  You can now **export and import agents** between environments via JSON files. Download an agent's full configuration — instructions, capabilities, documents, evals, and settings — and import it into another workspace or environment. The existing clone feature has been unified with this pipeline for consistency.

  ## Chat-Scoped Documents Replace Artifacts

  Artifacts have been replaced with **chat-scoped documents**. Agents now create visual content (HTML, React, SVG, Mermaid diagrams) as documents scoped to the current conversation. A new **Chat Documents panel** in the sidebar lets you browse all documents created during a chat session.

  ## Eval Agent Sandboxing

  Evals now run against a **temporary clone** of your agent by default, so document edits, instruction changes, and other mutations during eval runs don't affect the real agent or interfere with parallel runs. A per-eval toggle lets you disable sandboxing for read-only evals to skip cloning overhead.

  ## Eval UI Improvements

  * **Error/fail separation** — Errors are now visually distinct from fails with orange badges instead of being lumped together
  * **Run naming** — Name your eval runs via an inline pencil icon to easily identify them in history
  * **Abort runs** — Cancel running evals from both the results and history tabs

  ## Widget Style Customization

  The embeddable chat widget now supports **8 color override options** — customize bubble, input border, and submit button colors to match your brand. All defaults preserve the existing appearance, so no changes are needed for current embedders.

  ## Phone Call Instructions from Documents

  Phone call capabilities can now use an **agent document as their instructions source** instead of inline text. The voice session resolves the document content at call time. If the document is deleted, the capability automatically falls back to text instructions.

  ## Improvements

  * **Oversized tool output recovery** — When a tool response exceeds the output limit, it's now saved as a temporary document instead of being discarded. The agent gets a preview and can inspect the full result
  * **Email Reply-To support** — The send email tool now supports a Reply-To header, so email replies can go to a different address than the sender
  * **Scheduled Tasks relocated** — The Scheduled Tasks tab has moved from Settings to the Instructions page, sitting alongside the instructions it relates to
  * **Activity log fail reasons** — Failed executions now display a prominent error box with the specific reason the execution was evaluated as failed
  * **Disabled agents hidden** — Disabled agents no longer appear in the agent selector unless currently selected

  ## Bug Fixes

  * **Empty collection documents** — Creating a data collection document without providing initial data no longer shows an "Invalid JSON data" error
</Update>

<Update label="March 13, 2026" description="Automatic image scaling for faster vision processing, Gemini 3 multi-turn tool calling fix">
  ## Automatic Image Scaling

  Images sent to vision models (Claude, Gemini, GPT) are now **automatically scaled down** before uploading — matching each provider's internal target resolution. This reduces upload time and bandwidth without affecting quality, since the APIs would downscale them anyway. Images used as sources for image generation are also pre-scaled. Repeated images in a conversation are cached to avoid redundant downloads.

  ## Bug Fixes

  * **Gemini 3 multi-turn tool calling** — Fixed "missing thought signature" validation errors that could break multi-turn conversations using Gemini 3 models with function calls
  * **HTTP API secret display** — Expanding an HTTP API config row no longer briefly flashes a raw ID before the secret name loads
</Update>

<Update label="March 12, 2026" description="Slack file attachments, widget improvements, mobile responsive UI, agent-to-agent response fix">
  ## Slack File Attachments

  Agents can now **send file attachments** in Slack messages and DMs — including images (with automatic SVG-to-PNG conversion), PDFs, and other files. Attach files from agent documents, URLs, or raw SVG content directly in your Slack conversations.

  ## Smarter Slack User Matching

  Slack user lookup now matches on **display name, real name, email prefix, and username** — with partial match fallback and disambiguation when multiple users match. This makes it much more reliable to send DMs or mention users by name.

  ## Slack App Setup Improvements

  The Slack integration card now includes a **ready-to-use app manifest** you can copy-paste when creating your Slack app, plus a **"Test connection" button** to verify your token works before saving.

  ## Mobile Responsive UI

  The portal is now more **mobile-friendly**: tabs scroll horizontally instead of wrapping, document lists are compact with collapsible details, filters collapse into a toggle, and action buttons use icons only on small screens.

  ## Bug Fixes

  * **Agent-to-agent responses preserved** — Fixed a bug where text was silently discarded when an LLM response contained both text and tool calls, causing "No response" in agent logs
</Update>

<Update label="March 11, 2026" description="Embeddable chat widget, call transfer for voice calls, batch document edits, voice capability picker, scheduled trigger sync improvements">
  ## Embeddable Chat Widget

  A new **embeddable chat widget** lets you add an Abundly-powered chat bubble to any external website. The widget supports streaming responses, markdown rendering, and image display — with configurable allowed origins and daily credit limits per agent.

  ## Call Transfer for Voice Calls

  Agents can now **transfer live phone calls to a human** during Twilio voice sessions. Configure an escalation number in the phone call capability settings, and the agent will hand off the call when needed.

  ## Voice Call Capability Picker

  You can now choose **exactly which capabilities** are available during voice calls, instead of a simple on/off toggle. Select the specific tools your agent should have access to when on the phone.

  ## Batch Document Edits

  The `edit_document` tool now accepts **multiple edits in a single call**, making document updates faster and more reliable when agents need to make several changes at once.

  ## Improvements

  * **Agent timezone setting** — Agents now have a configurable timezone (auto-set from your browser) used for scheduling and time reasoning
  * **Stronger scheduled task sync** — Removing a scheduled task from instructions now reliably removes the corresponding trigger
  * **Subagent output limits** — Delegated tasks now have a higher output limit (200K) for more room during complex work

  ## Bug Fixes

  * **Voice capability toggles** — Fixed toggles doing nothing when the agent had never explicitly configured voice capabilities
  * **Voice session stability** — Fixed sessions dropping when routed to different server instances
</Update>

<Update label="March 10, 2026" description="Agent timezone setting, Steep analytics capability, Snowflake connector now available, per-agent Slack webhooks, Google Calendar for all customers">
  ## Agent Timezone Setting

  Agents now have a configurable **timezone** that's set automatically from the creating user's browser timezone. The agent uses this timezone when reasoning about times and scheduling triggers, so everything works in your local time rather than UTC. You can change the timezone anytime in the agent's Settings page.

  ## Steep Analytics Capability

  A new **Steep** capability lets agents browse and query metrics from the Steep analytics platform. Agents can explore available metrics, entities, and modules, and run queries with configurable time grains and date ranges.

  ## Snowflake Connector Now Available

  The **Snowflake** connector has been moved from hidden to **available** — it can now be enabled on any agent.

  ## Per-Agent Slack Webhook

  Each agent now has its own **dedicated Slack webhook URL**, so you can point each Slack app directly at its agent without relying on a shared global endpoint. The webhook URL is shown in the Slack integration settings for easy copy-paste.

  ## Google Calendar for All Customers

  The Google Calendar capability is now **available to all customers** — no longer requires admin access to enable.

  ## Improvements

  * **Subagent output limits** — Subagent tasks now have a higher output limit (200K), giving delegated work more room
  * **Safer Update Instructions default** — The Update Instructions capability now defaults to chat-only and editors-only when first enabled
  * **HTTP endpoint names show URLs** — Named HTTP endpoints now display both name and URL for clarity
  * **Cleaner agent selector** — Removed the redundant green/red status dot; disabled agents now show a ban icon overlay instead

  ## Bug Fixes

  * **HTTP secret admin controls** — Edit/disable/delete buttons now correctly hidden when user hasn't claimed admin on a secret
</Update>

<Update label="March 9, 2026" description="Google Calendar access control, private Slack channel fix">
  ## Google Calendar Access Control

  When connecting Google Calendar, you now choose **which calendars** the agent can access — both during OAuth setup and afterwards in your personal token settings. Calendar tools enforce these choices, so agents only see the calendars you've allowed.

  Event listings are also more compact, reducing output size when querying multiple calendars at once.

  ## Bug Fixes

  * **Private Slack channel messaging** — Fixed an issue where agents couldn't post to private Slack channels even when the bot was already a member
</Update>

<Update label="March 9, 2026" description="Usage data dashboard with charts and CSV export, configurable voice call language, improved voice call quality and reliability">
  ## Usage Data Dashboard

  The billing page now includes an interactive **usage data section** with a stacked bar chart showing credit consumption per agent or per team. Filter by date range, toggle individual agents or teams on and off, and switch between daily, weekly, and monthly views. A new CSV export modal lets you download usage data with different aggregation options (by agent, by team, or by time period).

  ## Voice Calls

  Voice call transcription language is now **configurable per capability** instead of being hardcoded to Swedish. Choose between Swedish, English, Norwegian, Danish, and Finnish in the phone call capability settings.

  ### Improvements

  * **Better voice call quality** — Upgraded speech detection to semantic VAD for smarter turn-taking, and improved the transcription model for better accuracy on phone audio
  * **Smoother voice call interruptions** — Interrupting the agent mid-sentence now works more reliably, with proper handling of buffered audio
  * **Cleaner voice call hangup** — The goodbye message now plays fully before the call disconnects, and call summaries are generated faster
  * **Stable usage chart colors** — Agent and team colors in usage charts no longer shift when toggling visibility
  * **Bookmarkable group state** — Expanded and collapsed groups on the agents page are now preserved in the URL, so you can bookmark or share a specific view
  * **Voice call summary narrated aloud** — Fixed an issue where the agent could narrate the full conversation summary on the phone after ending a call, instead of hanging up cleanly
</Update>

<Update label="March 6, 2026" description="Granular secrets access control, HTTP API configuration for agents, new Shared Assets sidebar section">
  ## Granular Secrets Access Control

  Secrets now support **fine-grained access control**. Scope secrets to specific groups or agents, claim admin rights as a Workspace or Team Admin, and see at a glance which agents use each secret. A reverse-lookup view highlights access mismatches so you can spot configuration issues early.

  ## HTTP API Configuration

  A new **HTTP API** management page lets you define external API endpoints with secure secret injection. Agents reference secrets using `[secret:id]` syntax, and the platform validates placement and enforces access rules automatically. Supported injection methods include bearer tokens, API key headers, basic auth, query parameters, and custom headers.

  The **HTTP capability card** now includes an API config selector, approval flow options, and an "Allow unrestricted requests" toggle — plus an inline Add API shortcut.
</Update>

<Update label="March 6, 2026" description="Smooth chat streaming, agent activity log tools, lazy-loaded skills, client-side profile image compression, repeat course move-outs">
  ## Smooth Chat Streaming

  Chat responses now appear with a **smooth character-by-character animation** instead of arriving in bursty chunks. The text flows at a steady pace with adaptive catchup when the display falls behind — creating a polished typewriter-like experience.

  ## Agent Activity Log

  Agents can now inspect their own trigger history and credit usage directly with new **Activity Log** tools. Enable the capability to give your agent self-service access to `read_activity_log`, `get_activity_log_entry`, and `get_activity_log_stats`. A **"Chat about this"** button on activity log entries lets you drill into any entry with your agent.

  Optionally enable cross-agent log access so agents can inspect connected agents' activity too.

  ## Lazy-Loaded Skills

  Agent prompts are now leaner. Detailed guidance for artifacts, documents, Slack formatting, and more has been moved into **on-demand skills** that agents load only when needed — reducing baseline prompt size and improving cache hit rates.

  ## Improvements

  * **Client-side profile image compression** — Profile images are now resized and compressed before uploading, making uploads instant and eliminating oversized-image errors
  * **Repeat course move-outs** — New option allowing members in ongoing courses to move out multiple times, e.g. to access new agents added over time
  * **Document editor stays open on save** — Saving a document in the maximized editor no longer unexpectedly closes the modal
</Update>

<Update label="March 5, 2026" description="Claude prompt caching for lower credit usage, structured JSON viewer for tool calls, memory and performance improvements, security hardening">
  ## Claude Prompt Caching

  Conversations using Claude models now benefit from **automatic prompt caching**. The system prompt is split into stable and volatile parts, allowing Claude to reuse cached content across turns — reducing credit usage on longer conversations and multi-step tool call sequences.

  ## Structured JSON Viewer for Tool Calls

  Tool call arguments and results now display in a **collapsible JSON tree** instead of raw text. This applies to the chat log, eval transcripts, and admin LLM logs — making it much easier to inspect complex tool interactions.

  ## Improvements

  * **Reduced memory usage** — Switching between agents now properly cleans up conversation state, preventing the browser from accumulating memory over time
  * **Fewer unnecessary re-renders** — The chat stream handler is now more stable, avoiding extra work on every incoming message

  ## Bug Fixes

  * **Multi-step tool calls with Claude** — Fixed an issue where cache control markers could exceed Anthropic's limits during long tool call sequences, causing errors
  * **Hot reload stability** — Fixed a "SocketManager not initialized" error that could appear during development hot reloads
</Update>

<Update label="March 4, 2026" description="Structured agent log events, workspace default daily credit limits, billing and error handling fixes">
  ## Structured Agent Log Events

  The agent activity log now displays trigger events in a **structured, readable layout** instead of raw JSON. Dedicated cards show key details for Slack, Microsoft Teams, Trello boards, GitHub PRs, SMS, MCP tool calls, and API endpoint events — each with the most relevant fields at a glance. A collapsible JSON viewer with expand/collapse all and copy to clipboard is available for full details.

  Event labels in the log list are now human-readable (e.g. "Slack message from John", "PR #42: Fix bug") making it easier to scan through activity.

  ## Workspace Default Daily Credit Limit

  Workspace admins can now set a **default daily credit limit** that automatically applies to all agents that don't have their own explicit limit. Agent-specific limits still take priority when set. This makes it easier to manage credit budgets across many agents without configuring each one individually.

  ## Bug Fixes

  * **Gemini cached token billing** — Fixed an issue where cached tokens were being double-counted, making cached Gemini requests cost more than uncached ones.
  * **Grok cached token billing** — Fixed a calculation error that could cause credits to be added instead of deducted when using Grok with token caching.
  * **Anthropic error handling** — Server errors and mid-stream failures from Anthropic now display a proper error message instead of raw JSON text in the chat.
</Update>

<Update label="March 3, 2026" description="Public audio toggle for Text to Speech, user avatar tooltips, credit calculation and caching fixes">
  ## Public Audio for Text to Speech

  The Text to Speech capability now has a **"Create public audio"** toggle. When enabled, agents can generate publicly accessible audio URLs — useful for embedding audio in emails and other external content. This works the same way as the existing "Create public images" toggle.

  ## User Avatar Tooltips

  Hovering over a user avatar in chat now shows a **tooltip with the user's name, email, and message timestamp**. This makes it easier to identify who said what in busy conversations.

  ## Microsoft Teams Integration

  Your agent can now appear in Microsoft Teams as its own named bot. Connect it via the Developer Portal and Azure, then @mention it in channels, group chats, or DMs to trigger it directly — similar to the Slack integration.

  ## Bug Fixes

  * **Credit calculation for GPT and Grok models** — Fixed an issue where cached tokens were being double-counted, resulting in slightly higher credit usage than expected for GPT and Grok model calls.
  * **Scheduled trigger sync** — Fixed a rare issue where scheduled triggers could incorrectly link to non-existent script documents.
  * **Improved caching for tool-heavy conversations** — Claude-based agents now cache more efficiently during multi-step tool call sequences, reducing credit usage for complex tasks.
</Update>

<Update label="March 2, 2026" description="MCP capability card fix, Slack mention trigger fix">
  ## Bug Fixes

  * **MCP capability cards in chat** — When an agent suggested an MCP capability in a chat message, the toggle card was always showing as disconnected regardless of the actual state. The cards now correctly reflect whether the capability is enabled or disabled, and toggling works as expected.
  * **Slack @mention triggers** — Fixed an issue where Slack @mention triggers were being incorrectly suppressed, preventing agents from responding to mentions.
</Update>

<Update label="February 28, 2026" description="Team monthly credit limits, bulk move agents between teams, model inheritance fix">
  ## Team Monthly Credit Limits

  Workspace admins can now set **monthly credit usage limits per team**. A new Team Limits tab on the Usage page lets admins configure budgets for each team. Team admins can view their team's usage and manage agent daily limits within their teams.

  Agents in a team that exceeds its monthly budget are blocked until the limit is raised or the month resets. Role-aware warnings appear in the top banner and chat input so everyone knows what's going on.

  ## Bulk Move Agents Between Teams

  Workspace admins can now move all agents from one team to another.

  ## Bug Fixes

  * **Incorrect model usage** — During the past few days, agents with no preferred model were silently using the wrong fallback model (Claude Sonnet 4.0) instead of the global default (currently Claude Sonnet 4.5). The model shown in the UI now always matches the model actually used by the agent.
</Update>

<Update label="February 27, 2026" description="Auto-collapsing document folders, improved voice recording with waveform visualization, mobile UI fixes, endpoint credit tracking">
  ## Auto-Collapsing Document Folders

  Agents with many documents no longer bloat the system prompt. Folders with more than 10 documents are now automatically collapsed — showing only the 3 most recent items. The agent can still browse and search all documents on demand.

  ## Improved Voice Recording

  The chat microphone now shows a **scrolling waveform** while recording, giving you visual feedback that your mic is working. If the microphone is muted or disconnected, you'll see an inline warning instead of silently sending empty audio. Specific error messages now appear for common mic issues (permission denied, no mic found, mic busy).

  ## Bug Fixes

  * **Mobile sidebar menus** — Dropdown menus no longer appear behind the sidebar overlay on mobile devices
  * **Walk & Talk on mobile** — Layout now fits properly on small screens with appropriately sized buttons
  * **Endpoint credit tracking** — MCP tool calls, HTTP API endpoints, and webhook triggers now correctly record credit consumption in the activity log
</Update>

<Update label="February 26, 2026" description="MCP disconnect button, token optimisation mode, credit tracking in activity log, default daily credit limits">
  ## MCP Disconnect Button

  OAuth and token-based MCP capabilities now have a **Disconnect** button in the advanced configuration section. Use it to disconnect and reconnect your MCP integrations without removing them.

  ## Token Optimisation Mode

  A new toggle reveals detailed credit consumption information per tool call and message — helping you understand exactly where your credits go and optimise accordingly.

  ## Credit Cost Tracking

  The agent activity log now shows credit costs for each entry, giving workspace admins full visibility into per-action credit consumption.

  ## Improvements

  * **Default daily credit limits** — New workspaces are now created with a sensible global default daily credit limit per agent
  * **Chat draft persistence** — Unsent messages and file uploads are preserved when navigating between chats, so you never lose work in progress
  * **Slimmer capability prompts** — Capability instructions sent to the model are now more compact, saving tokens on every request
  * **Conversation memory as % full** — The context window indicator now shows an intuitive percentage instead of raw token counts
</Update>

<Update label="February 25, 2026" description="React app 'Fix this' button, multi-image Gemini editing, default thinking preferences, conversation memory % indicator">
  ## React App 'Fix this' Button

  When a React app fails to render — whether from a transpile error or a runtime crash — a **'Fix this'** button now appears alongside the error message. Clicking it sends the error details straight to the chat so the agent can automatically attempt a fix.

  ## Multi-Image Gemini Editing

  The Gemini image generation tool now accepts multiple source images as references when editing. Pass several reference images to guide more complex image transformations.

  ## Default Thinking Preferences

  New agents are now created with thinking **on** by default. The chat thinking toggle is also visible immediately on the new-chat screen — before the first message — so you can adjust it up front.

  ## Improvements

  * **Conversation memory indicator** — The context indicator now shows an intuitive "% full" instead of raw token counts, with color-coded warning levels
  * **Faster capability activation** — Triggers, agent-to-agent calls, and endpoint handlers now pre-activate all capabilities instantly, removing unnecessary startup delays in automated contexts
</Update>

<Update label="February 24, 2026" description="Script execution on scheduled triggers, webhooks, HTTP API & MCP, smoother chat scrolling, bug fixes">
  ## Script Execution on Scheduled Triggers

  Scheduled triggers can now run scripts directly — bypassing the full agent/LLM pipeline for fast, credit-free execution. Scripts have access to `callTool()` and `askAgent()`, making them ideal for recurring automations that don't need an LLM in the loop.

  * **Webhooks, HTTP API & MCP too** — Script execution also extends to webhooks, HTTP API endpoints, and MCP tools, so any integration type can now be backed by a script
  * **Full tool access** — Scripts running on any trigger type can call tools and delegate to other agents

  ## Improvements

  * **Smoother chat scrolling** — Auto-scroll during streaming is now less aggressive. Scrolling up immediately pauses auto-scroll so you can read in peace; it re-enables when you scroll back to the bottom
  * **Flexible HTTP request bodies** — The `http_request` tool now accepts any valid JSON value (string, array, number) as the request body, not just objects — fixing issues with raw file uploads
  * **MCP capabilities in delegated tasks** — Sub-agents created via `delegateTask` now correctly receive MCP-backed capabilities and user-added MCP servers

  ## Bug Fixes

  * **Slack double triggers** — Fixed an issue where @mentioning the bot in a channel would trigger the agent twice
  * **Parallel code execution** — Fixed a bug where concurrent `code_execution` calls could silently corrupt each other's output
</Update>

<Update label="February 23, 2026" description="Unified LLM preferences, BigQuery for all, user profile images, chat-scoped documents, run-now for scheduled tasks">
  ## Unified LLM Preferences

  You can now control which LLM model and thinking settings are used in different contexts — chat, scheduled tasks, email triggers, Slack, and more. This gives advanced users fine-grained control over model selection per context.

  * **Per-context model selection** — Choose different models for chat vs. scheduled tasks vs. triggers
  * **Thinking toggle per context** — Enable or disable extended thinking independently for each context
  * **"Run now" button** — Recurring scheduled tasks now have a "Run now" button to trigger them on demand
  * **Activity log** — Model and thinking settings are now recorded in activity log entries

  ## BigQuery for Everyone

  The BigQuery capability is now available to all workspaces. You can add it to your agents and configure GCP service account credentials directly — no admin enablement required.

  ## User Profile Images

  Users can now set a profile image that appears across the platform:

  * **Upload or AI-generate** — Upload your own photo or generate one with AI
  * **SSO fallback** — If no profile image is set, your SSO provider avatar is shown automatically in all views

  ## Chat-Scoped Documents

  Agents can now create documents scoped to the current conversation. These documents are automatically cleaned up when the chat is deleted — useful for temporary working files.

  ## Improvements

  * **Smarter agent list filters** — Search bar and toggle filters only appear when relevant, keeping the page clean
  * **Better tool call logging** — GPT and Grok tool calls now appear in the Activity Log, matching Claude and Gemini
</Update>

<Update label="February 20, 2026" description="Graph database capability, LLM model awareness, Workspace/Team rename, diary deep-linking">
  ## Graph Database Capability

  Agents can now store and query entities and their relationships using a graph database (FalkorDB). This enables powerful knowledge modeling with nodes, relationships, and property queries.

  * **7 new tools** — Create/find/delete nodes, create relationships, find neighbors, view schema, and run raw OpenCypher queries
  * **Enable per agent** — Available as a new capability in agent settings

  ## Agents Aware of Available Models

  Agents now know which LLM models are available to them, shown directly in their system prompt. Eval tools validate model selection against available models.

  ## Workspace & Team Terminology

  User-facing terminology has been updated for clarity:

  * "Customer" / "Team" is now called **Workspace**
  * "Group" is now called **Team**

  ## Improvements

  * **Fresh new-chat page** — Opening an agent now starts with a clean new-chat page instead of resuming the last conversation
  * **Auto-select agent name** — The Create New Agent page now focuses and selects the name field, so you can start typing immediately
  * **Diary date deep-linking** — The activity diary tab now supports `?date=YYYY-MM-DD` in the URL for linking to a specific date
  * **Credit limit icon** — Agent cards now show a compact dollar sign icon instead of a text badge for credit limits
</Update>

<Update label="February 19, 2026" description="Smarter Slack channel handling, consistent user avatars, block alert default fix">
  ## Smarter Slack Channel Handling

  Slack integration improvements for agents that interact with Slack:

  * **Archived channels excluded** — Archived channels no longer appear in channel lists, reducing clutter
  * **Faster lookups** — Channel and user resolution is now paginated, improving performance for workspaces with many channels
  * **Abort support** — Long-running Slack operations can now be cancelled

  ## Bug Fixes

  * **Block alert defaults** — Legacy agents created before the block alert feature no longer have alerts accidentally enabled; alerts now default to off until explicitly turned on
  * **Consistent avatars** — Avatars across the sidebar, chat, and members page now consistently show profile images with proper initials fallback
</Update>

<Update label="February 18, 2026 (2)" description="Download folders as ZIP, paginated activity log, smarter agent list filters, React artifact fix, data document crash fix">
  ## Download Folders as ZIP

  You can now download all documents in a folder as a single ZIP file.

  * **"Download all" button** — Available on the documents page to download everything in the current view
  * **Folder context menu** — Right-click a folder and choose "Download folder as ZIP"
  * **Smart naming** — Files in the ZIP use the document name with auto-added extensions; duplicates get a counter suffix

  Also, downloading individual files now correctly uses the document name instead of the internal storage path, and file uploads no longer fail entirely when transcription fails.

  ## Bug Fixes

  * **React artifact rendering** — Fixed an issue where artifacts with helper functions could render the wrong component, showing "N/A" instead of the actual UI
  * **Data document editing** — Editing invalid JSON in a data document no longer crashes the page; validation errors are shown inline and Save is disabled until fixed
  * **Admin user search** — Rapid filter changes no longer cause stale data to flash in the results
</Update>

<Update label="February 18, 2026" description="Disable/enable agents, block alert notifications, paste profile pictures, smarter tool output limits, rich editor improvements">
  ## Disable/Enable Agents

  Agents can now be disabled instead of deleted when no longer needed. This is like archiving an agent.

  * **Hidden by default** — Disabled agents disappear from lists and sidebar, with a "Show disabled" checkbox to reveal them
  * **Quick toggle** — Enable or disable from the agent card gear menu
  * **Clear communication** — Other agents see disabled agents marked as "(currently disabled)" and get a clear error if they try to call one
  * **Sync scheduled tasks** — A new "Sync from instructions" button lets you re-parse scheduled tasks if instructions were edited while the agent was disabled

  ## Block Alert Notifications

  Admins can now configure email notifications for when agent trigger execution is blocked during planning or security validation.

  * **Per-agent settings** — Choose who gets notified and what content is included
  * **Covers both block types** — Alerts for plan blocks and security-check blocks

  ## Paste Images for Agent Profile Pictures

  The agent profile picture dialog now supports pasting images from clipboard (Ctrl/Cmd+V), in addition to file upload and drag-and-drop.

  ## Smarter Tool Output Limits

  Agents are less likely to be bogged down by very large tool call responses, for example an API call with a very long response.

  * When tool output is very long, agents don't receive the data. Instead agents receive tailored advice with working code examples for filtering, compressing, or saving data
  * Code execution sandbox now includes lodash, JSONPath, cheerio, and Papa Parse for handling large responses

  ## Improvements

  * **Rich Editor handles more markdown** — Documents with links, tables, images, and code blocks no longer cause parsing failures in the Rich Editor
</Update>

<Update label="February 17, 2026" description="BigQuery integration, download documents from URLs, sleep in code execution">
  ## BigQuery Integration

  Agents can now connect to Google BigQuery to explore and query datasets directly.

  * **Dataset exploration** — List datasets, tables, and view table schemas
  * **Query execution** — Run SELECT queries with built-in security controls (read-only enforcement, 30-second timeout, 1GB billing limit)
  * **Cost awareness** — Query results include bytes processed so you can monitor costs
  * **Team credentials** — Uses GCP service account keys managed through the existing secrets system

  ## Download Documents from URLs

  Agents can now download files from public URLs and save them as agent documents using the new `download_document` tool. Text files are stored inline; binary files go to cloud storage with automatic transcription.

  Additionally, `get_media_content` now supports viewing external images directly via URL without downloading them first.

  ## Sleep in Code Execution

  Agents can now use `await sleep(ms)` in code execution scripts, enabling polling patterns — for example, triggering an async API job and waiting for the result.
</Update>

<Update label="February 16, 2026" description="MCP server support, webhooks, HTTP API endpoints, enterprise bootstrap, Unicode filenames">
  ## Expose Agents as MCP Servers

  Agents can now be exposed as MCP (Model Context Protocol) servers, allowing external AI applications like Claude Desktop, Cursor, and n8n to connect and call agent-defined tools.

  * **Tool configuration** — Define which tools your agent exposes via MCP, each with its own prompt and optional model override
  * **Shared API keys** — Authentication uses the same API key system as the HTTP API
  * **Activity logging** — All MCP tool calls are tracked in the activity log

  ## Agent-Specific Webhooks

  Agents can now receive unauthenticated webhook callbacks from external services like Stripe, GitHub, and Trello.

  * **Per-integration endpoints** — Each integration gets its own URL at `/agents/:agentId/webhook/:integration`
  * **Fire-and-forget** — Webhooks respond immediately with 200 and process asynchronously
  * **Custom prompts** — Configure how your agent handles each integration's callbacks

  ## HTTP API Endpoints

  The HTTP API has been upgraded from a single generic endpoint to defined, per-path endpoints with individual prompts and model overrides.

  * **Defined endpoints** — Only configured paths respond; undefined paths return 404
  * **Per-endpoint prompts** — Each endpoint can have its own instruction prompt
  * **Model overrides** — Choose different models per endpoint for cost or capability tuning

  ## Unified API Access Settings

  API Keys, HTTP API, MCP Server, and Webhooks are now managed in a single "API Access" settings tab with a cleaner collapsible card layout.

  ## Improvements

  * **Unicode filenames preserved** — Document downloads now keep international characters (åäö, etc.) in filenames instead of replacing them with underscores
  * **User activity weekly graph fixed** — Weekly and monthly views now correctly group data into proper time buckets
</Update>

<Update label="February 13, 2026" description="Agent-to-agent evals, eval run tracking, clickable agent names, smarter scheduled runs">
  ## Agent-to-Agent Calls in Evals

  Evals can now test multi-agent workflows. When an eval triggers `call_other_agent`, the sub-agent executes with its own tools and configured model, and the full conversation is captured as a collapsible nested transcript in the eval viewer.

  * **Recursive support** — Agent chains are supported up to 3 levels deep
  * **Per-agent tool overrides** — Configure fake/exclude overrides for each sub-agent independently
  * **Run progress tracking** — Active eval runs now show persistent status indicators with elapsed time in the eval tab

  ## Eval Tools UX Improvements

  * **Bulk actions** — "Include All" / "Exclude All" buttons for quick tool override configuration
  * **Status badges** — Tools tab shows included (green) and faked (orange) counts at a glance
  * **Renamed tab** — "Trigger" tab renamed to "Test Prompt" for clarity

  ## Clickable Agent Names

  Agent names across the management page, usage limit cards, admin agent list, billing transactions, and feedback page are now clickable links to the respective agent pages.
</Update>

<Update label="February 12, 2026 (2)" description="Clickable links in artifacts, mobile agent cards, improved chat error recovery, new scheduling system, support access admins">
  ## Clickable Links & Downloads in Artifacts

  Links and file downloads now work inside rendered HTML documents and React components created by agents. Previously, the sandbox policy blocked all navigation.

  * **External links** — Links open in new browser tabs as expected
  * **File downloads** — Download buttons and links work inside artifacts

  ## Mobile-Responsive Agent Cards

  Agent cards on the management page now adapt to mobile screens with a collapsible info panel replacing desktop tooltips.

  * **Expandable details** — Tap the chevron to reveal agent name, description, access level, paused status, tags, and credit usage
  * **Smarter tooltips** — On desktop, the info icon only appears when the agent name is truncated or has extra metadata

  ## Better Chat Error Recovery

  Improved messaging when conversations hit token limits or experience service overloads.

  * **Clearer actions** — Choose between "Keep it, I'll guide you" or "Try again, shorter" instead of seeing technical error messages
  * **Reliable error states** — Overload failures now consistently end the loading spinner and show the error

  ## Support Access Members as Agent Admins

  Support access members can now be added as agent admins directly from the agent management page, with a visual "support" tag for easy identification.
</Update>

<Update label="February 12, 2026" description="Public image generation for emails, email URL validation">
  ## Public Image Generation for Emails

  Agents can now generate images with publicly accessible URLs, making them suitable for embedding directly in outgoing emails.

  * **New `makePublic` option** — Image generation tools (Gemini and OpenAI) now support publishing generated images as public documents
  * **Capability setting** — Controlled via a new "Create public images" setting in agent configuration

  ## Email URL Validation

  The send email tool now validates all URLs in HTML email content before sending, preventing recipients from receiving emails with broken links.

  * **Broken link prevention** — Rejects emails containing malformed or relative URLs
  * **Unpublished document warnings** — Warns agents when shared file URLs reference documents that aren't actually published
  * **Stricter attachment URLs** — Attachment URLs now require absolute HTTP/HTTPS protocol
</Update>

<Update label="February 11, 2026" description="Smarter editing tools, eval streaming, redesigned agent cards">
  ## Smarter Editing Tools for Agents

  Agents now have more precise editing capabilities with a new unified approach:

  * **Partial instruction editing** — Agents can now make targeted edits to their instructions using exact string matching, instead of rewriting the entire instruction each time
  * **Unified document editing** — Three separate document editing tools have been consolidated into a single, more intuitive tool that mirrors the editing pattern used by leading coding agents

  ## Real-Time Eval Streaming

  Eval results now stream in real-time as they complete, instead of waiting for the entire run to finish.

  * **Live progress** — Results appear progressively with a progress bar and spinners for pending items
  * **Full grid view** — The complete eval × model grid is shown from the start so you can monitor progress at a glance
  * **Higher throughput** — Concurrency doubled for faster eval runs

  ## Redesigned Agent Cards

  Agent cards on the management page have been redesigned with a compact horizontal layout featuring avatars, overflow-aware capability tags, and icon indicators for status and permissions.

  ## Improvements

  * **Groups no longer blocked by user limits** — Groups availability is now driven solely by the groups setting, no longer incorrectly restricted by the subscription user limit on Start/Pro tiers
  * **More reliable image generation** — Gemini image generation now includes automatic retry with backoff for rate limits

  ## Bug Fixes

  * Fixed eval tool calls crashing the service in certain configurations
  * Fixed JSON parsing errors when LLM responses contained curly braces in explanatory text
</Update>

<Update label="February 10, 2026" description="Fullscreen docs, CAD extraction, security isolation, eval fixes">
  ## Fullscreen Document Viewing

  Documents can now be expanded from the side panel into a fullscreen overlay. This makes it easier to work with large documents, dynamic React UIs, and other rich content without leaving the chat.

  ## CAD File Data Extraction

  Agents can now extract geometry (polygons and lines) from DWG and DXF CAD files, saving the result as a GeoJSON document for further analysis or visualization.

  ## User Content Security Isolation

  User-generated content such as dynamic React UIs and HTML documents is now served from a separate domain, isolating untrusted content from the main app via cross-origin sandboxing.

  ## Improvements

  * **Tab header wrapping** — Tab headers with longer labels now wrap gracefully on narrow viewports instead of getting clipped
  * **External stylesheets in iframes** — User-generated HTML in iframes can now load external CSS and fonts (e.g. for leaflet maps)
  * **Stricter tool parameter validation** — Tool calls with unknown parameter names now return a validation error instead of silently dropping the parameter

  ## Bug Fixes

  * Fixed evals not including the agent's start message and user timestamp in chat mode, which could cause different behavior compared to real chats
  * Fixed agent cloning not copying evals and eval settings
</Update>

<Update label="February 9, 2026" description="Speech-to-text editing, credit improvements, eval credits">
  ## Speech-to-Text for Message Editing

  You can now use the microphone button when editing a previously sent message, just like in the main chat input. Dictate corrections or additions hands-free.

  ## Improved Credit & Limit Warnings

  Credit and limit warning messages are now clearer and more actionable:

  * **Better wording** — Messages reworded for clarity from the agent's perspective
  * **Admin actions** — Admins see direct links to edit limits or reset usage right in the warning banner
  * **Role-aware guidance** — Different guidance shown based on whether you're an admin or regular user

  ## Eval Credits Breakdown

  Eval results now show test and grader credits separately instead of a combined total, making it easier to understand where credits are being spent.

  * **Run-level view** — Test credits shown primarily, with grader cost as a side note
  * **Detail modal** — Separate lines for test and grader credits

  ## Improvements

  * **Microphone while waiting** — The mic button is now available while waiting for an agent response, so you can start dictating immediately
  * **Agent-to-agent activity logging** — Communication between agents is now tracked in the activity log with a dedicated UI widget
  * **Document visibility on creation** — Agents can set document visibility when creating documents

  ## Bug Fixes

  * Fixed cross-tab socket issue where having multiple agents open in different browser tabs could cause errors
  * Fixed signup copy text on email flow
</Update>

<Update label="February 5, 2026" description="Google Drive and Evals">
  ## Google Drive Enhancements

  Google Drive integration now supports a wider range of file types and improved file management.

  * **Expanded file support** — Agents can now read PDFs, Word documents (.docx), Excel spreadsheets (.xlsx), PowerPoint presentations (.pptx), and text files from Google Drive. Previously limited to Google Workspace files (Docs, Sheets, Slides).
  * **Multi-file selection** — Share multiple files with your agent at once using the file picker, instead of adding them one at a time.

  ## Evals Improvements

  Enhanced testing capabilities for evaluating agent performance across different scenarios.

  * **Trigger execution mode** — Evals can now run in "trigger" mode to test how agents behave when triggered by events (like incoming emails or scheduled tasks), not just chat interactions. Defaults to "chat" mode for backward compatibility.
  * **Per-model thinking configuration** — Configure thinking/reasoning independently for each model in an eval, giving you finer control over test scenarios. Previously this was a global on/off setting.
</Update>

<Update label="February 3, 2026" description="Feedback, diary toggle, credit widget">
  ## User Feedback

  You can now submit feedback directly from the user menu. Click your profile, select "Give feedback", and share your thoughts with optional sentiment tags (positive, negative, question, or alert). Your feedback includes context about where you are in the app to help us understand and act on it.

  ## Diary Toggle

  You can now disable diary writing per agent to reduce LLM costs. Find the toggle in the Diary view — disabled agents won't write new diary entries but existing entries remain visible.

  ## Credit Usage Widget

  A new widget in the sidebar shows your agent's daily credit usage at a glance. When a daily limit is configured, you'll see a progress bar. The widget updates automatically after each chat response.

  ## Agent Cloning Improvements

  When cloning an agent to another workspace, you now see:

  * Available teams in the target workspace
  * Warnings about capabilities that will be removed if not available
  * Clear errors if the clone isn't allowed

  ## Course Move-out

  When moving from a course or hackathon workspace to your own workspace, you can now select which agents to bring along. Agents you don't select stay paused on the source workspace.
</Update>

<Update label="January 29, 2026" description="Attack detection, evals, and security alerts">
  ## Configurable Security Alert Recipients

  Admins can now control who receives attack detection notifications. Choose from three recipient modes: all admins, admins plus additional emails, or custom emails only. You can also optionally include the blocked content in alert emails for debugging.

  * **Recipient modes** — All admins (default), admins + additional, or custom only
  * **Include blocked content** — Optional toggle to see what was blocked

  ## Smarter Attack Detection

  Attack detection now focuses specifically on actual LLM manipulation attempts like prompt injection and jailbreaks. This reduces false positives for legitimate messages.

  * **No longer flagged** — Angry customers, rude tone, sensitive topics (medications, personal data), legal threats, urgent demands
  * **Still blocked** — Prompt injection, jailbreak attempts, fake system messages, hidden instructions

  <Tip>This is especially important if you're routing a support inbox to your agent — customers don't know they're talking to an AI and may express frustration or discuss sensitive matters.</Tip>

  ## Conversation Weight Indicator

  New visual indicator in the chat bar showing how much context has accumulated in your conversation. As context grows, each message costs more credits — this helps you know when it's time to start a fresh chat.

  * **Visual ring** — Circular progress that fills as context grows
  * **Color coding** — Gray → yellow → orange → red as cost increases
  * **Quick help** — Click the indicator to open optimization tips

  ## Enhanced Eval Testing

  The eval system now supports more powerful testing scenarios with tool-level control and flexible grading options.

  * **Fake tools** — Simulate external service responses during test runs
  * **Validation rules** — Assert that specific tools were (or were not) called
  * **Grader scope** — Choose whether the grader sees the full transcript or just the final response
  * **Smoke test mode** — Skip the grader entirely and pass if the eval runs without errors

  ## MCP + Gemini Compatibility

  MCP tools now work properly with Gemini models. Tool schemas are automatically sanitized to remove JSON Schema features that Gemini's function calling API doesn't support.
</Update>

<Update label="January 23, 2026" description="Security & admin controls">
  ## Attack Detection for Email and SMS Triggers

  Incoming emails and SMS messages are now automatically scanned for potential attacks before your agent processes them. The system detects prompt injection attempts, social engineering, data exfiltration, jailbreak attempts, and reconnaissance attacks.

  * **Blocked attacks** — Suspicious messages are blocked and logged in the activity log with a clear indicator
  * **Admin alerts** — Workspace admins receive email notifications when attacks are detected and blocked

  ## Update Instructions Restrictions

  Admins can now control when and who can use the "Update Instructions" capability, adding an extra layer of safety for production agents.

  * **Chat only** — Restrict instruction updates to chat conversations only (blocks autonomous triggers from modifying instructions)
  * **Editors only** — When chat-only is enabled, further restrict to users with Edit access or higher

  ## Batch User Invites

  Add multiple users to a team simultaneously instead of one at a time, making onboarding faster.

  ## Agent Evals Improvements

  Several enhancements to the evals experience:

  * **Run management** — Delete entire eval runs, duplicate existing evals
  * **Grader configuration** — Configure and display the grader model used for evaluation
  * **Transcript view** — Toggle between friendly and raw JSON transcript views
  * **Navigation** — Clickable result badges in the list view for quick access
</Update>

<Update label="January 21, 2026" description="Agent Evals & Template Creation">
  ## Agent Evals

  New comprehensive evaluation system for testing and validating agent behavior. Define test cases, run them across multiple models, and track results over time.

  * **New capability** — Enable the Evals capability to let your agent manage its own test cases
  * **Test definitions** — Create evals with trigger prompts and grader prompts
  * **Grading options** — Support for pass/fail and 1-10 rating scales
  * **Multi-model testing** — Run the same eval across different models to compare performance
  * **Results grid** — Visual comparison of results across models
  * **Credit tracking** — See credits used for both test and grader phases
  * **Agent tools** — When enabled, agents get access to `list_evals`, `create_eval`, `edit_eval`, and `run_evals` tools

  <Info>
    Learn more about testing your agents in the [Evals documentation](/features/evals).
  </Info>

  ## Template-based agent creation

  Create agents from external links with pre-filled prompts — perfect for use case pages and onboarding flows.

  * **URL parameters** — Support for `createAgent=true`, `name=`, and `prompt=` parameters
  * **Auth flow preservation** — Parameters preserved through signup and workspace selection
  * **Pre-filled chat** — Chat input automatically populated with the template prompt
  * **Clone support** — Also works when cloning agents with predefined name and prompt

  ## Agent management sorting

  The agent management table now supports column sorting:

  * **Sortable columns** — Click any column header to sort by name, status, privacy, usage, or admins
  * **Visual indicators** — Sort direction shown with arrow icons
  * **Alphabetical admin sorting** — Admin column sorts by name, not count

  ## Improvements

  * **Simplified agent creation** — Removed description field, pre-fills name with "New Agent"
  * **Better onboarding** — New users auto-redirected to workspace creation when that's the only option
  * **Login flow** — Users now return to their intended page after login (callbackUrl support)

  ## Bug fixes

  * Fixed inline code blocks rendering on separate lines instead of inline with text
  * Fixed login redirect incorrectly sending members to admin page
  * Handled deleted users gracefully in chat (shows neutral avatar instead of error)
</Update>

<Update label="January 14, 2026" description="Create Workspace & Document Bookmarks">
  ## Create new workspace

  Existing users can now create their own workspace directly from the portal:

  * **Trial support** — New users get a free trial period
  * **Immediate payment** — Users with existing billing history can subscribe directly
  * **Simple flow** — Just name your workspace and you're ready to go

  ## Document bookmarking

  Agent documents can now be bookmarked via URL, making it easier to share and navigate directly to specific documents.

  ## Capability documentation links

  Capabilities now include direct links to documentation, making it easier to learn how to use each integration.

  ## Improvements

  * **MCP server authentication** — Fixed modal blocking and improved OAuth validation
  * **Welcome page** — Refreshed design with resource link cards
  * **Profile pictures** — OAuth profile pictures now sync on every sign-in
</Update>

<Update label="January 9, 2026" description="Native Image Support & Memory UI">
  ## Enhanced data isolation

  Major architectural improvement for enterprise-grade security: the web portal no longer has direct database access. All data operations now go through a dedicated service layer, enabling physical separation between frontend and data storage for enhanced data protection and compliance.

  This is one step in an ongoing process to enable full data isolation for enterprise customers.

  ## Native image support

  Agents can now view images directly using Claude, GPT, and Gemini's vision capabilities:

  * Use the new `get_media_content` tool to analyze images natively
  * Supports JPEG, PNG, GIF, WebP, and HEIC/HEIF (iOS photos)
  * 5MB size limit per image for Claude
  * Agents automatically use this when working with image documents

  ## Improved chat file uploads

  File uploads in chat are now faster and smarter:

  * Files are stored as chat-scoped documents with on-demand transcription
  * Manual "Transcribe" button available in the document panel
  * Agents can promote chat documents to permanent agent documents
  * Chat documents are automatically cleaned up when the chat is deleted

  ## Memory management UI

  Browse and manage your agent's memories from a new dedicated interface:

  * Search memories using semantic similarity
  * Create, edit, and delete memories directly
  * Adjustable similarity threshold for better recall
  * Access from the Memory capability card

  ## Text-to-speech providers

  Choose your preferred voice provider for agent speech:

  | Provider       | Models                                |
  | -------------- | ------------------------------------- |
  | **ElevenLabs** | Multilingual v2, Turbo v2.5 (default) |
  | **OpenAI**     | Standard TTS voices                   |

  ## MCP server improvements

  MCP capabilities now support auto-connect and better agent understanding through LLM help text descriptions.
</Update>

<Update label="December 18, 2025" description="Document Controls & Delegate Tasks">
  ## Document visibility badge

  See and change document visibility with a single click:

  * New badge shows visibility level at a glance
  * Click to quickly change visibility via popover
  * Works in both list view and document panel

  ## Enhanced delegate task tool

  More control when delegating tasks to sub-agents:

  * Specify which LLM model to use
  * Set temperature and enable thinking mode
  * Better capability validation

  ## Email & SMS whitelisting

  Filter incoming messages before they reach your agent:

  * Configure sender whitelists for emails
  * Configure phone number whitelists for SMS
  * Filtering happens before LLM processing

  ## Audio transcription recovery

  New recovery modal when transcription fails — retry or dismiss failed transcriptions easily.

  ## Large data documents

  Data documents with thousands of items are now handled gracefully with automatic truncation.

  ## Folder conversations

  The "Talk about this" button now works for folders, letting you discuss entire folder contents with your agent.
</Update>

<Update label="December 12, 2025" description="MCP Server Integration">
  ## MCP server integration

  Connect your agents to external tools using Model Context Protocol (MCP) servers:

  * Access thousands of external tools through MCP
  * OAuth authentication support with automatic token refresh
  * MCP capabilities appear alongside regular capabilities
  * Agents can suggest MCP capabilities during conversations

  <Info>
    Learn more about connecting MCP servers in the [MCP integration guide](/integrations/mcp).
  </Info>
</Update>

<Update label="December 9, 2025" description="Rich Documents & Chat Sorting">
  ## Rich markdown documents

  Markdown documents now support more content types:

  * Inline SVG images and HTML elements
  * Mermaid diagrams embedded directly in code blocks
  * Better version selection for diagrams

  ## Chat improvements

  * **Activity sorting** — chats now sorted by most recent activity instead of creation date
  * **Copy messages** — copy any agent message to clipboard with one click (appears on hover)
</Update>

<Update label="December 4, 2025" description="New UI & Image Generation">
  ## New UI design

  Complete design refresh with improved layout and visual hierarchy:

  * Revamped expandable panels and drag-and-drop functionality
  * Full mobile responsiveness — the entire platform now works beautifully on phones and tablets
  * Refined color scheme and improved readability
  * Better modals and tooltip behavior

  ## AI image generation

  Gemini image generation is now the default provider with better quality and faster results:

  * Unified image generation and editing
  * Provider selection (Gemini vs OpenAI) with advanced options
  * Support for aspect ratio and image size parameters
  * Download button for generated images

  ## Visual PDF reading

  New tool extracts text from complex PDFs using AI vision:

  * Handles PDFs with images, tables, charts, and complex layouts
  * Extracts content that standard text extraction misses

  ## Publish original files

  When publishing documents, you can now include the original file alongside the published content.
</Update>

<Update label="November 27, 2025" description="Data Documents & Admin Tools">
  ## Data document import/export

  You can now import and export data documents with full control over how data is handled:

  * **Append or replace** — choose whether to add to existing data or start fresh
  * **Duplicate detection** — automatic ID conflict detection with option to replace with new IDs
  * **Auto-validation** — JSON is automatically cleaned and validated on import

  <Tip>
    When adding items via the `insertDocumentData` tool, you no longer need to manually assign IDs — they're generated automatically.
  </Tip>

  ## Data document versioning

  Data document changes are now saved as versions, just like regular agent documents. View history, compare versions, and track changes over time.

  ## Admin agent overview

  Workspace and team admins now have a dedicated overview page showing all agents at a glance — view capabilities, ownership, and team membership in one place.
</Update>

<Update label="November 19, 2025" description="Gemini 3 & Team Management">
  ## Gemini 3 is now available

  Google's latest Gemini 3 models are now fully supported in Abundly. Select Gemini 3 from the model picker to try it out.

  ## Team management improvements

  * **Collapsible teams** — organize your workspace by expanding and collapsing agent teams
  * **Better invitations** — improved invitation system with proper team role handling
  * **Role clarity** — clear distinction between workspace-level and team-level roles

  ## Drag files anywhere

  Drop files anywhere in the chat window, not just the input field.
</Update>

<Update label="November 12, 2025" description="Agent Teams & Document Visibility">
  ## Agent teams

  Organize your agents into custom teams for better workspace management:

  * **Shared billing** — common billing across teams while maintaining fine-grained access control
  * **Collapsible interface** — expand and collapse teams with smooth transitions
  * **Access control** — teams support proper permissions and are workspace-specific

  ## Document visibility controls

  New visibility settings help your agents focus on what matters:

  | Visibility     | Behavior                              |
  | -------------- | ------------------------------------- |
  | **Full**       | Always included in agent's context    |
  | **Summary**    | Basic info only — saves tokens        |
  | **Searchable** | Findable via search but not in prompt |
  | **Hidden**     | Not visible to the agent              |

  ## Agent-to-agent improvements

  * Conversational context maintained across multiple agent calls
  * Agents can temporarily grant other agents read/write access to specific documents
  * Feature is now stable — removed beta warning!

  ## Gemini integration fixed

  Gemini models now work reliably via Vertex AI with improved error handling.

  ## Advanced data queries

  Data documents now support MongoDB-style query operators: `$or`, `$and`, `$nor`, `$not`
</Update>

<Update label="October 29, 2025" description="Security & Gemini Support">
  ## Enhanced file security

  Files now require authenticated access via secure proxy endpoints with file access logging for audit trails.

  ## Google Gemini models

  Gemini models are now available as an LLM option. Switch to Gemini from the model selector in any chat.

  ## Document management improvements

  * **Semantic search** — find data documents using natural language queries
  * **Hidden documents** — mark documents as hidden (useful for large email attachments)
  * **Folder renaming** — rename folders to keep your documents organized
  * **Batch operations** — select all button for bulk actions

  ## 3-month data retention

  Internal chat threads logs are now retained for 3 months. Agent documents and configurations are not affected. This doesn't affect the user's own chat threads, only the internal logs.
</Update>

<Update label="October 15, 2025" description="GPT-5 & Smarter Controls">
  ## GPT-5 support

  GPT-5 is now available in Abundly. It works well for complex reasoning tasks, though it's noticeably slower than Claude 4.5.

  ## Smarter LLM controls

  * **Thinking toggle** — fully respected per chat, with per-call override when needed
  * **Model switching** — change models mid-conversation for flexibility
  * **Larger responses** — tool responses now support up to 400k characters

  ## No more daily token limits

  Removed the daily token limit — your daily credit limits already cover usage.
</Update>

<Update label="October 6, 2025" description="Phone Calling & Code Execution">
  ## Phone calling

  Your agents can now make and receive phone calls:

  * Real-time voice conversations powered by OpenAI's realtime API
  * Configurable phone numbers per agent
  * Call summaries automatically generated at end of calls
  * Full agent capabilities available during calls

  <Info>
    Find the new Make Phone Call and Receive Phone Call capabilities in your agent settings.
  </Info>

  ## Enhanced code execution

  * **Async support** — code execution now supports async/await operations
  * **Tool calling** — agents can call any other tool from within code using `callTool(name, params)`
  * **Code from documents** — load and execute code from agent documents using `codeDocumentId`

  ## Event codes

  New event code system for fast sign-up to courses and events. Share simple access codes for streamlined onboarding.
</Update>

<Update label="September 19, 2025" description="Data Documents">
  ## Data Documents

  Agents can now use documents as structured databases with MongoDB-style operations:

  * Query, insert, update, and delete data items within agent documents
  * Filters include `$in`, `$nin`, `$ne`, `$gt`, `$gte`, `$lt`, `$lte`, `$exists`, `$regex`
  * Automatic ID assignment for data items
  * New `dataSchema` field to describe the structure of document data
  * Dedicated "Data" tab in the document panel

  <Tip>
    React artifacts can access database functions: `queryInDocumentData`, `insertIntoDocumentData`, `updateDocumentData`, `deleteFromDocumentData`
  </Tip>

  ## Enhanced React artifacts

  * Charts now work properly in React artifacts
  * Added support for common React hooks
  * More robust error handling — failed React apps no longer block Abundly
</Update>

<Update label="September 15, 2025" description="Microsoft Integration & SMS">
  ## Microsoft integration

  * **SharePoint** — access and manage SharePoint documents and libraries
  * **Outlook** — enhanced email capabilities with CC support

  <Info>
    SharePoint and Outlook are available on request — contact Abundly to enable.
  </Info>

  ## SMS messaging

  Agents can now send and receive SMS text messages with configurable phone numbers.

  ## Agent activity logging

  New comprehensive agent log with real-time event tracking:

  * Tool call monitoring with detailed execution information
  * Better debugging capabilities for agent behavior analysis
  * Find it in the "Activity" section

  ## React artifacts enhanced

  * `loadAgentDocument` — React components can now access agent documents as data
  * `updateAgentDocument` — React artifacts can modify agent documents
</Update>

<Update label="September 1, 2025" description="API Access Keys & Model Selection">
  ## Agent API access keys

  Give external systems programmatic access to your agents:

  * Generate secure API keys (similar to OpenAI or Anthropic keys)
  * Build custom UIs using tools like Lovable
  * Integrate agents into external workflows
  * Manage multiple access keys per agent

  Find it in Agent Settings → API Access.

  ## LLM model selection

  Choose the right model for your agent:

  | Model                | Best for                        |
  | -------------------- | ------------------------------- |
  | **Claude 4 Sonnet**  | All-round performance (default) |
  | **Claude 4.1 Opus**  | Complex reasoning (expensive)   |
  | **Claude 3.5 Haiku** | Simple tasks (fast & cheap)     |

  Find it in Agent Settings → Advanced.

  ## Thinking mode toggle

  Thinking mode is now enabled by default for better reasoning. Disable it for simple tasks where extra thinking tokens aren't worth the cost.
</Update>

<Update label="August 28, 2025" description="Activity Tracking & Document Improvements">
  ## User activity tracking

  New admin analytics for understanding workspace usage:

  * Tracks meaningful user actions: chat activities, agent management, document operations
  * Page navigation tracking with smart categorization
  * View in Admin UI: Usage → User Activity

  ## Token-efficient document saving

  Improved handling of large documents in chat — better performance when working with substantial content.

  ## Google Calendar improvements

  Enhanced Google Calendar integration with improved authentication and more reliable operations.
</Update>

<Update label="August 25, 2025" description="Claude Thinking Mode & Personal Auth">
  ## Claude thinking mode

  Claude can now show its reasoning process with interactive thinking blocks:

  * See how Claude works through problems
  * Collapsible thinking sections
  * Real-time streaming of thought processes

  ## 1 million token support

  Claude now supports massive context windows — perfect for processing large documents and complex workflows.

  ## Personal authentication tokens

  Connect your personal Google accounts:

  * Personal Google Drive access
  * Personal Gmail integration
  * Personal Google Calendar

  ## Large file uploads

  Handle files larger than 32MB with direct transcription service integration.
</Update>

<Update label="August 12, 2025" description="Google Calendar & Credit Limits">
  ## Google Calendar integration

  Connect your agents with Google Calendar for scheduling and event management.

  ## Workspace-level credit limits

  Agents can now have daily credit limits set by workspace admins:

  * New usage limits section in agent settings
  * Progress indicators and status badges
  * Inheritance from workspace default limits

  ## CSV export for usage data

  Export usage analytics to CSV files from the Usage page.

  ## Better copy/paste from artifacts

  Fixed background color issues when copying from artifacts to Google Docs.
</Update>

<Update label="July 11, 2025" description="Diary Calendar & Batch Documents">
  ## Diary calendar view

  Visualize your agent's activity history with a beautiful calendar interface:

  * Day view with activity indicators
  * Today button for quick navigation
  * Cleaner diary interface with better organization

  ## Batch document operations

  Move multiple documents to folders simultaneously with improved folder navigation.

  ## Improvements

  * Auto-scroll to bottom of chat when opening
  * Better error messages for tool credentials
  * All workspaces and agents sorted by name by default
</Update>

<Update label="June 29, 2025" description="Canvas Always Available">
  ## Canvas always available

  Canvas (artifacts) is no longer a toggleable capability — it's always available in chat.

  ## Smarter agents

  Agents now better understand the current date, their capabilities, and their email address.
</Update>

<Update label="June 27, 2025" description="Walk & Talk Mode">
  ## Walk & Talk mode

  Renamed "Voice Mode" to "Walk & Talk mode" with a new icon and bigger, easier-to-tap microphone button.
</Update>

<Update label="June 25, 2025" description="Deep Research & Document Publishing">
  ## Deep Research

  Comprehensive research capability powered by Perplexity's deep research model:

  * Expert-level research with exhaustive searches
  * Generates comprehensive reports
  * Asks clarifying questions before diving deep

  <Warning>
    Deep Research is thorough but expensive — a single research session can cost 300+ credits.
  </Warning>

  ## Document publishing

  Make any agent document publicly accessible:

  * Save artifacts as agent documents, then publish
  * Works for React apps, HTML, and Mermaid diagrams
  * Anyone with the link can view the published document

  ## Enhanced web search

  New search models from Perplexity:

  | Model                   | Best for                             |
  | ----------------------- | ------------------------------------ |
  | **sonar**               | Quick, cost-effective searches       |
  | **sonar-pro**           | Complex queries and follow-ups       |
  | **sonar-reasoning**     | Fast real-time problem-solving       |
  | **sonar-reasoning-pro** | Deep reasoning with Chain of Thought |

  New parameters: `search_recency_filter` (hour/day/week/month), `search_mode` (academic/web), `search_domain_filter`.
</Update>

<Update label="June 19, 2025" description="Prompt Caching & Version Control">
  ## Prompt caching

  Up to 90% cost reduction for repeated interactions with the same agent. Token costs are dramatically reduced when you have ongoing conversations.

  ## Enhanced version control

  * **Approve/reject workflows** — review and control specific changes in document versions
  * **Visual diff viewer** — collapsible unchanged regions with better styling
  * **Version timeline** — beautiful interface for navigating document history
  * **Bulk actions** — "Approve all changes" and "Undo all changes" buttons

  ## Chat length warnings

  Helpful warnings when chats get too long, helping you maintain optimal performance.

  ## Account enforcement

  Workspaces without credits or active subscriptions will see clear messaging and can no longer use agents until resolved.
</Update>

<Update label="June 12, 2025" description="Multichat & Artifact Editing">
  ## Multichat

  Multiple people can chat with the same agent in the same thread at the same time — perfect for classrooms or onboarding sessions.

  ## Collaborative artifact editing

  You and your agent can now collaboratively edit artifacts:

  * Edit artifacts directly in the chat interface
  * Changes automatically included in your next message
  * Track changes with version navigation
  * Revert edits when needed

  ## Rich text editing

  Enhanced markdown editor with better formatting and real-time preview for documents, instructions, and artifacts.

  ## Coupon codes

  Create coupon codes for a class and share via public URL for easy sign-ups.

  ## Agent document links

  Agents can create direct links to documents that open in side panels for seamless viewing.
</Update>

<Update label="May 26, 2025" description="Abundly Rebrand & Security">
  ## Welcome to Abundly

  We've rebranded with updated logos and a fresh look throughout the platform.

  ## Enterprise billing

  * Enterprise plan option with dedicated support
  * Promotion codes in subscription checkout
  * Manual billing option for enterprise customers

  ## Security enhancements

  * Application-level encryption for secrets and API keys
  * New secret management UI for integration credentials
  * reCAPTCHA protection on signup

  ## Improvements

  * Document version history increased from 10 to 100 versions
  * Enhanced web scraping with improved Firecrawl tool
  * Better Slack message threading
</Update>

<Update label="May 16, 2025" description="Agent Tagging">
  ## Agent tagging

  New clickable tag system makes it easier to categorize and filter agents.

  ## Simplified agent editing

  Users with "Edit" access can now modify agent names and profile pictures — dedicated renaming section in settings.
</Update>

<Update label="May 14, 2025" description="Tool Approval & Document Folders">
  ## Tool approval system

  New approval system for safer agent interactions:

  * In-chat approval UI for reviewing and controlling tool actions
  * Enhanced approval workflows for HTTP requests and email sending

  ## Document folders

  Organize agent documents with folders and subfolders.

  ## Memory capability

  Agents can now save and retrieve related memories for better context over time.

  ## Notion integration

  Connect your agents with Notion databases.

  ## Credits system

  New credit system for tracking and managing resource usage with automatic allocation for paid subscriptions.
</Update>


# API Access
Source: https://docs.abundly.ai/features/agent-api-endpoint

Expose your agent to external systems via HTTP API, MCP, webhooks, or an embeddable chat widget

Your agent can be called from outside Abundly. In the agent's **Settings → API Access** tab, you can expose the agent as an HTTP API, as an [MCP](/integrations/mcp) server, as a webhook receiver, and as an embeddable chat widget — each with its own configuration, but sharing the same API keys.

<Frame>
  <img alt="API Access settings tab showing collapsible sections for API Keys, HTTP API, MCP Server, Chat Widget, and Webhooks" />
</Frame>

## When to use each

| Exposure mode    | Use for                                                                               | Authentication                         |
| ---------------- | ------------------------------------------------------------------------------------- | -------------------------------------- |
| **HTTP API**     | Calling your agent from a script, backend service, or automation tool                 | API key                                |
| **Document API** | Reading and writing agent documents from external systems                             | API key                                |
| **MCP server**   | Letting AI apps like Claude Desktop, Cursor, or n8n use your agent as a tool provider | API key                                |
| **Chat widget**  | Embedding an AI chat bubble on your own website                                       | API key (stored on your backend proxy) |
| **Webhooks**     | Receiving callbacks from external services like Stripe, GitHub, or Trello             | None (URL is the secret)               |

All modes run through the same agent, respect the same capabilities and guardrails, and log activity to the [activity log](/features/monitoring).

## API keys

HTTP API and MCP calls are authenticated with API keys. Open the **API Keys** section to create, list, and revoke keys for the agent.

Include a key on every request using either header:

```
X-Agent-Access-Key: your_key_here
Authorization: Bearer your_key_here
```

<Warning>
  Keys grant full access to the endpoints you've defined. Never commit them to source control or expose them in client-side code.
</Warning>

### Workspace API keys

By default, API keys are scoped to a single agent. If you have multiple agents that external systems need to access, you can create **workspace API keys** instead — a single key that authenticates across all agents in the workspace that opt in.

1. Go to **Workspace → API Keys** tab and create a workspace key (requires workspace admin role)
2. On each agent that should accept the key, open **Settings → API Access → API Keys** and enable **Allow Workspace Keys**

Workspace keys use the same headers as agent keys and work with all exposure modes (HTTP API, Document API, MCP, widget). They start with `wk_` to distinguish them from per-agent keys.

## HTTP API

Turn on **Enable HTTP API** to expose your agent at:

```
https://your-service-domain/agents/{agentId}/api/{path}
```

You explicitly define which paths are available. Any request to an undefined path returns 404.

For each endpoint, you choose:

* **Path** — e.g. `summarize` or `ingest/orders`
* **Handler** — either a natural-language prompt for the agent, or a [script](/features/scripts) that runs directly without invoking the LLM
* **Model override** (optional) — run this endpoint on a specific model, separate from the agent's default

```bash theme={null}
curl -X POST https://your-service-domain/agents/{agentId}/api/summarize \
  -H "X-Agent-Access-Key: your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"text": "..."}'
```

The JSON body is passed to the agent (or script) as input. The response contains the agent's reply.

## Document API

The Document API provides built-in REST endpoints for reading and writing your agent's [documents](/features/documents) — no custom scripts needed.

Turn on **Enable Document API** and select which documents to expose. For each document, you choose read and/or write permissions.

Documents are accessed at:

```
https://your-service-domain/agents/{agentId}/documents/{documentId}
```

The document ID is shown in the settings card for each exposed document.

### Reading a document

```bash theme={null}
curl https://your-service-domain/agents/{agentId}/documents/{documentId} \
  -H "X-Agent-Access-Key: your_key_here"
```

Returns the document's name, description, content, and data (if it's a [database](/features/agent-databases)).

### Querying a database document

For database-type documents, you can query, insert, update, and delete records:

<CodeGroup>
  ```bash Query theme={null}
  curl -X POST https://your-service-domain/agents/{agentId}/documents/{documentId}/data/query \
    -H "X-Agent-Access-Key: your_key_here" \
    -H "Content-Type: application/json" \
    -d '{"filter": {"category": "Electronics"}}'
  ```

  ```bash Insert theme={null}
  curl -X POST https://your-service-domain/agents/{agentId}/documents/{documentId}/data/insert \
    -H "X-Agent-Access-Key: your_key_here" \
    -H "Content-Type: application/json" \
    -d '{"items": [{"name": "New Product", "price": 29.99}]}'
  ```
</CodeGroup>

<Tip>
  The Document API is useful when external systems need to read or update the same data your agent works with — for example, syncing a product catalog, populating a knowledge base, or pulling reports.
</Tip>

## MCP server

[MCP](/integrations/mcp) (Model Context Protocol) lets external AI applications — Claude Desktop, Cursor, n8n, and others — call tools you define on your agent.

Turn on **Enable MCP server** to expose the agent at:

```
https://your-service-domain/agents/{agentId}/mcp
```

For each **tool** you add, you specify:

* **Name and description** — shown to MCP clients so they know when to call it
* **Handler** — a prompt for the agent, or a [script](/features/scripts) for deterministic execution
* **Model override** (optional)

Connect from Claude Desktop by adding your server to the MCP config:

```json theme={null}
{
  "mcpServers": {
    "my-agent": {
      "url": "https://your-service-domain/agents/{agentId}/mcp",
      "headers": {
        "X-Agent-Access-Key": "your_key_here"
      }
    }
  }
}
```

## Chat widget

The chat widget embeds your agent as a floating chat bubble on any external website. Visitors can ask questions and get streaming answers — without logging in or leaving your site. Turn on **Enable widget** in the API Access tab to configure it.

The widget uses a secure backend proxy pattern: a small server-side endpoint on your server holds the API key and forwards requests to Abundly, so your key never appears in client-side code.

For setup instructions, backend proxy examples, configuration options, and the headless `createClient` API, see the [Chat Widget](/features/chat-widget) page.

## CORS settings

If you want to call your agent's HTTP API directly from a browser (for example, from a web application), you need to configure which origins are allowed to make cross-origin requests. In the **HTTP API** section, use **Allowed origins (CORS)**:

* **Allow requests from Abundly apps** — Automatically allows requests from Abundly's user-facing domains
* **Custom origins** — Add any additional origins (e.g. `https://yourapp.com`) that should be permitted to make browser requests

<Note>
  CORS restrictions apply to browsers only. If you're calling the API from a server-side environment, you can leave CORS unconfigured. The chat widget always goes through your backend proxy (server-to-server), so it doesn't need a CORS entry either.
</Note>

## Webhooks

Many external services send notifications by calling a URL you provide — Stripe after a payment, GitHub when a PR opens, Trello when a card moves. Turn on **Enable Webhooks** to receive these callbacks.

Each webhook you define gets its own URL:

```
https://your-service-domain/agents/{agentId}/webhook/{integration}
```

Unlike HTTP API calls, webhooks don't use API keys — the URL itself provides routing. Requests to undefined integration names return 404.

For each integration, configure:

* **Integration name** — used as the URL path, e.g. `stripe` or `github`
* **Handler** — a prompt describing how the agent should process the callback, or a [script](/features/scripts) for fast, deterministic handling
* **Model override** (optional)
* **Signature verification** (optional) — enable HMAC signature verification to ensure requests are authentic. Many services (Jira, GitHub, Stripe, Linear, Shopify, etc.) sign webhook payloads with a shared secret. When configured, Abundly verifies the signature before processing — requests with invalid or missing signatures are rejected with a 401. Store the signing secret as a "Webhook signing secret" in your workspace's Secrets page.

Webhooks respond immediately with `200 OK` and process the payload asynchronously, so external services never time out.

<Tip>
  Paste the webhook URL into the external service's webhook settings. The agent will run each time the service fires an event — driven entirely by your prompt or script.
</Tip>

## Agent self-management

If you want the agent to manage its own exposed surfaces programmatically — for example, spinning up a webhook endpoint in response to a user request, or exposing a new MCP tool after learning a new skill — enable the **Endpoint Management** capability on the agent's [capabilities](/features/capabilities) page.

With this capability, the agent can list, add, update, and remove entries across HTTP API endpoints, Document API entries, MCP server tools, and webhooks, as well as toggle surface settings like enabling/disabling, activity logging, and CORS configuration.

## Activity log

All exposure modes record entries in the [activity log](/features/monitoring), including the full request headers and body, the agent's response, and any credits consumed. You can disable logging per section if you have a high-volume endpoint where logging isn't useful.

## Learn more

<CardGroup>
  <Card title="MCP Servers" icon="plug-circle-plus" href="/integrations/mcp">
    Add external MCP servers to your agent's capabilities
  </Card>

  <Card title="Code Execution" icon="code" href="/features/scripts">
    Back endpoints, tools, and webhooks with scripts for faster handling
  </Card>

  <Card title="Monitoring" icon="gauge-high" href="/features/monitoring">
    Review what your agent does when triggered externally
  </Card>
</CardGroup>


# Agent Databases
Source: https://docs.abundly.ai/features/agent-databases

Instant structured data stores your agent can create, query, and update—no setup required

Agent databases (also known as data documents) let your agent store and query structured records. Use agent databases to track invoices, manage inventories, store research findings, or maintain structured logs of anything it needs to remember across conversations. Pair them with [interactive apps](/features/apps) to browse, filter, and edit your data visually.

<Frame>
  <img alt="Agent database showing a product catalog in table view with columns for name, category, price, stock, SKU, and rating" />
</Frame>

Technically, an agent database is a special type of [agent document](/features/documents) that stores data as a collection of JSON records—like a lightweight MongoDB collection. The agent has tools to insert, query, update, and delete records.

## Creating an agent database

Creating a database is simple - just ask your agent to create it. No setup needed.

* "Create a database to track processed invoices"
* "Set up a database for customer feedback with fields for date, rating, and comments"
* "Make a database to log all emails you send"
* "Turn these notes from our product brainstorm workshop into a database of ideas".

The agent creates the database instantly and defines an appropriate data schema based on your description. It shows up alongside your other agent documents in the Documents panel.

## When to use an agent database

Think of it like choosing between a word processor and a spreadsheet. You'd personally write an article or report in a document, but you'd track a list of products, transactions, or tasks in a spreadsheet or database table.

**Use a database when you have many items of the same type**—products with prices and stock levels, support tickets with status and priority, research findings with sources and dates.

**Use a text document when the content is prose**—articles, meeting notes, guidelines, templates. If you're writing something you'd read top to bottom, it belongs in a document.

Could you store a product list in a text document? Sure—it works fine for a handful of items. But once you have dozens or hundreds, text gets inefficient: every insert or update requires the agent to find the right spot, parse the format, and rewrite the text. That's slow, token-expensive, and error-prone. Databases handle this with precise operations—faster, cheaper, and no risk of corrupting the format. Plus you get a visual table view instead of scrolling through text.

Not sure which to use? Just pick one—you can always ask your agent to convert a text document into a database or vice versa.

Agent databases are a simple way to create an instant data store, but you can also connect to an external database if needed—see [Agent database vs external database](#agent-database-vs-external-database) below.

## Editing data

You can ask the agent to make direct changes. The agent will interpret your request and use database operations to make the changes. Examples:

* **Insert data**: "Add this invoice to the database" (drop PDF in chat)
* **Update data**: "Update the status of invoice 123 to 'paid'"
* **Delete data**: "Delete last week's invoice from company Y"

If the database should be used as part of the agent's workflow, describe that in the agent's instructions:

* "When you receive an invoice, add it to the database"
* "Keep the product catalog up to date whenever you hear about changes"
* "After your weekly research, update the database with the new findings"

## Querying data

Ask your agent questions about the data in natural language, and it will execute database queries to answer your question. Examples:

* "How many invoices did you process this week?"
* "Show me all pending tickets from the last 24 hours"
* "What's the total amount of approved expenses?"

**Direct queries** find records by exact field values—status equals "pending", amount greater than \$1000, category is "urgent".

**Semantic search** finds records by meaning, not just exact matches. This uses RAG (retrieval-augmented generation) to understand what you're looking for:

* "Find all feedback with negative sentiment"
* "Show me bug reports that mention data loss"
* "Which research findings are about competitor pricing?"

Semantic search is especially useful when you don't know the exact wording used in the records.

You can also make search part of the agent's workflow:

* "When you receive a customer support email, check the FAQ database for similar questions before suggesting an answer"

## Viewing agent database contents

The agent database is listed alongside the other documents in the Documents panel. Click on it to see the contents.

<Frame>
  <img alt="Agent database Data mode showing raw JSON data and the Show data schema toggle" />
</Frame>

* **Data mode** — Raw JSON data plus schema (editable)
* **View mode** — A simple database viewer. You can toggle between table view (like a spreadsheet) and card view (grid of cards for each record). Nested objects and arrays appear as clickable links that expand for details.

This is useful for small databases, or to get an overview of what kind of data is in the database.

For larger databases, consider asking the agent to create an [interactive app](/features/apps) to browse the data. One or two prompts is usually enough to get you something like this:

<Frame>
  <img alt="Product Catalog Dashboard app showing inventory statistics, category filters, and product cards with prices and stock levels" />
</Frame>

## Importing and exporting data

When viewing an agent database, you can click on the menu to import or export data in JSON text format.

* **Export data** — Download the data as a JSON text file.
* **Import data** — Upload a JSON file from your computer. When importing data, you can choose to append to the existing data or replace it.

## Connecting to apps and scripts

Databases pair naturally with apps and scripts (which also are agent documents):

* **[Interactive apps](/features/apps)** — Dashboards, forms, and data tables that read from and write to databases
* **[Executable scripts](/features/scripts)** — Code that performs database operations programmatically

Example:

1. "Create a database to track beta feedback"
2. "Build a dashboard showing feedback trends and top issues"
3. "Add a form so users can submit new feedback"
4. "Write a script to import feedback from this google spreadsheet"

<Tip>
  If you want the data to be visible to other users, you can publish the dashboard app. Open the dashboard app and click "Publish" in the menu.
</Tip>

## Data schema

A database can have a schema that describes the shape of its records (which fields etc). The agent defines the schema automatically when creating the database, so you don't need to manage this unless you want to.

To view or edit the schema, open the database and click **Schema** in Data mode.

Schemas can be written in any format, but the agent uses TypeScript by default—a concise way to express data structures. For example here is a simple schema for a task list:

```typescript theme={null}
interface Task {
  title: string;
  assignee: string;
  status: 'todo' | 'doing' | 'done';
}
```

The schema helps the agent understand how to read and write records consistently. It's a guide, not a strict constraint—the agent can add fields or adjust as needed.

Schemas can include nested objects for richer data:

```typescript theme={null}
interface Task {
  title: string;
  assignee: {
    name: string;
    email: string;
  };
  status: 'todo' | 'doing' | 'done';
}
```

<Tip>
  Need to change the schema? Ask your agent—it can update the schema and migrate existing records in one step. For example: "Add a priority field to all tasks, default to 'medium'."
</Tip>

## Security

Agent databases follow the same access model as other [agent documents](/features/documents):

* By default, the agent can read and write to databases (disable via the **Edit Documents** capability)
* Users with access to the agent can view database contents
* For detailed permissions, see [access control](/security/access-control)

## Agent database vs external database

Agent databases are included by default—instant to create, no setup required, and tightly integrated with your agent. However, sometimes you may want your agent to connect to an external database, such as an existing company database, a shared data source used by multiple systems, or when you need enterprise-grade scaling. External databases (PostgreSQL, MongoDB, etc.) can be connected via [MCP](/integrations/mcp) or other integrations.

| Agent database                             | External database                         |
| ------------------------------------------ | ----------------------------------------- |
| Instant creation, no setup                 | Requires provisioning and configuration   |
| Great for prototyping and simple workflows | Enterprise-grade with backups and scaling |
| Hundreds of records                        | Millions of records                       |
| Data stays within the agent                | Multiple systems can access the same data |
| Query within single collection (no joins)  | Complex queries, joins, transactions      |
| RAG included (semantic search)             | No RAG (in most cases)                    |

You can change your mind later. If you started with an agent database and later decide to store the data in an external database, or vice versa, you can ask the agent to convert/migrate the data. Enable the Code Execution capability in that case, so it can write a script for it.

You can also combine both approaches - store some data in an agent database and some in an external database.

<Tip>
  If you aren't sure that you need an external database, then start with an agent database. It is the simplest way to get you agent up and running quickly, and you can always migrate to an external database later if your needs grow.
</Tip>

## Agent database vs spreadsheet

Your agent can access external spreadsheets via capabilities like [Google Drive](/integrations/google-drive). This works well for small data (a few hundred cells), for example to maintain a product catalog with a few dozen products.

But with large and complex spreadsheets with 1,000+ cells, agents struggle—reading and updating cells requires parsing row/column coordinates, which gets slow, expensive, and error-prone. In that case it is better to use an agent database or external database.

## Learn more

<CardGroup>
  <Card title="Interactive Apps" icon="browser" href="/features/apps">
    Build dashboards and forms backed by database data
  </Card>

  <Card title="Code Execution and Scripts" icon="code" href="/features/scripts">
    Write code that performs database operations programmatically
  </Card>

  <Card title="Agent Documents" icon="folder-open" href="/features/documents">
    Learn about other types of agent documents
  </Card>
</CardGroup>


# Agent management
Source: https://docs.abundly.ai/features/agent-management

Centralized administration for managing agents, filters, and limits across your workspace

The Agent Management page gives workspace and team administrators a centralized view of the agents they can manage. You can search, filter, sort, and update agents in bulk, including status, team assignment, model, tags, owners, and daily limits.

<Frame>
  <img alt="Agent management page showing the Agents tab with a list of agents, their status, privacy settings, and capabilities" />
</Frame>

## Access levels

* **Workspace administrators** can manage all agents in the workspace.
* **[Team](/features/teams) administrators** (when teams are enabled) can manage agents in their teams.

## Agents tab

The Agents tab is a flat, searchable table with sorting, filtering, and bulk actions.

| Column                  | Description                                                                                                        |
| ----------------------- | ------------------------------------------------------------------------------------------------------------------ |
| **Agent**               | Name, avatar, privacy icon, and compact badges for team and tags                                                   |
| **Usage (today)**       | Credits used today, with limit context                                                                             |
| **Last activity**       | Latest entry in the agent activity log                                                                             |
| **Last updated**        | Most recent configuration change date                                                                              |
| **Credit details**      | Whether credit details are shown in logs/chat sidebar                                                              |
| **Value & criticality** | Category badges and criticality level when value capture is enabled; warning icon when required fields are missing |

Click a row to expand details for:

* Tags
* Enabled capabilities (including whether a capability is disabled by admin policy)
* API capabilities
* MCP servers
* Agent owners

### Search, filters, and sorting

You can search by agent name and open a filter panel to narrow by status, privacy, teams, tags, model, owner, capabilities, API capabilities, MCP servers, usage range, and date ranges for last activity and last updated.

When filters are active, a **Reset** button appears next to the filter panel button, letting you clear all filters in one click.

Column headers support sorting for key fields like usage and activity dates, and active filters are counted in the filter button.

### Bulk actions

Select one or more agents with the checkboxes to run bulk actions:

* Trigger (run each selected agent once with a message you provide)
* Send notice (message the owners of each selected agent)
* Enable
* Disable
* Move to team (if you have move permissions)
* Set daily limit
* Set model
* Turn credit details on or off

Triggered runs are queued and start as capacity allows. Disabled agents are skipped.

**Send notice** reaches the people responsible for a set of agents — for example to announce a policy change or ask them to review something. Each owner gets notified, and the notice is listed under [Owner notices](/features/monitoring#owner-notices) on each agent, clearly marked as coming from you rather than from the agent. Agents with no owner notify their team or workspace admins instead.

### Per-agent quick actions

Each agent row includes an actions menu where you can:

* Enable or disable the agent
* Add owner
* Edit tags
* Move to team
* Open agent settings
* Show or hide credit details

<Tip>
  You can only remove an owner if at least one other owner remains on that agent.
</Tip>

## Agent limits tab

The Agent limits tab lets you manage daily credit limits directly from Agent Management:

* Set or edit a workspace default daily limit
* Add or change custom per-agent limits
* Remove a custom limit (fall back to the default)
* Reset an agent's daily usage counter

For details on how limits behave when reached, see [Usage & Limits](/features/usage#agent-limits).

## Capability policies

Capability availability policies are configured outside this page:

* **Capabilities** for workspace-wide defaults plus per-team overrides (when teams are enabled)
* **Workspace management → Model selection** for which LLM models agents may use in this workspace (you can turn families off, for example; models disabled by Abundly for all workspaces stay off)

## FAQ

<AccordionGroup>
  <Accordion title="What happens when I disable an agent?">
    Disabled agents stop processing triggers, schedules, and incoming work until re-enabled.
  </Accordion>

  <Accordion title="Can I update many agents at once?">
    Yes. Use row selection plus bulk actions to trigger agents with a message, enable/disable them, move teams, or set limits/models for multiple agents in one step.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Value dashboard" icon="gauge" href="/features/value-dashboard">
    Cross-agent overview of triggers, credits, and value
  </Card>

  <Card title="Usage & limits" icon="chart-line" href="/features/usage">
    Track credit usage and configure limits in detail
  </Card>

  <Card title="Teams" icon="users-rectangle" href="/features/teams">
    Organize agents and administration by team
  </Card>

  <Card title="Sharing" icon="clone" href="/features/sharing">
    Clone and share agents across workspaces
  </Card>

  <Card title="Access control" icon="shield" href="/security/access-control">
    Understand workspace, team, and agent permissions
  </Card>
</CardGroup>


# Agent value capture
Source: https://docs.abundly.ai/features/agent-value-capture

Document the business value and operational criticality of your agents so admins can prioritize and report on impact

Agent value capture helps workspace and team administrators understand what each agent delivers. Agent owners describe value in structured categories with quantified descriptions, and optionally rate how critical an agent is to day-to-day operations.

<Frame>
  <img alt="Agent settings Value capture section showing a Speed increase value entry and Major impact criticality selected" />
</Frame>

## Who configures what

| Role                     | What they can do                                                                                                                                                           |
| ------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Workspace admin**      | Turn value capture on or off, define value categories, enable each capture method, set workspace-wide required/optional rules, and control the owner-facing warning banner |
| **Team admin**           | Tighten requirements for their team (never loosen below workspace rules) and view compliance on the team settings tab                                                      |
| **Agent owner / editor** | Fill in value entries and criticality under **Settings → Main Settings → Value capture**                                                                                   |

## Capture methods

When enabled, a workspace can use one or both methods:

### Quantifiable value

Each entry combines a **category** (for example Cost reduction, Revenue increase, Speed increase) with a **free-text description**. Use monthly figures where possible (e.g. "\~\$8,400 / month") so values stay comparable across agents.

Every entry has two figures: an **expected** value (the estimate) and an optional **actual** value you fill in later once the agent is running. For actual value you can either:

* Choose **per interaction** (the default for new entries) — set a numeric amount and optional unit (for example `$` or `USD`), pick **all interactions** or a specific channel (for example Email or Scheduled triggers), and let the [Value dashboard](/features/value-dashboard) compute the period total from usage counts, or
* Enter a **manual** figure in free text (same style as expected).

The expected and actual figures each track their own last-edited time — shown as a small clock you can hover — so you can tell at a glance whether either figure is fresh or has gone stale, independently of the other.

Workspace admins manage the category list — rename, reorder, recolor, add, or delete unused categories. Categories in use by an agent cannot be deleted until those agents are updated.

### Criticality for operations

A single choice: *If this agent was removed…*

* Impossible to operate
* Major impact
* Minor impact
* No impact

These four levels are fixed in the current release.

## Required fields and compliance

Workspace admins can require at least one quantifiable value entry and/or a criticality level for all agents. Team admins may **tighten** these rules for their team but cannot make a field optional if the workspace requires it.

When fields are required but empty:

* Agent owners see a warning banner in the **Value capture** section on Main Settings (when enabled by the workspace admin)
* Main Settings shows a warning indicator when required fields are missing
* The agent management table shows a warning icon and summarizes captured value in the **Value & criticality** column
* Team admins can view compliance counts on the team's **Settings** tab

## Where to find it

* **Workspace → Settings → Value capture** — master toggle, categories, workspace requirements, and owner warning banner
* **Team → Settings → Agent value capture** — team requirement overrides and compliance summary (when teams are enabled)
* **Agent → Settings → Main Settings → Value capture** — owner entry form (when the feature is enabled for the workspace)
* **Agent management** — column summary and missing-field indicators for admins

<Tip>
  Workspace Explorer and Team Explorer agents can list agents with `valueEntries` and `criticality`, and call `summarize_workspace_value_capture` or `summarize_team_value_capture` for aggregate reports.
</Tip>

## Learn more

<CardGroup>
  <Card title="Agent management" icon="sliders" href="/features/agent-management">
    Centralized view of agents including value summaries
  </Card>

  <Card title="Teams" icon="users" href="/features/teams">
    Team settings and per-team governance
  </Card>
</CardGroup>


# Interactive Apps
Source: https://docs.abundly.ai/features/apps

Agents build web applications on demand—dashboards, forms, calculators, and custom tools

Apps are [agent documents](/features/documents) containing executable code that renders as a user interface. Ask your agent to create a dashboard, form, or custom tool, and it writes the code and displays the result. You can see and tweak the code yourself, or ask the agent to make changes.

<Frame>
  <img alt="Feedback Hub dashboard aggregating customer feedback from Twitter, Email, Survey, App Review, and Support Calls—showing sentiment trends, key metrics, and a live feed of categorized feedback" />
</Frame>

## Two types of apps

Agents can create two types of applications:

| Type           | Best for                                                                                           |
| -------------- | -------------------------------------------------------------------------------------------------- |
| **React apps** | Rich interactive tools using Chakra UI, Recharts, and Lucide icons as built-in component libraries |
| **HTML apps**  | Custom layouts, 3D and map visualizations, and anything that fits better as plain HTML/JS/CSS      |

Both types run in a sandboxed iframe and can read from and write to [agent databases](/features/agent-databases), load other documents, pull in libraries from public CDNs, generate and trigger file downloads (Excel, Word, ZIP, and more using standard CDN libraries), and call a curated set of agent tools — for example, HTTP requests through an [API capability](/features/http) or queries against BigQuery, Snowflake, Notion databases, and similar data sources. Document and agent-database helpers on `abundly` (such as `loadAgentDocument` and `queryDocumentData`) use the same toolbox execution path as `abundly.callTool` with the matching tool names when the app runs in an agent context. The agent chooses the appropriate type based on your request, but you can tell it to use one or the other if you have a preference.

## Multi-file apps

Simple apps live in a single document, but larger projects can be split across multiple files in a folder — for example, an `index.html` entry file alongside separate CSS, JavaScript, and image files. When a folder contains an entry file named `index.html`, `index.jsx`, or `index.tsx`, a **Run app** button appears directly on the folder row, letting you launch the whole folder as one app without navigating into it.

Multi-file apps work across all document scopes — agent, workspace, team, and chat. Any folder with an entry file shows the Run app button, regardless of which library it lives in.

<Tip>
  Ask the agent to "build this as a multi-file app in a folder" when you want a cleaner structure or when the app needs several supporting files.
</Tip>

## How apps work

Interactive apps are shown alongside your other agent documents in the Documents view.

When you open an app document, you use a mode selector:

* **View mode** — The live rendered application (default view)
* **Text mode** — The source code (React/JSX or HTML)
* **Changes mode** — A history of all changes made to the app

You can edit code directly or ask the agent to make changes. Every edit is versioned automatically. When an app is open, a **reload** button re-runs it with the latest code, and a **fullscreen** button expands the preview to fill the window.

## App scopes

Apps can live in multiple scopes depending on where they're created:

| Scope                         | Persistence                                                              | Best for                                                  |
| ----------------------------- | ------------------------------------------------------------------------ | --------------------------------------------------------- |
| **Chat document**             | Scoped to the current conversation; listed in the Chat Documents panel   | One-off visualizations, quick prototypes, temporary tools |
| **Agent document**            | Saved to the agent's Documents panel, available outside the conversation | Reusable apps, apps you want to publish or share          |
| **Workspace / Team document** | Shared across the workspace or team                                      | Shared tools, team dashboards, cross-agent apps           |

When you ask the agent to create an app in a conversation, it creates a **chat-scoped document** and opens it in a side panel next to the chat. Follow-up edits keep the same document, so you can iterate without the panel re-rendering from scratch. You can browse everything created in the current chat by clicking the **Chat documents** button in the chat action bar.

To keep an app around beyond the current conversation, promote it to an **agent document**:

* Ask the agent: "Save this as an agent document"
* Or open the document menu in the panel and choose **Save as agent document**

Once promoted, the app appears in your main Documents panel. You can open it anytime, the agent can reference it in any conversation, and it becomes eligible for publishing. Chat-scoped documents are automatically deleted when the chat is deleted.

<Tip>
  Use chat documents for exploratory work: "Show me a chart of last month's sales." When you have something worth keeping, save it as an agent document.
</Tip>

## Publishing apps with data

Like other agent documents, apps can be [published](/features/documents#publishing-documents) with a public URL. Anyone with the link can view and interact with the app.

If your app reads from an agent database, you need to publish that database too—otherwise the app won't have data to display. When publishing a database, you choose what public viewers can do:

<Frame>
  <img alt="Publish Document dialog showing permission checkboxes: Read (checked), Create, Update, and Delete operations for public viewers" />
</Frame>

| Permission | What it allows                                        |
| ---------- | ----------------------------------------------------- |
| **Read**   | View data (required for dashboards and data displays) |
| **Create** | Submit new records (enable for public forms)          |
| **Update** | Modify existing records                               |
| **Delete** | Remove records                                        |

<Warning>
  Publishing a database makes its entire contents publicly readable, regardless of which permissions you enable. The Read/Create/Update/Delete toggles control which **operations** public viewers can perform — they don't hide rows. Anyone with the document link can fetch the full dataset, even if your app's UI only displays a slice or aggregates. Only publish databases where every row is safe to share publicly.
</Warning>

<Tip>
  For a public feedback form, publish both the app and the database. Enable **Read** and **Create** on the database so visitors can submit entries. Leave **Update** and **Delete** disabled unless you want the public to modify existing data.
</Tip>

<Note>
  Apps that call agent tools (for example, HTTP requests through an API capability) work in the agent chat but not on public share links — the publish dialog warns you if this applies to your app.
</Note>

## Typical use cases

**Dashboards and data browsers**

* Sales metrics with filters by region and date
* Project status boards showing tasks by stage
* Inventory views with search and category filtering

**Data collection**

* Customer feedback forms
* Event registration
* Order intake forms that write to a database

**Internal tools**

* Approval workflows with action buttons
* Scheduling views with drag-and-drop
* Comparison tools for evaluating options

**Public utilities**

* Pricing calculators for your website
* Quote generators for sales teams
* Interactive reports for clients

## Real-world example

A TV production company used Abundly to build a photographer scheduling system. The agent created its own database and interactive dashboard for visualizing and managing schedules. Staff could view the schedule, and the agent could suggest and apply changes through conversation.

<Frame>
  <img alt="Screenshot of the SVT Scheduler Agent dashboard" />
</Frame>

<Card title="Behind the Scenes: Making a Scheduler Agent for SVT" icon="tv" href="https://www.abundly.ai/blog/svt-scheduler-agent">
  Full case study showing how the agent built its own database and dashboard for a real production workflow
</Card>

## Learn more

<CardGroup>
  <Card title="Agent Databases" icon="database" href="/features/agent-databases">
    The data layer that powers interactive apps
  </Card>

  <Card title="Agent Documents" icon="folder-open" href="/features/documents">
    How apps fit into the document system
  </Card>

  <Card title="Code Execution and Scripts" icon="code" href="/features/scripts">
    Code that runs for processing and automation
  </Card>
</CardGroup>


# Configurable Capabilities
Source: https://docs.abundly.ai/features/capabilities

Give your agent the tools it needs to do its job

A **capability** is a feature or integration that your agent can use to accomplish tasks. Each agent starts with a basic set of commonly used capabilities, and you can add more as needed. Some capabilities also have configuration options that can be adjusted to match your needs.

<Frame>
  <img alt="Capability card showing toggle and configuration options" />
</Frame>

## Types of capabilities

To browse available options, use the tabs on the agent's capabilities page: **Built-in** for platform features and integrations, **API** for custom API capabilities, and **MCP** for custom MCP servers (if enabled in your workspace).

* Some capabilities represent **Platform features** - such as the ability to generate images, execute code, edit documents, etc.
* Other capabilities represent **external integrations**—such as Slack, Google Drive, HubSpot, and GitHub. You can also add your own integrations using [MCP servers](/integrations/mcp). Learn more about integrations in the [integrations overview](/integrations/overview).

## Managing capabilities

Capabilities are an important part of managing the behavior of your agent.

* If your agent has too few capabilities, it might not be able to do the task you asked it to do.
* If your agent has too many capabilities, or capabilities that it doesn't need, it might get confused about when to use which capability.

Agents can combine capabilities in powerful ways. Here are some examples:

**With Web Search, Edit Documents, and Send Email:**

* "Every morning, search for recent news about our competitors and add a summary to my research document."
* "Research the top 10 AI startups in Europe and email me a summary."

**With Google Drive, Generate PDF, and Send Email:**

* "Every friday at 9am, generate a PDF report of this week's sales data in google drive and email it to the team."

**With Receive Email and Edit Documents:**

* "When you receive an email with an invoice attached, extract the key details and add them to the invoice database."

**With Twitter and Slack:**

* "Monitor Twitter for mentions of our brand and post a daily summary to the #social-listening Slack channel."

<Tip>
  Start with minimal capabilities and add more as needed. This lets you get started faster.
</Tip>

## Capabilities your admin has blocked

If your workspace admin has disabled certain capabilities workspace-wide (or for your team), they don't disappear from the **Built-in** tab — they show up as greyed-out cards labelled **Not available**, sorted below the capabilities you can use. Click a greyed card to see why it's unavailable. Category badges on the tab show an available/total count when some capabilities in that category are blocked, so you can tell at a glance where restrictions apply.

See [Guardrails](/security/guardrails) for how workspace-level capability controls work.

## On-demand capabilities

You don't need to configure all capabilities upfront. If you ask the agent to do something that requires a capability that isn't enabled, it will automatically request it by showing a capability card in the conversation.

<Frame>
  <img alt="Agent requesting a capability with a capability card" />
</Frame>

## FAQ

<AccordionGroup>
  <Accordion title="What's the difference between capabilities, tools, and MCP servers?">
    * A **tool** is a lower-level technical function exposed to an LLM. For example, `postSlackMessage` is a tool that posts a message to a Slack channel.
    * A **capability** groups related tools together. The Slack capability includes posting messages, listing channels, reading threads, etc. Capabilities are toggled on/off by the user and may require authentication.
    * An **MCP server** is an external service that exposes tools to an LLM. Adding an MCP server to your agent creates a capability for those tools. See [MCP Servers](/integrations/mcp).
  </Accordion>
</AccordionGroup>


# Advanced Chat
Source: https://docs.abundly.ai/features/chat

A feature-rich chat interface for interacting with your agent

Each agent comes with an advanced chat interface where you can talk to your agent, upload files, generate and edit documents, and use voice input. This chat is similar to apps like ChatGPT and Claude. The main difference is that Abundly agents don't live exclusively in the chat, the chat is just one of many possible communication channels. See [Multi-channel Communication](/features/communication) for more details.

<Frame>
  <img alt="Chat interface showing a conversation with an agent" />
</Frame>

<Tip>
  Think of each conversation as a local context for the agent, which includes all messages in that conversation, plus the agent's instructions, documents, and capabilities. If you want any chat content to be available to the agent outside of that specific conversation, you need to ask the agent to save it in its instructions or as an agent document.

  [Diary entries](#diary) are an exception—when diary writing is enabled, the agent writes these automatically and can read them later, outside the chat.
</Tip>

## Conversation context indicator

On desktop, the chat action bar includes a small circular indicator that shows how full the current conversation context is. As the conversation grows, the ring fills and changes color from gray to yellow, orange, and red.

Hover the indicator to see the current percentage and guidance. Click it to open a popover with detailed context usage information and links to optimization tips.

Once the conversation reaches 80% of the context limit, a **New chat with summary** button also appears next to the indicator. Click it to start a fresh, summarized chat proactively — the same handoff flow as the warning bar, but before the conversation gets very long.

The indicator is specific to each conversation. Starting a new chat gives you a fresh context window.

When a conversation grows very long, a warning bar appears above the chat input suggesting you continue in a new chat. Click **New chat with summary** and the agent generates a handoff recap of the current conversation — your goal, progress so far, pending items, key decisions, and links to any documents involved — and drops it into a fresh chat so you can pick up where you left off without dragging the full history along.

## When to use the chat

The role of the chat depends a lot on what the agent's job is.

During initial setup and tuning, the chat is particularly useful for:

* **Discussing and updating the agent's instructions** "I want you to help me handle incoming invoices, check that they are valid, and then route them to the right team. What do you need from me to do this?"
* **Testing the agent**: "Check this incoming order PDF and process it according to the instructions, as if you received it in an email. Don't interact with any external systems, just tell me what you would have done."
* **Understanding and improving the agent's behavior**: "You flagged the latest invoice as potentially fraudulent, even though you've processed similar invoices before without any issues. Can you analyze your instructions and docs and figure out why? And then suggest how we can improve your instructions to avoid this in the future?"
* **Adding more context**: "Since you keep asking me about people's contact info, here is a link to our contacts page. Update your instructions to use this as a source of information."

Once the agent is up and running, the role of the chat will vary.

* **Some agents are very chat-centric**, where most of the interaction happens in the chat. For example if the job of the agent is to analyze the competitive landscape, you might use the chat as the primary interface for discussing this.
* **Some agents use other channels mostly**, such as email or Slack. In that case the built-in chat has a secondary role.
* **Some agents handle a workflow automatically** and rarely need to chat at all. For example, once the invoice router is up and running and doing its job, it will work silently in the background for the most part, except for when it needs tuning.

## File uploads

You can upload files directly in the chat, using drag drop or by clicking the upload button.

* **Images** — The agent can analyze and describe images, extract text, or use them as context
* **Documents** — PDFs, Word documents, spreadsheets, and presentations
* **Audio** — Recordings that can be transcribed and analyzed
* **Data files** — CSV, JSON, and other structured data

The agent processes these files in context and can reference them throughout the conversation.

## Working with agent documents

From the chat, you can ask the agent to read, create, or update its [agent documents](/features/documents) and [databases](/features/agent-databases). This makes it easy to build up the agent's knowledge base through natural conversation.

* **Saving documents**: "Save these updated guidelines as an agent doc please"
* **Reading documents**: (drop a PDF) "Check if this PDF complies with the guidelines document"
* **Updating databases**: (after brainstorming) "Those are great ideas, please add them to the brainstorm database"
* **Creating databases**: (paste a whiteboard photo) "Here are some notes from our product planning session, please create a database and store the product ideas, with suitable fields like category, prio, etc".

When the agent creates visual or rich content—an HTML page, a React app, a chart, a diagram, a formatted report—it appears as a **chat document** that opens in a side panel next to the chat. Chat documents are scoped to the current conversation and are deleted with the chat. Click the **Chat documents** button in the chat action bar to browse everything created in the current conversation, and promote anything worth keeping to a permanent [agent document](/features/documents).

From the chat side panel you can also browse every document library the agent has access to — its own agent library, its team's library (if any), and the workspace library — making it easy to reference shared materials mid-conversation.

## Rich responses

Agents can respond with more than plain text. They can respond with formatted text, diagrams, images, and voiceovers.

* **Formatted text**: "Make a nice-looking draft blog post based on my messy brainstorm notes"
* **Diagrams**: "Show me a flowchart of your workflow"
* **Images**: "Create a nice-looking infographic of the given meeting notes"
* **Voiceovers** - "Email me a voiceover of this article, using a casual female british accent"
* **Mixed formats**: "Create a nice-looking draft blog post based on my messy brainstorm notes, include a suitable image and visual overview, and a voiceover link."

The visualization below was generated by our own Release Noter agent that monitors GitHub and writes release notes. The prompt was: "Create a nice looking visual timeline of the top 5 features released during the past month, as a compact visual. Make it look like a poster, landscape aspect ratio."

<Frame>
  <img alt="Example of a visual timeline generated by an agent" />
</Frame>

## Interactive applications

Agents can generate interactive applications directly in the chat. This allows agents to create custom interfaces for specific tasks:

* **Dashboards**: "Create a nice-looking interactive dashboard for our OKRs, where you can browse progress per team, and aggregate the progress by quarter."
* **Forms** — "Create a user feedback form that saves data to the user feedback database"

See [Interactive Apps](/features/apps) for more on interactive applications.

<Note>
  Interactive applications are rendered using React components that the agent generates on the fly. The agent can iterate on these based on your feedback.
</Note>

## Code execution

If you enable the **Code execution** capability, the agent can execute code directly in the chat. This is useful for situations where large amounts of data need to be processed, or where the agent needs to perform complex calculations.

* "Calculate the total cost of the project, including materials and labor"
* "Generate a report of the sales data, including charts and tables"
* "Aggregate the sales data in the database and store in a new database called 'Sales by product category'"

This allows the number crunching and data processing to be done using code instead of the LLM, which is faster, cheaper, and more reliable. This is especially useful when working with large datasets or complex calculations.

See [Code Execution and Scripts](/features/scripts) for more on code execution.

## Voice input

You don't have to type—press the microphone button to speak, and your speech is transcribed to text. This works in most languages, and is often a huge time saver.

<Tip>
  You don't need to think carefully about what to say, just ramble your thoughts freely and the agent will transcribe them to text. It may look messy, but the agent will usually understand what you meant.
</Tip>

## Editing sent messages

You can edit a message you already sent if you want to correct it or add missing details. This is useful when a small wording change would otherwise require restarting the discussion.

When you send an edited message, the conversation continues from that updated point and messages below it are removed so the agent can respond to the revised context.

You can also use voice input while editing, so you can dictate changes instead of typing.

## Text-to-speech

If you hover over a chat message, you can click on the "speak" button below to have it read aloud. This is useful for situations such as taking a walk, where you don't want to look at your phone more than necessary. There is also a walk-and-talk mode that is optimized for this use case.

See [Voice communication](/features/voice) for more on voice features, including walk-and-talk mode.

## Multi-user chat

The chat is collaborative—different users in your workspace can talk to the same agent, and multiple people can participate in the same conversation. Responses are live-streamed, so if several users are watching the same chat, they all see the response as it generates.

## Conversation history

All conversations are saved automatically. The sidebar lists every chat you have with the agent, and a filter at the top lets you switch between all chats, **my chats** (chats where you've sent at least one message), and starred chats only. A **List view** link opens a full chats page where you can search, sort, bulk-manage conversations, and apply the same filter.

From the chats page or the sidebar you can:

* Continue a previous conversation by opening it
* Rename a chat
* Star important chats so they stay at the top in a dedicated **Starred** section
* Delete chats individually, or select multiple on the chats page and bulk delete
* Share conversation links with teammates

## Diary

When diary writing is enabled, the agent automatically writes diary entries when significant events happen during a conversation. This is a high-level record of the agent's activities and internal reasoning. Click "Activity" in the sidebar to view the diary entries.

The agent is also able to read its own diary entries, but will only do so if asked, or if it needs some specific information from the diary.

<Tip>
  The agent decides when to write a diary entry based on the conversation content and the agent's internal reasoning. For triggers and other events, diary entries are also written automatically as long as diary writing is enabled for that agent.
</Tip>

See [Activity monitoring](/features/monitoring) for more on diary entries and activity monitoring.

## Model switching

You can switch LLM in the middle of a conversation. For example if you are having a conversation that doesn't require deep reasoning, you can switch to a cheaper model to make the responses faster and cheaper.

See [Model selection](/features/model-selection) for more on model switching.


# Chat Widget
Source: https://docs.abundly.ai/features/chat-widget

Embed an AI-powered chat bubble on any website with a secure backend proxy

Embed your agent as a floating chat bubble on any website. Visitors open the chat, ask questions, and receive streamed AI responses with markdown and images — without leaving your page or needing an Abundly account. Widgets are typically embedded on pages that require login — a customer portal, internal dashboard, or help center — so that only authenticated users can chat. However, you can also use the widget on public-facing pages. A separate daily credit limit can be configured for the widget to control usage independently of the agent's main limit. Every incoming widget message is also screened by built-in attack detection, which blocks prompt-injection and data-exfiltration attempts before a response reaches the visitor. Widget conversations are stored as full platform chats — they appear in the portal chat list alongside regular and Slack conversations, with complete message history and admin tooling.

<Frame>
  <img alt="Chat widget panel showing a streamed conversation with markdown, links, images, and a processing indicator" />
</Frame>

Want to see it in action? The [live widget showcase](https://app.abundly.ai/widget-showcase) has two interactive demos — the ready-made bubble and a fully custom UI built with `createClient`.

The widget has two modes:

* **Ready-made widget** (`mountChatWidget`) — a complete floating chat bubble and panel with streaming responses, a processing indicator (animated dots while the agent works), markdown rendering, and image support. One function call and you're done.
* **Custom UI** (`createClient`) — a headless client that handles backend communication, streaming, and markdown rendering while you build the UI entirely yourself.

If you're using plain HTML, PHP, WordPress, or any non-React stack, you only need a `<script>` tag pointing at the [Abundly-hosted loader](https://app.abundly.ai/widget.js) — no npm install or build step required. If you're building a custom UI or working in a React/Next.js project, install the [`@abundly/widget`](https://www.npmjs.com/package/@abundly/widget) npm package instead. Both approaches share the same backend proxy contract.

<Warning>
  The chat widget is gated behind a workspace feature flag. Contact [support@abundly.ai](mailto:support@abundly.ai) to have it enabled for your workspace.
</Warning>

## How it works

The widget never talks to Abundly directly. A server-side proxy on your backend holds the API key and forwards requests to Abundly server-to-server:

```
Browser  -->  Your backend proxy  --server-to-server-->  Abundly API
              (holds API key)                     X-Agent-Access-Key
```

This keeps the API key out of client-side code. If the widget should only be available to logged-in users, protect the proxy with your existing authentication.

The widget calls three routes on your backend:

| Route            | Method | Forwards to                  |
| ---------------- | ------ | ---------------------------- |
| `/status`        | GET    | `…/widget/status`            |
| `/chat`          | POST   | `…/widget/chat` (SSE stream) |
| `/images/<path>` | GET    | `…/widget/images/<path>`     |

All upstream requests use header `X-Agent-Access-Key: {API_KEY}`. The full upstream URL pattern is `{SERVICE_URL}/agents/{AGENT_ID}/widget/{route}`.

**CORS:** the widget runs in the visitor's browser. If `backendUrl` points to a different origin than the page, your proxy must allow cross-origin requests. A same-origin route (e.g. `/api/abundly` on the same host) avoids this entirely.

## Setting up

<Steps>
  <Step title="Configure the agent">
    In the Abundly portal, open your agent's **Settings → API Access** tab. Turn on **Enable widget** and choose which capabilities the widget may use. The agent's API Capabilities and MCP servers are listed separately in the same picker — like capabilities, they are unavailable in widget conversations unless selected. Then create an **API access key** (starts with `ak_`) in the API Keys section — you'll store this on your server only.

    You can also configure:

    * **Daily credit limit** — a separate cap for widget usage, independent of the agent's main limit
    * **Widget instructions** — dedicated instructions for widget conversations, replacing the agent's main instructions (if empty, the agent's default instructions are used)
    * **Demo response override** — a fixed reply for presentations (skips the LLM)
    * **Model override** — run widget conversations on a specific model

    <Tip>
      Depending on your setup, widget users may range from logged-in employees on an internal dashboard to anonymous visitors on a public page. The primary defense is the **capability list on the widget itself** — only enable what the widget actually needs. As an extra layer against misconfiguration, it's a good idea to use restrictive settings on the agent's capabilities in general. For example, setting **Update Instructions** to **chat only** ensures it is automatically blocked in widget conversations, even if someone accidentally adds it to the widget's capability list.
    </Tip>
  </Step>

  <Step title="Implement the backend proxy">
    Create a server-side endpoint that the widget can call. It must expose `/status`, `/chat`, and `/images/<path>` under a single base URL. Each route forwards the request to Abundly with your API key.

    You need three environment variables (values shown in the portal):

    | Variable              | Description              |
    | --------------------- | ------------------------ |
    | `ABUNDLY_SERVICE_URL` | Abundly service endpoint |
    | `ABUNDLY_AGENT_ID`    | Your agent's ID          |
    | `ABUNDLY_API_KEY`     | API access key (`ak_…`)  |

    See [backend proxy examples](#backend-proxy-examples) below for ready-to-use code in Node.js, Express, Next.js, and PHP.
  </Step>

  <Step title="Embed the widget">
    Add the widget script to your page and mount it, pointing at your backend:

    ```html theme={null}
    <script src="https://app.abundly.ai/widget.js"></script>
    <script>
      Abundly.mountChatWidget({
        backendUrl: "/api/abundly",
        textOptions: {
          bubbleLabel: "Chat with us",
          welcomeMessage: "Hello! How can I help you today?",
        },
      });
    </script>
    ```

    The loader injects the current versioned widget bundle automatically — your snippet stays the same across releases.

    For **React / Next.js**, install the npm package instead:

    ```bash theme={null}
    npm install @abundly/widget
    ```

    ```tsx theme={null}
    "use client";
    import { useEffect } from "react";
    import { mountChatWidget } from "@abundly/widget";

    export default function Page() {
      useEffect(() => {
        const handle = mountChatWidget({ backendUrl: "/api/abundly" });
        return () => handle.unmount();
      }, []);
      return null;
    }
    ```
  </Step>
</Steps>

## Backend proxy examples

Pick your backend language — each example creates a self-contained project you can run locally. The frontend embed code is the same for all of them.

<Tabs>
  <Tab title="Node.js">
    ```bash theme={null}
    mkdir abundly-widget && cd abundly-widget
    mkdir public
    ```

    Create `server.js`:

    ```js theme={null}
    const http = require("http");
    const fs = require("fs");
    const path = require("path");

    const SERVICE_URL = process.env.ABUNDLY_SERVICE_URL;
    const AGENT_ID = process.env.ABUNDLY_AGENT_ID;
    const API_KEY = process.env.ABUNDLY_API_KEY;
    const BASE = `${SERVICE_URL}/agents/${AGENT_ID}/widget`;

    const MIME = { ".html": "text/html", ".js": "application/javascript", ".css": "text/css" };

    const server = http.createServer(async (req, res) => {
      if (req.method === "GET" && req.url === "/api/abundly/status") {
        const resp = await fetch(`${BASE}/status`, {
          headers: { "X-Agent-Access-Key": API_KEY },
        });
        res.writeHead(resp.status, { "Content-Type": "application/json" });
        res.end(await resp.text());
        return;
      }

      if (req.method === "POST" && req.url === "/api/abundly/chat") {
        const chunks = [];
        for await (const chunk of req) chunks.push(chunk);
        const body = Buffer.concat(chunks).toString();

        const resp = await fetch(`${BASE}/chat`, {
          method: "POST",
          headers: { "Content-Type": "application/json", "X-Agent-Access-Key": API_KEY },
          body,
        });
        res.writeHead(resp.status, {
          "Content-Type": "text/event-stream",
          "Cache-Control": "no-cache",
          Connection: "keep-alive",
        });
        const reader = resp.body.getReader();
        const decoder = new TextDecoder();
        while (true) {
          const { done, value } = await reader.read();
          if (done) break;
          res.write(decoder.decode(value, { stream: true }));
        }
        res.end();
        return;
      }

      if (req.method === "GET" && req.url.startsWith("/api/abundly/images/")) {
        const subPath = req.url.substring("/api/abundly/images/".length);
        const resp = await fetch(`${BASE}/images/${subPath}`, {
          headers: { "X-Agent-Access-Key": API_KEY },
        });
        const headers = {};
        ["content-type", "content-length", "cache-control"].forEach((h) => {
          const v = resp.headers.get(h);
          if (v) headers[h] = v;
        });
        res.writeHead(resp.status, headers);
        const reader = resp.body.getReader();
        while (true) {
          const { done, value } = await reader.read();
          if (done) break;
          res.write(value);
        }
        res.end();
        return;
      }

      const filePath = path.join(__dirname, "public", req.url === "/" ? "index.html" : req.url);
      try {
        const data = fs.readFileSync(filePath);
        res.writeHead(200, { "Content-Type": MIME[path.extname(filePath)] || "text/plain" });
        res.end(data);
      } catch {
        res.writeHead(404);
        res.end("Not found");
      }
    });

    server.listen(3000, () => console.log("Server running at http://localhost:3000"));
    ```

    Create `public/index.html`:

    ```html theme={null}
    <!DOCTYPE html>
    <html lang="en">
    <head>
      <meta charset="UTF-8" />
      <meta name="viewport" content="width=device-width, initial-scale=1.0" />
      <title>My Site</title>
    </head>
    <body>
      <h1>My Website</h1>
      <p>The chat widget should appear in the bottom-right corner.</p>

      <script src="https://app.abundly.ai/widget.js"></script>
      <script>
        Abundly.mountChatWidget({
          backendUrl: "/api/abundly",
          textOptions: {
            placeholder:    "Type a message...",
            bubbleLabel:    "Chat with us",
            headerTitle:    "Chat",
            welcomeMessage: "Hello! How can I help you today?",
          },
        });
      </script>
    </body>
    </html>
    ```

    Run:

    ```bash theme={null}
    ABUNDLY_SERVICE_URL=https://service.abundly.ai \
    ABUNDLY_AGENT_ID=your_agent_id \
    ABUNDLY_API_KEY=ak_your_key \
    node server.js
    ```
  </Tab>

  <Tab title="Express.js">
    ```bash theme={null}
    mkdir abundly-widget && cd abundly-widget
    npm init -y
    npm install express
    ```

    Create `server.js`:

    ```js theme={null}
    const express = require("express");
    const path = require("path");
    const app = express();

    const SERVICE_URL = process.env.ABUNDLY_SERVICE_URL;
    const AGENT_ID = process.env.ABUNDLY_AGENT_ID;
    const API_KEY = process.env.ABUNDLY_API_KEY;
    const BASE = `${SERVICE_URL}/agents/${AGENT_ID}/widget`;

    app.use(express.static(path.join(__dirname, "public")));
    app.use(express.json());

    app.get("/api/abundly/status", async (_req, res) => {
      const resp = await fetch(`${BASE}/status`, {
        headers: { "X-Agent-Access-Key": API_KEY },
      });
      res.status(resp.status).json(await resp.json());
    });

    app.post("/api/abundly/chat", async (req, res) => {
      const resp = await fetch(`${BASE}/chat`, {
        method: "POST",
        headers: { "Content-Type": "application/json", "X-Agent-Access-Key": API_KEY },
        body: JSON.stringify(req.body),
      });
      res.writeHead(resp.status, {
        "Content-Type": "text/event-stream",
        "Cache-Control": "no-cache",
        Connection: "keep-alive",
      });
      const reader = resp.body.getReader();
      const decoder = new TextDecoder();
      while (true) {
        const { done, value } = await reader.read();
        if (done) break;
        res.write(decoder.decode(value, { stream: true }));
      }
      res.end();
    });

    app.get("/api/abundly/images/*", async (req, res) => {
      const subPath = req.params[0];
      const resp = await fetch(`${BASE}/images/${subPath}`, {
        headers: { "X-Agent-Access-Key": API_KEY },
      });
      ["content-type", "content-length", "cache-control"].forEach((h) => {
        const v = resp.headers.get(h);
        if (v) res.setHeader(h, v);
      });
      res.status(resp.status);
      const reader = resp.body.getReader();
      while (true) {
        const { done, value } = await reader.read();
        if (done) break;
        res.write(value);
      }
      res.end();
    });

    app.listen(3000, () => console.log("Server running at http://localhost:3000"));
    ```

    Create `public/index.html`:

    ```html theme={null}
    <!DOCTYPE html>
    <html lang="en">
    <head>
      <meta charset="UTF-8" />
      <meta name="viewport" content="width=device-width, initial-scale=1.0" />
      <title>My Site</title>
    </head>
    <body>
      <h1>My Website</h1>
      <p>The chat widget should appear in the bottom-right corner.</p>

      <script src="https://app.abundly.ai/widget.js"></script>
      <script>
        Abundly.mountChatWidget({
          backendUrl: "/api/abundly",
          textOptions: {
            placeholder:    "Type a message...",
            bubbleLabel:    "Chat with us",
            headerTitle:    "Chat",
            welcomeMessage: "Hello! How can I help you today?",
          },
        });
      </script>
    </body>
    </html>
    ```

    Run:

    ```bash theme={null}
    ABUNDLY_SERVICE_URL=https://service.abundly.ai \
    ABUNDLY_AGENT_ID=your_agent_id \
    ABUNDLY_API_KEY=ak_your_key \
    node server.js
    ```
  </Tab>

  <Tab title="Next.js">
    ```bash theme={null}
    npx create-next-app@latest abundly-widget --app --typescript
    cd abundly-widget
    npm install @abundly/widget
    ```

    Create `app/api/abundly/[...path]/route.ts`:

    ```ts theme={null}
    const SERVICE_URL = process.env.ABUNDLY_SERVICE_URL!;
    const AGENT_ID = process.env.ABUNDLY_AGENT_ID!;
    const API_KEY = process.env.ABUNDLY_API_KEY!;
    const BASE = `${SERVICE_URL}/agents/${AGENT_ID}/widget`;

    export async function GET(
      _req: Request,
      context: { params: Promise<{ path: string[] }> },
    ) {
      const { path } = await context.params;
      const [first, ...rest] = path;

      if (first === "status") {
        const resp = await fetch(`${BASE}/status`, {
          headers: { "X-Agent-Access-Key": API_KEY },
        });
        return Response.json(await resp.json(), { status: resp.status });
      }

      if (first === "images") {
        const resp = await fetch(`${BASE}/images/${rest.join("/")}`, {
          headers: { "X-Agent-Access-Key": API_KEY },
        });
        const headers = new Headers();
        ["content-type", "content-length", "cache-control"].forEach((h) => {
          const v = resp.headers.get(h);
          if (v) headers.set(h, v);
        });
        return new Response(resp.body, { status: resp.status, headers });
      }

      return Response.json({ error: "Not found" }, { status: 404 });
    }

    export async function POST(
      req: Request,
      context: { params: Promise<{ path: string[] }> },
    ) {
      const { path } = await context.params;
      if (path[0] !== "chat")
        return Response.json({ error: "Not found" }, { status: 404 });

      const resp = await fetch(`${BASE}/chat`, {
        method: "POST",
        headers: {
          "Content-Type": "application/json",
          "X-Agent-Access-Key": API_KEY,
        },
        body: JSON.stringify(await req.json()),
      });

      return new Response(resp.body, {
        status: resp.status,
        headers: {
          "Content-Type": "text/event-stream",
          "Cache-Control": "no-cache",
        },
      });
    }
    ```

    Add to `.env.local`:

    ```
    ABUNDLY_SERVICE_URL=https://service.abundly.ai
    ABUNDLY_AGENT_ID=your_agent_id
    ABUNDLY_API_KEY=ak_your_key
    ```

    Mount the widget in `app/page.tsx`:

    ```tsx theme={null}
    "use client";
    import { useEffect } from "react";
    import { mountChatWidget } from "@abundly/widget";

    export default function Home() {
      useEffect(() => {
        const handle = mountChatWidget({
          backendUrl: "/api/abundly",
          textOptions: {
            bubbleLabel: "Chat with us",
            welcomeMessage: "Hello! How can I help you today?",
          },
        });
        return () => handle.unmount();
      }, []);

      return (
        <main>
          <h1>My Website</h1>
          <p>The chat widget should appear in the bottom-right corner.</p>
        </main>
      );
    }
    ```

    Run with `npm run dev`.
  </Tab>

  <Tab title="PHP">
    ```bash theme={null}
    mkdir abundly-widget && cd abundly-widget
    mkdir api
    ```

    Create `api/abundly.php`:

    ```php theme={null}
    <?php
    $SERVICE_URL = getenv("ABUNDLY_SERVICE_URL");
    $AGENT_ID    = getenv("ABUNDLY_AGENT_ID");
    $API_KEY     = getenv("ABUNDLY_API_KEY");
    $BASE        = "$SERVICE_URL/agents/$AGENT_ID/widget";

    $path   = parse_url($_SERVER["REQUEST_URI"], PHP_URL_PATH);
    $method = $_SERVER["REQUEST_METHOD"];

    if ($method === "GET" && $path === "/api/abundly/status") {
        $ch = curl_init("$BASE/status");
        curl_setopt_array($ch, [
            CURLOPT_HTTPHEADER     => ["X-Agent-Access-Key: $API_KEY"],
            CURLOPT_RETURNTRANSFER => true,
        ]);
        $body = curl_exec($ch);
        $code = curl_getinfo($ch, CURLINFO_HTTP_CODE);
        curl_close($ch);
        http_response_code($code);
        header("Content-Type: application/json");
        echo $body;

    } elseif ($method === "POST" && $path === "/api/abundly/chat") {
        $input = file_get_contents("php://input");
        header("Content-Type: text/event-stream");
        header("Cache-Control: no-cache");
        header("Connection: keep-alive");

        $ch = curl_init("$BASE/chat");
        curl_setopt_array($ch, [
            CURLOPT_POST       => true,
            CURLOPT_POSTFIELDS => $input,
            CURLOPT_HTTPHEADER => [
                "Content-Type: application/json",
                "X-Agent-Access-Key: $API_KEY",
            ],
            CURLOPT_WRITEFUNCTION => function ($ch, $data) {
                echo $data;
                if (ob_get_level() > 0) ob_flush();
                flush();
                return strlen($data);
            },
        ]);
        curl_exec($ch);
        curl_close($ch);

    } elseif ($method === "GET" && strpos($path, "/api/abundly/images/") === 0) {
        $subPath = substr($path, strlen("/api/abundly/images/"));
        $ch = curl_init("$BASE/images/$subPath");
        curl_setopt_array($ch, [
            CURLOPT_HTTPHEADER    => ["X-Agent-Access-Key: $API_KEY"],
            CURLOPT_HEADERFUNCTION => function ($ch, $header) {
                if (preg_match("/^(content-type|content-length|cache-control):/i", $header))
                    header(trim($header));
                return strlen($header);
            },
            CURLOPT_WRITEFUNCTION => function ($ch, $data) {
                echo $data;
                return strlen($data);
            },
        ]);
        curl_exec($ch);
        http_response_code(curl_getinfo($ch, CURLINFO_HTTP_CODE));
        curl_close($ch);

    } else {
        http_response_code(404);
        echo "Not found";
    }
    ```

    Create `router.php`:

    ```php theme={null}
    <?php
    if (preg_match("#^/api/abundly/(status|chat|images/)#", $_SERVER["REQUEST_URI"])) {
        require __DIR__ . "/api/abundly.php";
        return true;
    }
    return false;
    ```

    Create `index.html` in the project root:

    ```html theme={null}
    <!DOCTYPE html>
    <html lang="en">
    <head>
      <meta charset="UTF-8" />
      <meta name="viewport" content="width=device-width, initial-scale=1.0" />
      <title>My Site</title>
    </head>
    <body>
      <h1>My Website</h1>
      <p>The chat widget should appear in the bottom-right corner.</p>

      <script src="https://app.abundly.ai/widget.js"></script>
      <script>
        Abundly.mountChatWidget({
          backendUrl: "/api/abundly",
          textOptions: {
            placeholder:    "Type a message...",
            bubbleLabel:    "Chat with us",
            headerTitle:    "Chat",
            welcomeMessage: "Hello! How can I help you today?",
          },
        });
      </script>
    </body>
    </html>
    ```

    Run:

    ```bash theme={null}
    ABUNDLY_SERVICE_URL=https://service.abundly.ai \
    ABUNDLY_AGENT_ID=your_agent_id \
    ABUNDLY_API_KEY=ak_your_key \
    php -S localhost:8080 router.php
    ```
  </Tab>
</Tabs>

Open `http://localhost:3000` (port `8080` for PHP). You should see the chat bubble in the bottom-right corner. Click it, type a message, and see the streamed response from your agent.

<Tip>
  These examples are minimal — they forward requests to Abundly but do not authenticate the browser caller. In production, add your own auth middleware (session cookies, JWT verification, etc.) to protect these endpoints.
</Tip>

## Configuration

`mountChatWidget` requires only `backendUrl` — everything else is optional. Use `textOptions` to customize labels and `uiOptions` to match your site's colors.

### Text options

All text options are optional and have sensible defaults.

| Property         | Default             | Description                                         |
| ---------------- | ------------------- | --------------------------------------------------- |
| `placeholder`    | "Type a message..." | Input field placeholder text                        |
| `bubbleLabel`    | "Chat about this"   | Text on the floating bubble                         |
| `headerTitle`    | "Chat"              | Title in the chat panel header                      |
| `welcomeMessage` | *(none)*            | Initial assistant message shown when the chat opens |

### UI options

All color options are optional and default to brand colors. Values accept any valid CSS — named colors, hex, rgba, or gradients.

| Property              | Default        | Description                        |
| --------------------- | -------------- | ---------------------------------- |
| `bubbleColor`         | Brand gradient | Floating bubble background         |
| `bubbleHoverColor`    | —              | Bubble background on hover         |
| `bubbleTextColor`     | `#fff`         | Text and icon color on the bubble  |
| `userBubbleColor`     | `#319795`      | User message bubble background     |
| `inputBorderColor`    | `#319795`      | Text input border when focused     |
| `submitColor`         | Brand gradient | Send button background             |
| `submitHoverColor`    | —              | Send button on hover               |
| `submitDisabledColor` | —              | Send button when disabled          |
| `headerColor`         | —              | Chat panel header background       |
| `headerTextColor`     | `#888`         | Header text and close button color |
| `linkColor`           | `#0d7a6e`      | Link color in assistant messages   |

## What the ready-made widget handles

`mountChatWidget` takes care of these concerns automatically:

* **Availability check** — calls `/status` before showing the bubble; stays hidden if the agent is unavailable or the daily limit is exhausted
* **SSE streaming** — character-by-character reveal animation as the response arrives
* **Processing indicator** — animated dots while the agent is working (tools, sub-agents, or waiting for a response). Internal reasoning text is never shown in the widget.
* **Markdown rendering** — safe HTML output with XSS protection
* **Image support** — URL rewriting, preloading, and authenticated image loading through the proxy
* **Mobile layout** — soft-keyboard detection and layout adjustment
* **Auto-resizing input** — textarea grows up to three lines
* **Rate limiting** — displays errors and suspends input on HTTP 429

`mountChatWidget` returns a handle with two methods:

| Method             | Description                                                                                           |
| ------------------ | ----------------------------------------------------------------------------------------------------- |
| `handle.reset()`   | Clears conversation history, errors, and suspended state. Restores the welcome message if configured. |
| `handle.unmount()` | Removes bubble, panel, backdrop, and injected styles. Idempotent, safe to call at any time.           |

## Custom UI with createClient

If you need full control over the chat interface — your own layout, animations, or interaction model — use `createClient` instead of `mountChatWidget`. It provides the same streaming and protocol engine without any DOM or styles.

```ts theme={null}
import { createClient } from "@abundly/widget";

const client = createClient("/api/abundly");
```

Or via script tag:

```html theme={null}
<script src="https://app.abundly.ai/widget.js"></script>
<script>
  const client = Abundly.createClient("/api/abundly");
</script>
```

The client exposes, among other things:

* **`client.checkStatus()`** — check if the agent is available
* **`client.sendMessage(text, callbacks)`** — stream a response via SSE with `onChunk`, `onProcessing`, `onDone`, and `onStreamError` callbacks
* **`client.resetSession()`** — clear the server-side session so the next message starts a new conversation
* **`client.renderMarkdown(text)`** — convert markdown to safe HTML (optional — use your own library instead if you prefer)
* **`client.stripPartialMd(text)`** — clean up half-formed markdown tokens during streaming

Conversation history is persisted server-side — you only need a local `Message[]` for rendering your UI. The client handles protocol details, session management, URL rewriting, and optionally markdown rendering. You own all markup and styling.

<Note>
  The methods listed above are the most commonly used — the full client API includes additional helpers for image authentication, blob URL management, and more. See the [`@abundly/widget` package on npm](https://www.npmjs.com/package/@abundly/widget) for the complete reference.
</Note>

The usual flow for a custom UI:

1. **Backend first** — same proxy as above (`/status`, `/chat`, `/images/<path>`)
2. **Install the SDK** — `npm install @abundly/widget`
3. **Create a client** — `const client = createClient("/api/abundly")`
4. **Build your component** — use `client.sendMessage` for streaming, `client.renderMarkdown` for replies, `client.checkStatus` for availability. Authenticated proxies and inline images require additional client helpers — see the [npm package docs](https://www.npmjs.com/package/@abundly/widget) for details
5. **Ship it** — mount wherever it should live; no `mountChatWidget` required

<Accordion title="Minimal React example">
  ```tsx theme={null}
  "use client";
  import { useState } from "react";
  import { createClient, type Message } from "@abundly/widget";

  const client = createClient("/api/abundly");

  export function MyChat() {
    const [messages, setMessages] = useState<Message[]>([]);
    const [streaming, setStreaming] = useState("");
    const [draft, setDraft] = useState("");

    async function send() {
      const text = draft.trim();
      if (!text) return;
      setDraft("");
      setMessages((prev) => [...prev, { role: "user", content: text }]);
      setStreaming("");
      // No need to pass conversation history — the server tracks it
      // via a session ID that the client manages automatically.
      const final = await client.sendMessage(text, {
        onChunk: (accumulated) => setStreaming(accumulated),
      });
      setStreaming("");
      setMessages((prev) => [...prev, { role: "assistant", content: final }]);
    }

    function resetChat() {
      client.resetSession();
      setMessages([]);
    }

    return (
      <div>
        {messages.map((m, i) => (
          <div key={i} className={m.role}>
            {m.role === "user" ? (
              m.content
            ) : (
              <div dangerouslySetInnerHTML={{
                __html: client.renderMarkdown(m.content),
              }} />
            )}
          </div>
        ))}
        {streaming && (
          <div className="assistant">
            <div dangerouslySetInnerHTML={{
              __html: client.renderMarkdown(client.stripPartialMd(streaming)),
            }} />
          </div>
        )}
        <input value={draft} onChange={(e) => setDraft(e.target.value)} />
        <button onClick={send}>Send</button>
        <button onClick={resetChat}>New chat</button>
      </div>
    );
  }
  ```
</Accordion>

For the full API reference, implementation help, TypeScript types, image auth helpers, and markdown renderer comparison, see [`@abundly/widget` on npm](https://www.npmjs.com/package/@abundly/widget).

## Authentication

If your proxy is same-origin and uses session cookies, no extra configuration is needed — cookies flow automatically.

For token-based or cross-origin setups, pass `headers` and/or `credentials` on the mount call. Both `mountChatWidget` and `createClient` accept the same options.

<Tabs>
  <Tab title="Bearer / JWT">
    Use the function form so the token is re-read on every request:

    ```tsx theme={null}
    mountChatWidget({
      backendUrl: "/api/abundly",
      headers: async () => ({
        Authorization: `Bearer ${await getAccessToken()}`,
      }),
    });
    ```
  </Tab>

  <Tab title="Static header">
    Use the object form when the value doesn't change:

    ```tsx theme={null}
    mountChatWidget({
      backendUrl: "/api/abundly",
      headers: { "X-API-Key": "ak_live_..." },
    });
    ```
  </Tab>

  <Tab title="Cross-origin cookies">
    Set `credentials: "include"` and ensure the proxy responds with `Access-Control-Allow-Credentials: true`:

    ```tsx theme={null}
    mountChatWidget({
      backendUrl: "https://api.example.com/abundly",
      credentials: "include",
    });
    ```
  </Tab>
</Tabs>

<Note>
  `headers` is applied to every request the widget makes (`/status`, `/chat`, and image loads). User-supplied headers override built-in headers with the same name. `credentials` maps directly to the `fetch` option.
</Note>

## Learn more

<CardGroup>
  <Card title="API Access" icon="key" href="/features/agent-api-endpoint">
    Configure API keys, HTTP endpoints, MCP, and webhooks alongside the widget
  </Card>

  <Card title="@abundly/widget on npm" icon="cube" href="https://www.npmjs.com/package/@abundly/widget">
    Full API reference, TypeScript types, and advanced customization
  </Card>

  <Card title="Monitoring" icon="gauge-high" href="/features/monitoring">
    Review widget conversations in the portal chat list and monitor credit usage
  </Card>
</CardGroup>


# Multi-channel Communication
Source: https://docs.abundly.ai/features/communication

Agents communicate through chat, email, SMS, voice calls, and more

Agents are not stuck in a chat window—they communicate through multiple channels depending on what's most convenient for you and the task at hand.

This flexibility means an agent can receive a request via email, update you on Slack, and call you if something urgent needs your attention. You decide which channels your agent can use.

## Available channels

| Channel                | Description                                                               | Setup                                                |
| ---------------------- | ------------------------------------------------------------------------- | ---------------------------------------------------- |
| **Chat**               | Feature-rich AI chat with voice input, file uploads, and interactive apps | Built-in, always available                           |
| **Email**              | Send and receive emails from the agent's own address                      | Built-in, toggle to enable                           |
| **SMS**                | Send and receive text messages                                            | Beta — contact support                               |
| **Voice calls**        | Make and receive phone calls with real-time conversation                  | Beta — contact support                               |
| **WhatsApp**           | Hold 1:1 WhatsApp conversations from a dedicated number                   | Requires [WhatsApp setup](/integrations/whatsapp)    |
| **Slack, Teams, etc.** | Communicate through collaboration tools                                   | Requires [integration setup](/integrations/overview) |

<Note>
  By default, only chat is enabled. This is a security measure—you decide how your agent is allowed to communicate. Go to **Settings → Capabilities** to enable additional channels.
</Note>

## Enabling communication channels

If you ask your agent to use a communication channel that isn't enabled, it will show you a capability card where you can toggle it on.

You can also proactively enable channels in **Settings → Capabilities** before you need them.

<Tip>
  If your agent has multiple communication channels enabled, it will make a best guess about which channel to use. For more control, give guidance in your agent's [instructions](/features/instructions) or during conversation—for example, "Always email me weekly reports" or "Only call me for urgent issues."
</Tip>

### Example use cases

* "Every Monday morning, email me a summary of last week's incoming book manuscripts"
* "Post a summary of this conversation to the #engineering Slack channel"
* "When you receive support tickets via email: Check the knowledgebase. For simple questions with obvious answers, reply yourself. Otherwise, text me and create a ticket in Notion."
* "When a new lead fills out the contact form, research their company, send them a personalized welcome email using the template I gave you, and post a summary to #sales on Slack."

## Multi-modal content

Beyond text, agents can work with different types of media:

* **Images** — Generate, analyze, and include images in messages
* **Audio** — Transcribe recordings, generate voiceovers
* **Documents** — Read and create PDFs, presentations, spreadsheets
* **Diagrams** — Generate visual diagrams and charts

### Generating voiceovers

Your agent can generate audio files and embed them in emails, documents, and messages. This is useful when you want content that can be consumed on the go—during a commute, while exercising, or between meetings.

**How to enable:** Go to **Settings → Capabilities** and enable **Text to Speech**.

**Example use cases:**

* "Generate a voiceover for this script using a male Scottish voice"
* "Every morning, generate a 'daily briefing' that I can listen to while driving to work"
* "In the weekly competitive analysis email, include an audio clip of the executive summary"

Here is an example of a voiceover embedded in an email report:

<Frame>
  <img alt="Weekly competitor analysis with embedded audio summary" />
</Frame>

Voice generation is powered by [ElevenLabs](/integrations/elevenlabs) and [OpenAI](/integrations/openai), giving you natural-sounding voices in multiple languages.

### Other examples

* "When you receive an email with a voice memo attached, transcribe it and add the notes to my meeting log"
* "When a customer sends a photo of a damaged product, analyze it and draft a response"

See [Document management](/features/documents) for more details on working with files.

## Learn more

<CardGroup>
  <Card title="Advanced chat" icon="message" href="/features/chat">
    Rich chat interface with voice input, file uploads, and interactive apps
  </Card>

  <Card title="Email & SMS" icon="envelope" href="/features/messaging">
    Send and receive emails and text messages
  </Card>

  <Card title="Voice" icon="phone" href="/features/voice">
    Voice input, text-to-speech, and phone calls
  </Card>
</CardGroup>


# Agent Documents
Source: https://docs.abundly.ai/features/documents

Agents can read, edit, and create documents, apps, databases, and executable scripts

Documents are first-class storage in Abundly. Each workspace, each team, and each agent has its own document library, and your agents can read across all of them—just like a colleague who has access to a shared drive plus their own files. The platform tracks every change with full version history.

Documents can be pure content, such as a text file or PDF or image, or they can be [Interactive Apps](/features/apps) or [Agent Databases](/features/agent-databases).

<Frame>
  <img alt="Documents panel showing a list of agent documents organized in folders" />
</Frame>

Documents work out of the box with no configuration. For external files like Google Drive or SharePoint, enable the corresponding [capabilities](/features/capabilities).

## Where documents live

Documents are organized into three scopes, mirroring how teams share information in real life:

* **Workspace documents** — Shared with everyone in your workspace. Find them on the **Workspace** page under the **Documents** tab. By default, only workspace admins can add, remove, and edit files; admins can adjust this on the Documents tab to also let all members and/or agents edit.
* **Team documents** — Shared within a single [team](/features/teams). Find them on the team's home page under **Documents**. By default, only team admins can edit; team admins can adjust this on the Documents tab to also let all team members and/or team agents edit. Workspace admins can always manage any team's documents.
* **Agent documents** — Belong to a single agent. Find them in the agent's **Documents** page. Anyone with edit access to the agent can manage them.

When working, an agent can read documents from all three scopes—its own, its team's, and the workspace. By default, it can only create or edit documents in its own library, but workspace and team admins can grant agents edit access to those libraries too—either all agents, or only [admin-restricted agents](/security/access-control#admin-restricted-agents).

The full-page **Documents** views (workspace, team, and agent) all use the same tree-based file explorer. The agent's Documents page shows the workspace and team scopes the agent can reach alongside its own library, so you can browse everything the agent works with from one place.

You can move a document between scopes from its menu, or by **dragging it** onto another scope or folder in the file explorer (for example, promote an agent document to a workspace document so other agents can use it). On desktop you can also drop files from your computer directly onto any scope to upload them there.

## Supported file types

Documents are plain text by default—typically Markdown for simple formatting. You can also upload other file types:

| File type                  | How it works                                                                    |
| -------------------------- | ------------------------------------------------------------------------------- |
| **Text & Markdown**        | Native format, directly editable by you and the agent                           |
| **Code files**             | JavaScript, TypeScript, HTML, CSS, JSON, YAML, and more—stored as editable text |
| **SVG graphics**           | Stored as editable text, rendered as vector graphics in preview                 |
| **PDFs, Word, PowerPoint** | Transcribed to text—original file preserved for reference                       |
| **Spreadsheets**           | Transcribed to text—original file preserved for reference                       |
| **Images**                 | Vision AI generates a text description the agent can read                       |
| **Audio**                  | Automatically transcribed to text                                               |
| **Video**                  | Audio track extracted and transcribed (video frames not analyzed)               |

<Note>
  For uploaded files (PDF, audio, etc.), the platform creates a text version that the agent works with. Transcription runs **automatically in the background** as soon as the file is uploaded to an agent, workspace, or team library, and a short description is auto-generated afterwards so the file is easier to find later. You can access the original via the **Original file** button. If you edit the text, it will diverge from the original—the agent always works with the text version.
</Note>

## Working with documents

Click any document to open it in a modal with a mode selector. Available modes depend on document type and context.

<Tip>
  Use **Copy link** from a document's menu to grab a direct link to it. Anyone
  with access to the agent — or to the workspace/team scope the document lives in
  — can open the link straight to that document or app, making it easy to
  reference a specific item in chat, Slack, or a ticket. This is different from
  [publishing](#publishing-documents), which creates a public link that needs no
  account.
</Tip>

### Text mode

The text representation of the document—this is what the agent sees and works with. You can edit the text directly here.

### View mode

A visual rendering of the document, depending on its type:

* **Markdown** → Rendered with formatting, headings, and links
* **React or HTML** → Live rendered application or web page
* **Mermaid diagrams** → Rendered flowcharts, sequence diagrams, etc.
* **SVG** → Rendered vector graphic (editable as text)
* **Database documents** → Interactive table view of the data

Here is an example of a preview of an [Interactive App](/features/apps) document.

<Frame>
  <img alt="Document view mode showing a live Launch Dashboard app with countdown timer, metrics, charts, and progress bars" />
</Frame>

### Data mode (for databases)

For [agent databases](/features/agent-databases), you can open a dedicated **Data** mode to view and edit raw JSON plus the schema.

### Changes mode

A version timeline showing every change made to the document. Click any version to view it, or click between two versions to see a diff of what changed. You can approve or undo individual changes, or revert to any previous version.

<Frame>
  <img alt="Changes mode showing version timeline and inline diff with additions and removals highlighted" />
</Frame>

## Visibility levels

Documents have two separate controls:

* **Visibility to the agent** controls what your agent can discover and read while working.
* **Publishing** controls whether people outside your workspace can open a public link.

By default, agents know which documents they have (name and summary) and can read the full content when needed. You can adjust visibility per document:

| Level                                | Behavior                                                                                                                                      |
| ------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------- |
| <Icon icon="list" /> **Full**        | Agent always knows the full content—included in every prompt alongside instructions. Use for essential context the agent should never forget. |
| <Icon icon="eye" /> **Summary**      | (default value) Agent knows the document exists and its summary. Reads the full content only when relevant. Good for most documents.          |
| <Icon icon="eye-slash" /> **Hidden** | Agent cannot see or find this document. Use for your own notes or materials not relevant to the agent's job.                                  |

To change visibility, click the visibility icon next to the document name in the Documents panel.

<Frame>
  <img alt="Document visibility level dropdown in the Edit Document modal" />
</Frame>

<Tip>
  Start with the default **Summary** visibility. You can add many documents without overloading the agent's context—it will read full content only when needed.
</Tip>

### Folder visibility

Folders have their own visibility control with just **Visible** and **Hidden** (folders have no content of their own, so **Full** doesn't apply). Click the visibility icon next to a folder name to change it.

Hiding a folder cascades to everything inside it:

* Every document and sub-folder underneath is treated as **Hidden** for the agent, regardless of the child's own visibility setting.
* The folder and its contents disappear from the agent's system prompt, from search results, and from folder browsing tools.
* In the Documents panel and file explorer, descendant documents and sub-folders show their visibility badge greyed out, with a tooltip explaining that an ancestor folder is hiding them — so you can tell at a glance which items are hidden because of a folder rather than their own setting.
* People who can see the workspace still see the folder normally in the Documents panel and the in-chat file explorer—only the agent's view is affected.

Use folder hiding when you want to keep a whole area of materials (drafts, archives, personal notes) out of an agent's reach without flipping visibility on each file.

## Adding documents

You can add documents in three ways. The flow is the same for workspace, team, and agent documents—open the corresponding **Documents** page or tab and use the same controls.

**1. Upload or create directly**

Each Documents view has **Create document** and **Upload document** buttons at the top—use them to start a new text document or upload files from your computer. You can also drag files straight onto the list. Each section and folder still has a kebab menu with **New Folder**, **New Document**, and **Upload** when you want to add content into a specific folder, and individual documents and folders have a row-level menu (⋮) with **Rename**, **Move**, and **Delete** so you can manage items right from the list without opening them first.

**2. Save from a conversation**

Drag a file into the chat, discuss it with the agent, then ask it to save the result. For example:

* Drag a meeting transcript, ask for a summary, then "save this as an agent document"
* Ask the agent to draft a newsletter, refine it together, then "save this to the newsletters folder"

<Tip>
  When the agent opens content in the side panel (a chat-scoped document), click the menu and choose **Save as agent document** to promote it to a permanent agent document without asking the agent to do it. The same menu also offers **Save to Google Drive**, which creates a native Google Doc in your My Drive.
</Tip>

**3. Instruct the agent to create documents automatically**

Add instructions like:

* "When processing an invoice, add it to the processed invoices database"
* "Save all incoming emails to a message log"
* "After posting release notes, save a copy to the 'release notes' folder"
* "When you receive an email with an attachment, save it as an agent document"

**4. Download from a public URL**

Ask the agent to fetch a file from a URL and save it—for example, "Download the PDF at [https://example.com/report.pdf](https://example.com/report.pdf) and save it to the reports folder." Images, PDFs, Markdown, and plain text are all supported.

## Finding documents

Each Documents view—agent, workspace, or team—has a free-text **search** field at the top. It matches by document name, description, or type, and results appear as a flat list with folder-path breadcrumbs so you can see where each match lives. The search is scoped to the current section, so a workspace-documents search won't surface agent or team documents.

## Organizing with folders

Documents can be organized into folders. Create folders in the Documents panel and drag documents to organize them. Folders can be nested as deeply as you need, and nesting works the same way across workspace, team, agent, and chat document libraries. Reference folders in your agent's instructions: "Save meeting notes to the 'meetings' folder."

Deleting a folder that still contains documents or sub-folders prompts for confirmation and shows what will be removed, so you don't lose files by mistake.

## Version control

Every edit—by you or the agent—creates a new version automatically. The Changes mode lets you:

* **Browse history** — See all versions with timestamps and who made each change
* **Compare versions** — Side-by-side diff view to see exactly what changed
* **Approve or reject** — Optionally mark changes as approved or rejected (rejecting creates a new version with that change undone)
* **Revert** — Roll back to any previous version if needed

## Publishing documents

You can publish a document to make it available to anyone with the link. Open the document menu and choose **Publish** (or **Publish details** if it is already published). This works for documents in any scope—workspace, team, or agent.

Anyone with the link can view the document—no Abundly account required. If it is an app, they can interact with it. Published documents are marked with a **Published** badge in the Documents view. Logged-in visitors can also use **Copy to my agent** from the viewer to clone the published document into one of their own agents—handy for sharing templates, dashboards, or reference material across workspaces.

Use cases:

* Share a live dashboard with colleagues
* Publish an input form for data collection
* Share reports, drafts, or documentation externally
* Create a public-facing interactive app

You can also publish [Interactive Apps](/features/apps) and [Agent Databases](/features/agent-databases). For agent databases, choose what public users can do:

* **Read** to query and view data
* **Create** to insert new items
* **Update** to modify existing items
* **Delete** to remove items

<Warning>
  Publishing a database exposes its entire contents. The Read/Create/Update/Delete toggles control which **operations** public viewers can perform; they don't restrict which rows are visible. See [publishing apps with data](/features/apps#publishing-apps-with-data) for details.
</Warning>

For uploaded files (like PDF, audio, or spreadsheets), you can optionally enable **Also publish original file** so anyone with the link can download the original file without logging in.

If a published app depends on other documents, publish those documents too (with at least **Read** enabled) so the app can access them from the public link.

<Frame>
  <img alt="Publish document dialog" />
</Frame>

<Tip>
  Published apps update automatically. Create a dashboard, share the link, and your colleagues always see the latest data.
</Tip>

## Editing and searching documents

Agents don't need to rewrite entire documents to make changes. They can:

* **Edit precisely** — Replace, insert, or append text
* **Search within a document** — Exact match, full-text, or semantic (meaning-based) search
* **Find documents** — Fuzzy keyword matching across names, descriptions, and content
* **Rename and describe** — Update document names and summaries

<Tip>
  Use semantic search within documents to find content by meaning, not just keywords. For example, searching "global warming" can find sections about "climate change."
</Tip>

## Downloading documents

You can download individual documents in several formats:

* **Download text** — Get the text content as Markdown, plain text, or other formats
* **Download as MS Word** — Convert Markdown content to a `.docx` file
* **Download as PDF** — Convert a document to a PDF. Markdown, rendered documents (HTML, SVG, React apps), and embedded images are all included.
* **Download original** — For non-text files (PDF, audio, etc.), download the original file

To download multiple documents at once, use **Download all** in the Documents panel to get a ZIP file containing everything in the current folder, including any nested subfolders. You can also choose **Download folder as ZIP** from a folder's menu to download just that folder's contents.

On the full-page Documents views, you can also multi-select items with the row checkboxes and then bulk **move**, **download**, or **delete** the whole selection in one go.

To delete multiple documents at once, use **Delete all** from the Documents panel menu to remove everything in the current view.

## Documents as extended context

Agent documents are a powerful way to give your agent additional context without bloating its [instructions](/features/instructions). Instead of pasting everything into the instructions, store reference materials as documents:

* **Checklists** — Step-by-step procedures for specific tasks
* **Templates** — Formats for reports, emails, or communications
* **Guidelines** — Company policies, compliance rules, or style guides
* **Reference data** — Product catalogs, contact directories, FAQs

With **Summary** visibility (the default), the agent knows these documents exist and reads them when relevant—keeping its context focused while still having access to everything it needs.

For best results, guide your agent on when to use specific documents:

* "Before responding to compliance inquiries, read the compliance guidelines document first"
* "When drafting emails, use the email template in the templates folder"
* "All meeting notes are saved by date—search for them when you need context"

<Tip>
  If you don't want the agent to create or edit documents, disable the **Edit Documents** capability in settings.
</Tip>

## Chatting about documents

Click **Chat about this** next to any document or folder to start a new conversation with that content as context. The agent has immediate access to the selected document(s), making it easy to ask questions, request edits, or work through the content together.

## FAQ

<AccordionGroup>
  <Accordion title="Where are my documents stored?">
    It depends on the document type:

    * **Agent documents** (created directly in Abundly) are stored in EU data centers in Stockholm, Sweden. They remain fully under your control—you can delete them at any time.
    * **External documents** (Google Drive, SharePoint, etc.) remain on their original provider's servers. The agent accesses them through your permission but doesn't copy them to Abundly's infrastructure.

    For example, if you share a Google Doc with an agent, the document stays on Google's servers and is accessed only through your authorized connection.

    See [Infrastructure & Compliance](/security/infrastructure) for details on data residency and security.
  </Accordion>
</AccordionGroup>


# Agent Evals
Source: https://docs.abundly.ai/features/evals

Test and evaluate your agents with automated test cases

You've built an agent that works. But what happens when you tweak the instructions? Change the model? Add a new capability?

Agent Evals let you define "working correctly" once, then verify it automatically anytime. Instead of manually testing after every change, run your evals and know in seconds if something broke—or improved.

<Frame>
  <img alt="Evals tab showing a complete eval suite for a target screening agent" />
</Frame>

## Why use evals

**Catch regressions.** You improve one behavior and accidentally break another. Evals catch this before your users do.

**Compare models.** Run the same tests across Claude, GPT, and Gemini to find the best fit for your use case—or discover a cheaper model works just as well.

**Ensure consistency.** Unlike manual testing, evals check the same criteria every time. No more "it worked when I tried it."

**Iterate with confidence.** Refine instructions, adjust capabilities, experiment freely—evals tell you if you're making progress.

## Quick start

The fastest way to understand evals is to create one.

1. Navigate to your agent's **Evals** tab and click **Add Eval**
2. Give it a name: "Knows its own name"
3. Enter a trigger prompt: "What is your name?"
4. Enter a grader prompt: "Pass if the agent identifies itself as **Scout** (or whatever your agent is called). Fail if it gives a different name or says it doesn't have one."
5. Click **Save**, then click **Run**

Check the **Eval Results** tab. You'll see whether it passed or failed, and the grader's reasoning for its decision.

That's the core of evals: define what you expect, run the test, verify the result. Everything else—validation rules, tool overrides, model comparison—builds on this foundation.

<Tip>
  Replace "Scout" with your agent's actual name. This is what makes graders deterministic—you include the expected answer.
</Tip>

## How evals work

Each eval follows a simple flow:

1. **Trigger** — You define a prompt that simulates a user message or scenario
2. **Execute** — The agent processes the prompt using its regular configuration
3. **Validate** — Optional validation rules check tool usage (was tool X called?)
4. **Grade** — An optional grader LLM evaluates the response against your criteria

Validation rules give you fast, deterministic checks. Grading gives you nuanced evaluation. You can use either or both.

## Creating an eval

Navigate to your agent's Instructions page and **Evals** tab and click **Add Eval**. The eval editor has three tabs: Test Prompt, Tools, and Grader.

<Frame>
  <img alt="Evals tab showing list of configured evaluations" />
</Frame>

### Test prompt tab

Write the prompt your agent will receive when the eval runs. This simulates a user message or scenario.

<Frame>
  <img alt="Edit eval modal showing the Trigger tab" />
</Frame>

For a target screening agent called "Deal Scout", you might write:

* "Evaluate this company: Acme Industrial, \$45M revenue, manufacturing sector, Germany"
* "Research Nordic Components AB and assess fit against our investment criteria"
* "Is TechStartup Inc (\$2M revenue, SaaS, San Francisco) worth pursuing?"

You can also simulate scenarios by including context directly: "Based on this company profile, evaluate fit against our criteria: Acme Corp, \$30M revenue, logistics sector, Netherlands..."

Under **Execution Settings** (inside this tab), you can also choose:

* **Execution mode**
  * **Chat** (default): Runs like a normal chat message
  * **Trigger**: Runs through the trigger pipeline used for autonomous events (useful for testing trigger-based behavior)
* **Clone the agent for this eval**
  * **On** (default): Runs on a temporary clone to isolate state changes
  * **Off**: Runs against the real agent state
* **Include the agent's system prompt**
  * **On** (default): The model receives the agent's system prompt and identity, just like a real chat or trigger
  * **Off**: Sends only the test prompt with no agent identity or instructions. Useful for benchmarking raw model capabilities side-by-side, independent of your agent's instructions. Tools are controlled separately on the **Tools** tab — exclude them there if you want a fully tool-free run.

### Tools tab (optional)

Here you can exclude or fake specific tools during the eval.

<Frame>
  <img alt="Edit eval modal showing the Tools tab with included, excluded, and faked options" />
</Frame>

| Mode         | Description                                                  |
| ------------ | ------------------------------------------------------------ |
| **Included** | Tool runs normally (default)                                 |
| **Excluded** | Tool is unavailable to the agent during this eval            |
| **Faked**    | Tool returns a canned response instead of actually executing |

**When to fake tools:**

* **Prevent side effects** — Fake `send_email` so the agent thinks it sent an email, but nothing actually goes out
* **Control inputs for deterministic testing** — Fake `perplexity_research` to return specific company data, so you know exactly what answer to expect
* **Speed up tests** — Fake slow external API calls

**When to exclude tools:**

* Test how the agent handles missing capabilities
* Ensure the agent doesn't use certain tools for specific scenarios

<Note>
  Faked tools return a default message: "This tool was faked for testing. Assume it succeeded and continue normally." You can customize this per-tool.
</Note>

### Grader tab

This is where you define how the eval result is determined.

<Frame>
  <img alt="Edit eval modal showing the Grader tab with grading options and validation rules" />
</Frame>

**Grader prompt** (optional) — If provided, an LLM examines the transcript and uses your prompt to determine the result.

For a screening agent:

* "Pass if the agent correctly identified this target doesn't meet our minimum revenue criteria (\$20M) and rejected it."
* "Rate the research completeness from 1-10. Award 2 points each for: revenue verified, ownership structure found, industry position assessed, geographic presence confirmed, M\&A history checked."

**Grading type:**

| Type              | When to use                                          |
| ----------------- | ---------------------------------------------------- |
| **Pass/Fail**     | Clear-cut criteria: "Did it reject non-fits?"        |
| **Rating (1-10)** | Quality assessment: "How thorough was the research?" |

**Grader scope** — What should the grader see?

| Scope                   | When to use                                                                  |
| ----------------------- | ---------------------------------------------------------------------------- |
| **Full transcript**     | You care about the process—what tools were called, what steps the agent took |
| **Final response only** | You only care about the end result, not how the agent got there              |

**Validation rules** (optional) — Code-based assertions about tool usage. These run before the grader and fail fast if not met.

| Rule                | Example use case                                                             |
| ------------------- | ---------------------------------------------------------------------------- |
| **Tool called**     | When doing research, it should use the `perplexity_search` tool              |
| **Tool not called** | When asked to *draft* an email, it should NOT use `send_email`—just draft it |

To configure validation rules in today’s UI:

1. Click **Edit Rules** in the Grader tab.
2. For each tool, click **Must be called** or **Must NOT be called**.
3. Use **Filter tools** to quickly find a tool in larger tool lists.
4. Click **Done** to return to the summary list, where each rule shows a checkmark/X icon and can be removed individually.

If a validation rule fails, the eval fails immediately and the grader is skipped (saving credits).

**What if you have no grader prompt and no validation rules?** Then it's a smoke test—the eval passes if the agent responds without errors. This is more of a technical platform check, not typical for most use cases.

### Managing eval definitions

From the eval list, you can quickly:

* **Run** an eval
* **Edit** an eval
* **Duplicate** an eval (useful when creating variants)
* **Export** an eval as JSON
* **Delete** an eval (including all results linked to it)

## Running evals

Once your eval is defined, click the **play button** on that row to run it, or click **Run All** to run all evals at once.

A run summary modal opens immediately and shows live progress while results arrive. You can also abort an in-progress run from this modal.

When the run completes, you can inspect results there or later in the **Eval Results** tab.

<Note>
  Evals consume credits for both execution and grading. Running against multiple models multiplies usage proportionally.
</Note>

### Model settings

Expand **Model Settings** in the Evals tab to configure:

<Frame>
  <img alt="Model Settings showing eval models and grader model configuration" />
</Frame>

| Setting          | Description                                                                                                                                                                       |
| ---------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Eval models**  | Which model configurations to test with. You can add multiple entries, including the same model twice with different thinking settings, to compare quality and cost side-by-side. |
| **Grader model** | Which model evaluates the results.                                                                                                                                                |
| **Thinking**     | Configure thinking separately for each eval model entry, and separately for the grader model.                                                                                     |

<Note>
  If your workspace admin disables a model, it is excluded from eval runs and removed from the model picker. Previously selected disabled models remain in your config but are shown as orange warning tags — re-select a different model to keep that comparison active.
</Note>

<Tip>
  **Which model for grading?** Consider how hard the grading task is. Look at your grader prompts—if they require nuanced judgment, use a more capable model. For simple checks ("does the response contain X?"), a faster model works fine.
</Tip>

## Viewing results

The **Eval Results** tab shows your eval history organized by run.

<Frame>
  <img alt="Eval Results tab showing history of eval runs" />
</Frame>

Each run displays pass/fail counts (or average rating), duration, and credits consumed. Results that couldn't be graded because the agent run itself errored (for example, an invalid model or an execution failure) are shown as a separate orange **Error** count so they don't get confused with genuine fails.

Credits are shown with test and grader usage separated:

* The main value shows test credits (agent execution).
* If grader costs are present, they appear as a side note (for example, `7.98 credits (+8.66 grader)`).

**Switch between views:**

* **List view** — Results listed by eval and model
* **Grid view** — A matrix of eval × model, where each cell shows pass/fail/rating. Useful for comparing how different models perform on the same tests.

<Frame>
  <img alt="Eval run summary in grid view comparing results across models" />
</Frame>

### Result details

Click on the Pass/Fail/Rating badge to see more details about the eval run.

* The agent's complete response
* Grader reasoning (why it passed or failed)
* Validation rule results
* Test credits and grader credits shown as separate values
* Full conversation transcript including tool calls

Use the **Rerun** button to re-run that specific eval and model combination directly from the result view, or use **Rerun** from the run summary modal to re-run the full eval lineup.

In the transcript view, you can switch between:

* **Friendly** format for quick human review
* **Raw JSON** for full structured data

Example of a successful eval run:

<Frame>
  <img alt="Eval result showing grader reasoning and transcript" />
</Frame>

Example of a failed eval run:

<Frame>
  <img alt="Eval result showing validation rule failure" />
</Frame>

You can examine the full conversation transcript to see the tool calls and the agent's response.

### Sub-agent calls in transcripts

If your agent delegates or calls another agent during an eval, the sub-agent runs with its own tools and configured model. Its full conversation is captured and shown as a nested, collapsible transcript inside the parent's tool call, labelled with the sub-agent's name and model.

Tool actions performed by a sub-agent are real (e.g. document edits) unless you fake or exclude those tools in the eval configuration. Agent-to-agent calls can recurse up to 3 levels deep within an eval.

### Managing run history

In **Eval Results**, you can manage runs as units:

* Give a run a custom name for easier tracking
* Delete an entire run (all results in that run)
* Open a run summary grid to compare evals across models

## Writing effective evals

Now that you understand the mechanics, here's how to write evals that actually catch problems.

### Write deterministic grader prompts

Include the expected answer so the grader can simply check—don't make it figure out what "correct" means.

Think of it like a teacher intern grading a test. The intern shouldn't solve the problems themselves—they should have an answer key. You're testing the agent, not the grader.

**Good (deterministic):**

* "The target should be rejected. Revenue 2 million USD is below our 20 million USD minimum threshold."
* "Research completeness should include: revenue source identified, ownership structure mapped, industry position assessed."

**Bad (requires grader to evaluate):**

* "Check if the investment analysis is accurate."
* "Verify the research is thorough."

### Control inputs for deterministic testing

To write deterministic graders, you need to control what the agent sees. Two approaches:

**1. Fake tools with predetermined responses:**

For a screening agent, fake `perplexity_research` to return specific company data:

* **Tool override:** `perplexity_research` → Faked with: "Nordic Components AB: €52M revenue, family-owned since 1985, automotive parts manufacturing, headquarters in Sweden with plants in Germany and Poland"
* **Grader:** "Check that the agent identified: revenue €52M (above threshold ✓), family-owned (succession opportunity ✓), manufacturing sector (matches criteria ✓), European presence (✓). Should recommend adding to shortlist."

**2. Simulate scenarios in the trigger:**

* **Trigger:** "Evaluate this target based on the following data—do not perform additional research: Company: Acme Industrial, Revenue: \$45M, Sector: Manufacturing, Ownership: Family (founder retiring), Location: Munich, Germany"

### Example: Target Screening Agent eval suite

Here's how a complete eval suite might look for an investment screening agent:

#### "Reject obvious non-fits" (Pass/Fail + Validation)

**Scenario:** Agent should reject targets that don't meet basic criteria.

* **Trigger:** "Research and evaluate this target: TechStartup Inc, \$2M revenue, SaaS, VC-backed, San Francisco"
* **Validation:** `perplexity_research` must be called
* **Grader:** "Pass if the agent correctly identified this doesn't meet criteria (revenue below \$20M minimum, wrong sector, wrong geography) and rejected it without extensive research."

**What it catches:** Agent being too optimistic, flagging everything "for review." Agent not doing actual research.

***

#### "Score against investment criteria" (Rating 1-10)

**Scenario:** Assess the quality of the agent's scoring logic.

* **Trigger:** "Evaluate: Acme Industrial, \$45M revenue, manufacturing sector, family-owned (founder age 68), based in Germany"
* **Tool overrides:** `perplexity_research` → Faked with predetermined company profile
* **Grader:** "Rate scoring accuracy 1-10. A qualified target should score 7-8/10 based on: revenue above threshold (✓), manufacturing sector (✓), European location (✓), succession situation (✓). Deduct points if scoring seems arbitrary or doesn't reference specific criteria."

**What it catches:** Inconsistent or inflated scoring. One team found their agent was giving everything 8+/10—this eval revealed the problem.

***

#### "Research completeness" (Rating 1-10)

**Scenario:** Ensure the agent gathers sufficient information before scoring.

* **Trigger:** "Research and evaluate: Nordic Components AB"
* **Tool overrides:** `perplexity_research` → Faked with partial company data
* **Grader:** "Rate research completeness 1-10. Award 2 points each for: revenue verified, ownership structure identified, industry position assessed, geographic footprint mapped, M\&A history checked."

**What it catches:** Agent rushing to conclusions. Also useful for model comparison—does a cheaper model still do thorough research?

***

#### "Handle missing data gracefully" (Pass/Fail)

**Scenario:** When information is unavailable, the agent should acknowledge uncertainty.

* **Trigger:** "Evaluate: Private Holdings GmbH"
* **Tool overrides:** `perplexity_research` → Faked with very limited results
* **Grader:** "Pass if the agent explicitly notes which data points couldn't be verified and either: (a) recommends direct outreach before scoring, or (b) provides a tentative score with clear caveats. Fail if it presents guesses as facts or gives a confident score despite missing information."

**What it catches:** Agent hallucinating data or overconfident with incomplete information.

### Using evals to improve your agent

**Scenario: Agent was too optimistic**

The "Reject obvious non-fits" eval kept failing—the agent was adding everything to the shortlist "for further review."

**Fix:** Added instruction: "If a target clearly fails to meet minimum criteria (revenue below \$20M, outside target sectors, wrong geography), reject immediately. Don't shortlist for 'potential' or 'further review.'"

**Result:** Eval now passes. Analysts spend less time reviewing obvious non-fits.

***

**Scenario: Research quality varied by model**

Ran "Research completeness" eval across three models:

* Claude Sonnet: 8/10
* Claude Haiku: 5/10
* Gemini Flash: 6/10

**Decision:** Kept Sonnet for screening. The cost savings from cheaper models weren't worth incomplete research—analysts had to redo the work manually.

***

**Scenario: Scoring was inconsistent**

The "Score against criteria" eval showed wild variance—same target got 6/10 one run, 9/10 the next.

**Fix:** Added explicit scoring rubric to instructions: "Score 2 points for each criterion met: revenue above \$20M, target sector, European presence, succession situation, strategic fit. Maximum 10 points."

**Result:** Scores now consistent within ±1 point across runs.

## Import and export

Share evals between agents using JSON export/import.

**To export:**

* Click the **Export** icon on an eval row to export that single eval.
* Open the top-right actions menu and click **Export All** to export every eval.

Exports contain eval definitions (without internal IDs), so they can be safely imported into other agents.

**To import:** Click Import and paste JSON or upload a file. By default, evals with duplicate names are skipped. You can optionally overwrite existing evals by name.

When you overwrite by name, Abundly updates the existing eval in place, so historical results remain linked to that eval.

<Tip>
  Tool overrides reference tools by name, so the target agent needs the same tools for those settings to apply.
</Tip>

## Letting the agent manage its own evals

Toggle the **Evals capability** on to let your agent manage evals end to end.

With this capability enabled, your agent can:

* Create, replace, remove, and run evals
* Review eval run history and open full result details
* View and update eval model settings (including grader model and thinking settings)

This is useful for bootstrapping your eval suite. Ask the agent to create evals covering edge cases you might have missed:

<Frame>
  <img alt="Agent creating new evals to cover edge cases" />
</Frame>

Then ask it to run them across different models to compare performance:

<Frame>
  <img alt="Agent running evals and reporting results across models" />
</Frame>

The agent can identify which models struggle with specific tasks—in this case, Haiku failed the comparison task (didn't look up both documents), while Gemini Flash failed the borderline case (rejected instead of shortlisting with caveats).

<Tip>
  **Eval-driven development:** Define the evals first, then ask the agent to iterate on its own instructions until the evals pass. It's like test-driven development for AI agents.
</Tip>

## Eval design tips

* **Start simple, add complexity.** Begin with basic evals to verify functionality. Add grading criteria once you understand what "good" looks like.
* **Control inputs for deterministic grading.** Fake tools or simulate scenarios so you know exactly what answer to expect.
* **Use validation rules for tool behavior.** They're faster and cheaper than graders for checking which tools were called.
* **Fake tools with side effects.** Always fake or exclude tools that send emails, make API calls, or modify external systems.
* **Run evals after changes.** Change instructions → run evals. Add capability → run evals. Update model → run evals.
* **Use ratings for quality tracking.** Pass/fail tells you if something works. Ratings (1-10) tell you if it's getting better.

## FAQ

<AccordionGroup>
  <Accordion title="Do I need a grader prompt?">
    No. Without a grader prompt, the eval operates in "smoke test" mode (passes if no errors) or "validation-only" mode (passes if validation rules pass). Add a grader when you need LLM-based evaluation.
  </Accordion>

  <Accordion title="Do eval runs have side effects?">
    By default, evals run on a temporary cloned agent, so state changes are isolated. If you disable cloning for an eval, then side effects are possible depending on enabled tools (for example, document edits).
  </Accordion>

  <Accordion title="Can I test that a tool was called with specific arguments?">
    Yes, if you mention that in the grader prompt. The normal tool validation rules only check whether a tool was called, not the arguments.
  </Accordion>

  <Accordion title="Can I schedule evals to run automatically?">
    Evals are normally triggered manually. But if you enable the "Evals capability" on your agent, you can ask it to run them on a schedule.
  </Accordion>

  <Accordion title="How many evals can I create?">
    No limit. Create as many test cases as needed to thoroughly validate your agent's behavior.
  </Accordion>

  <Accordion title="Should I use a more capable model for grading?">
    That depends on the complexity of the grading criteria. If the criteria are simple, you can use a cheaper model for grading. If the criteria are complex, you should use a more capable model for grading. Experiment!
  </Accordion>
</AccordionGroup>


# HTTP Requests: Connect to Any API
Source: https://docs.abundly.ai/features/http

Connect your agent to any online service with a REST API

The HTTP capability is your agent's "swiss army knife"—enabling it to talk to virtually any online service that has an API. If a service exists on the internet and exposes an API, your agent can interact with it.

<Frame>
  <img alt="Agent checking real-time weather in Stockholm using an external API" />
</Frame>

In this example, we simply asked the agent to "check the current weather in Stockholm using a suitable API." The agent found a free weather API, made the HTTP request, parsed the JSON response, and presented the results in a readable table—all without any configuration.

<Warning>
  This capability is hidden by default. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable it for your workspace. Due to its power and flexibility, we enable it on a case-by-case basis.
</Warning>

## How it works

Once enabled, you simply tell your agent which API you want to interact with—either in the chat or in your [instructions](/features/instructions). The agent figures out how to call the API and handles the request for you.

<Tip>
  Combine with the [Web Search](/features/search) capability for even more power. Your agent can search for APIs it doesn't already know about, read the documentation, and then use them.
</Tip>

## How the agent learns APIs

Think about what a human developer would need to work with an API—your agent needs the same things:

**Well-known APIs** — For popular services (weather APIs, social platforms, major SaaS tools), the agent often knows how to use them already from the LLM's training data. Things usually work right off the bat.

**Less familiar APIs** — For niche or custom APIs, you have several options:

* **Let it experiment** — Ask the agent to try the API and figure it out through trial and error
* **Point to documentation** — Give the agent a URL to the API reference, or upload a PDF with the docs
* **Use web search** — Enable [Web Search](/features/search) and ask the agent to research the API online
* **Describe it yourself** — If you know the API, just explain the endpoints and expected responses in the chat

The more context you provide, the faster the agent will succeed—especially for internal or proprietary APIs that don't exist in public documentation.

## Guardrails and approval

The HTTP capability card on the agent has two main guardrails:

* **Allow requests to any URL (without secrets)** — When on, the agent can hit any public URL. When off, requests are only permitted to endpoints defined by an enabled [API Capability](/integrations/overview#extending-integrations), so the agent is locked down to a specific set of APIs.
* **Require Approval** — Choose **No**, **Yes, for all requests**, or **Yes, except GET requests**. When approval is required, the agent must wait for a workspace member to approve each outgoing request before it executes.

<Frame>
  <img alt="HTTP requests settings showing Require Approval and allow unrestricted requests options" />
</Frame>

This is similar to how approval works for [email and SMS](/features/messaging).

## When to use HTTP requests

The HTTP capability is ideal when:

* **No existing capability** — The service you need isn't available as a built-in [capability](/features/capabilities) or [MCP server](/integrations/mcp)
* **Quick prototyping** — You want to test an integration before investing in a proper setup
* **Internal APIs** — You have proprietary services or internal databases with REST endpoints
* **Combining APIs** — Your workflow requires calling multiple APIs in sequence

<Note>
  If a dedicated capability or MCP server exists for the service you need, prefer those—they're more reliable and better tested. Use HTTP as a fallback or for experimentation.
</Note>

## Example use cases

**Real-time data lookups:**

* "Check the current Bitcoin price using a cryptocurrency API"
* "Get the latest exchange rate between USD and EUR"
* "Look up the shipping status for order #12345 via the carrier's API"

**Combining multiple APIs:**

* "Convert this address to GPS coordinates, then find the nearest coffee shops using a places API"
* "Fetch the weather forecast, then post a summary to Slack"
* "Look up the company info from this domain, then add it to my CRM"

**Prototyping integrations:**

* "Connect to our internal inventory API and check stock levels for product SKU-1234"
* "Query our company's time tracking API and summarize my hours this week"

## Real-world example: Office location analyzer

We built an HR agent that knows where all our employees live and helps analyze potential new office locations by calculating transit times for each person.

<Frame>
  <img alt="Interactive map showing team member locations and calculated transit times to potential office locations" />
</Frame>

This required chaining multiple APIs:

1. **Geocoding API** — Convert street addresses to GPS coordinates
2. **Transit time API** — Calculate commute times between coordinates
3. **Map visualization** — The agent built an interactive map to display the results

It took some iteration for the agent to discover which APIs to use and how to call them. But once it worked, we asked the agent to document the process—essentially "codifying" those API calls into its [instructions](/features/instructions) so future requests are fast and reliable.

## Handling credentials

For authenticated APIs, store credentials as **Secrets** in the workspace's **Shared Assets → Secrets** page, then define an **API Capability** (Shared Assets → API Capabilities) that describes the endpoints and how each secret is injected. The platform supports bearer tokens, API key headers, basic auth, query parameters, and custom headers — plus managed flows including OAuth 2.0 client credentials, GCP service accounts, and signed JWT bearer tokens (RS256 / ES256). The agent references the API by name; the actual secret value is injected server-side and never reaches the LLM.

Both Secrets and API Capabilities support fine-grained access control — scope them to specific teams or agents, and the platform will only let those agents use them.

<Warning>
  For quick experiments you can paste a credential into chat or instructions, but those values are **not encrypted** and are visible to anyone with access to the agent. Use Secrets for anything sensitive.
</Warning>

See [Credentials](/security/credentials) for more on how secrets are stored and protected.

## Available methods

The agent can use any standard HTTP method:

| Method                 | Use case                    |
| ---------------------- | --------------------------- |
| **GET**                | Retrieve data from APIs     |
| **POST**               | Create records, submit data |
| **PUT**                | Replace existing records    |
| **PATCH**              | Partial updates             |
| **DELETE**             | Remove records              |
| **HEAD** / **OPTIONS** | Check API capabilities      |

## Tips for reliability

**Ask the agent to remember what worked.** Once your agent successfully calls an API, ask it to update its instructions with the details—endpoint URLs, required headers, expected responses. This "codifies" the API and makes future calls instant.

**Start simple.** Test with basic GET requests before building complex multi-step workflows.

**Provide documentation.** If you have API docs, share them with the agent. The more context it has, the faster it can figure out the correct calls.

## FAQ

<AccordionGroup>
  <Accordion title="Why is this capability hidden by default?">
    The HTTP capability is extremely powerful—it can interact with any API on the internet. We enable it on a case-by-case basis to ensure organizations understand the security implications and have appropriate guardrails in place.
  </Accordion>

  <Accordion title="Can the agent access any URL?">
    By default, yes — the agent can make requests to any publicly accessible URL. If you want to restrict it to a known set of endpoints, turn off **Allow requests to any URL** on the capability card and define the allowed endpoints as an [API Capability](/integrations/overview#extending-integrations). Approval can still be required on top of either setup.
  </Accordion>

  <Accordion title="What happens if the response is very large?">
    Large responses won't overwhelm the agent's context. When a response is too big, the platform steps in and guides the agent on how to handle it—for example, previewing a portion to understand the shape of the data, or processing the full response with [code execution](/features/scripts) so only the useful parts end up in the conversation. If the agent needs to work through a big payload in one pass, it can [delegate](/features/multi-agent) the task to a sub-agent, which has more room to work with.
  </Accordion>

  <Accordion title="Why did my agent try multiple APIs before finding one that works?">
    When you ask for something general like "check the weather using an API," the agent may try several APIs before finding one that works without authentication or meets your needs. Once it finds a working solution, ask it to document the approach in its instructions.
  </Accordion>

  <Accordion title="Can I restrict which APIs the agent can call?">
    Yes. Define an [API Capability](/integrations/overview#extending-integrations) for each endpoint the agent should be able to call, then turn off **Allow requests to any URL** on the HTTP capability card. The agent will only be able to call those defined endpoints.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Web Search" icon="magnifying-glass" href="/features/search">
    Let your agent find APIs and documentation online
  </Card>

  <Card title="MCP Servers" icon="plug" href="/integrations/mcp">
    Connect to external tools via Model Context Protocol
  </Card>

  <Card title="Instructions" icon="scroll" href="/features/instructions">
    Codify API knowledge in your agent's instructions
  </Card>
</CardGroup>


# Agent Instructions
Source: https://docs.abundly.ai/features/instructions

Configure the behavior of your agent

Instructions are the foundation of your agent's behavior. This is where you describe what the agent should do, how it should behave, and provide the context needed to do its job effectively. Think of it like a working agreement between the agent and its owner.

<Frame>
  <img alt="Overview of agent instructions" />
</Frame>

A new agent starts with empty instructions, making it a general-purpose chatbot. Adding instructions gives your agent a purpose—similar to writing a job description for a new colleague.

The instructions are always "top of mind" for the agent. They are included in the agent's context in every situation, whether it is responding to a chat message, an external trigger (such as an email), or waking up to a scheduled event.

<Frame>
  <img alt="Agent instructions editor showing an invoice processing agent's role and workflow" />
</Frame>

## Editing instructions

You have two ways to create and update agent instructions:

* **Direct editing** — Click the **Instructions** link in the agent's navbar to view and edit instructions directly. This gives you full control over the exact wording and structure.
* **Conversational configuration** — Chat with your agent and then ask it to update its own instructions based on your conversation. The agent will draft instructions and ask for your approval before saving.

<Tip>
  Getting an agent from "This works pretty OK" to "Wow this is awesome!" usually takes a few rounds of refinement. You can tweak the instructions directly, or give the agent feedback and ask it to analyze its own instructions and suggest improvements.
</Tip>

## Typical workflow

Here is a typical workflow for instructing a new agent:

1. **Start a conversation:** Describe what you want the agent to do in plain language, as if you're onboarding a new colleague.
2. **Agent asks for what it needs:** The agent requests capabilities, documents, or clarification based on your requirements.
3. **Agent drafts instructions:** The agent writes its own instructions using your conversation and presents them for your review.
4. **Review and tweak:** Read through the draft, make any tweaks, and approve. The agent begins working.
5. **Iterate over time:** Return anytime to improve the agent—either by editing instructions directly or through further conversation.

This approach works because:

* Writing concise, clear instructions is hard. The agent often writes better instructions than humans would, and when setting up a new agent, this is often easier than trying to write perfect instructions from the start.
* You don't need to be a technical expert, you focus on what problem the agent should solve, and the agent helps you figure out how to solve it.
* The agent tells you what information and which capabilities it needs, so you don't need to guess.
* Instructions are written in clear language you can review and tweak
* You can iterate and improve the agent's behavior through conversation, so you can start simple and improve the agent as you go.

<Note>
  The agent will only update its instructions if you ask it to, or if it recognizes that an update would help it serve you better and asks for permission. The **Update Instructions** capability is enabled by default, with guardrails you can configure.
</Note>

## Update instructions capability settings

The **Update Instructions** capability has settings that control when the agent is allowed to edit its own instructions:

* **Chat only** (enabled by default) — the agent can only update instructions from chat conversations in the Abundly web portal. Slack chats, Teams chats, and autonomous triggers like email, webhook, or scheduled runs are all blocked.
* **Edit access required** (enabled by default) — when chat only is enabled, only users with **Edit** or **Admin** access to the agent can use instruction updates.

You can change these settings from the **Capabilities** page by opening the **Update Instructions** capability settings.

## Instruction versioning

Agent instructions are versioned automatically, so you don't need to worry about losing information when you or the agent make changes.

With versioning you can:

* Track changes over time
* Revert to previous versions if needed
* See who made changes (you or the agent)
* Compare versions side-by-side to see exactly what changed

<Frame>
  <img alt="Version timeline showing instruction changes between versions with additions and changes highlighted" />
</Frame>

## FAQ

<AccordionGroup>
  <Accordion title="Do updated instructions apply to existing chats?">
    Yes. The agent will start using the new instructions immediately.
  </Accordion>

  <Accordion title="What happens if I approve a change?">
    Nothing really—it just marks the change as approved and removes it from the diff view.
  </Accordion>

  <Accordion title="What happens if I undo a change?">
    It edits the instruction, reversing that change. Just like if you did it yourself, manually. When you click "Save changes" it will be saved as a new version.
  </Accordion>

  <Accordion title="How do I revert to an earlier version?">
    Click on that version in the timeline, then select "Restore" in the hamburger menu. It will create a new version that matches that older version. Click "Save changes" to confirm.
  </Accordion>
</AccordionGroup>


# Context and Memory
Source: https://docs.abundly.ai/features/memory

How agents know what they know, and remember across conversations

Context management drives your agent's behavior. What the agent knows at any moment determines how it responds—so understanding how context works helps you design more effective agents.

<Frame>
  <img alt="Diagram showing global context (instructions and documents) always available to the agent, with local context from individual chats and triggers" />
</Frame>

## Global context

Some information is always available to your agent, no matter what it's doing. Whether responding to a chat message, processing an incoming email, or waking up for a scheduled task, the agent has access to:

* **[Instructions](/features/instructions)** — The agent's mission, guidelines, and core knowledge
* **[Agent documents](/features/documents)** — Files, databases, and reference materials (depending on visibility settings—by default the agent knows which documents it has and can read them when needed)

These are the foundation for everything the agent knows across interactions.

## Local context

Each interaction has its own isolated context that the agent sees alongside the global context:

* **Chat conversations** — When you're in a chat, the agent knows all previous messages in that conversation. But it doesn't know about other chats—each conversation is independent.
* **Triggers** — When responding to an email, scheduled task, or webhook, the agent sees the trigger details (email content, task instructions, etc.) but not any chat conversations.

The platform optimizes what context to include for each trigger type. For example, when triggered by a Slack message, the agent receives the 5 previous messages in that channel for context. If needed, the agent can actively search for more information.

## What is memory?

Memory is the ability to remember information across different contexts. For example, the agent might remember which company you work for, your communication preferences, or decisions made in previous conversations.

### Instructions and documents work as memory

Since agents can read and write their own [instructions](/features/instructions) and [documents](/features/documents), these already serve as memory. The agent can store facts, preferences, and accumulated knowledge—and access it in any future context.

You can guide this behavior in your agent's instructions:

* "Create a database to store personal preferences for everyone in my workspace. Whenever someone expresses a preference, store it. When communicating with someone, check their preferences first."
* "Keep a running log of all decisions we make in our conversations"
* "When you learn something important about a client, add it to the client notes document"

<Info>
  Your agent can also read its own [diary](/features/monitoring#agent-diary)—a log of its activities across all contexts. The agent won't read its diary automatically, but will access it when asked or when it determines that reviewing past activities would help with the current task.
</Info>

## When to use what

| Approach                                         | Best for                                                           |
| ------------------------------------------------ | ------------------------------------------------------------------ |
| **Instructions**                                 | Core knowledge the agent needs in every interaction                |
| **Documents**                                    | Reference materials, templates, detailed procedures                |
| **[Agent databases](/features/agent-databases)** | Structured records (invoices, contacts, logs) with semantic search |

For use cases like accumulating facts over time or deduplicating recurring information (e.g. "remember news items you've already sent so similar stories aren't repeated"), use an agent database with semantic search. Databases find entries by meaning, not exact text, so the agent can recognize when two items are essentially the same even if worded differently.

## The Memory capability

<Note>
  The Memory capability is retired for new agents. It remains fully functional for agents that already have it enabled, but it can no longer be added to new agents. Use [agent databases](/features/agent-databases) instead—they offer the same semantic search with more flexibility.
</Note>

If your agent has the Memory capability enabled, it has a dedicated memory store with semantic search. Saved memories are converted to vector embeddings so they can be retrieved by meaning rather than exact text match.

To view and manage memories, open the Memory capability card and click **View memories**. You can search semantically, add new memories, or remove existing ones.

<Frame>
  <img alt="Agent Memories dialog showing semantic search results with match percentages" />
</Frame>

## Learn more

<CardGroup>
  <Card title="Instructions" icon="scroll" href="/features/instructions">
    The foundation of your agent's behavior
  </Card>

  <Card title="Agent Documents" icon="folder-open" href="/features/documents">
    Files and reference materials for your agent
  </Card>

  <Card title="Agent Databases" icon="database" href="/features/agent-databases">
    Structured data stores with semantic search
  </Card>
</CardGroup>


# Email & SMS
Source: https://docs.abundly.ai/features/messaging

Send and receive emails and text messages

Agents can communicate through email and SMS. Email works out of the box—just enable the capabilities and go. For SMS, you need to contact [support@abundly.ai](mailto:support@abundly.ai) (for now).

## How it works

Every agent automatically gets its own email address (e.g., `myagent.myworkspace@agent.abundly.ai`). The agent uses this as its own identity—it doesn't access your email account or send emails on your behalf (unless you enable a capability to access your email account).

<Frame>
  <img alt="Email capabilities showing Send Email and Receive Email toggles" />
</Frame>

Email capabilities:

* **Send Email** — Allows the agent to send emails from its address
* **Receive Email** — Triggers the agent when emails arrive at its address

No configuration required. Just toggle the capabilities on in **Settings → Capabilities**.

<Tip>
  You can customize the first part of the agent's email address in the agent settings, under Main Settings -> Communication.
</Tip>

## Handling incoming messages

When the agent receives an email or SMS, it needs instructions on what to do—otherwise it will likely ignore it. Add guidance in your agent's [instructions](/features/instructions).

By default there is no mailbox layer—each incoming message triggers the agent, which processes it according to your instructions. The platform also keeps a copy of every sent and received email and SMS so you can review them in the UI (see [Activity monitoring](/features/monitoring#saved-messages-email-and-sms)); message logging can be paused per agent from that tab. You can also ask the agent to keep its own record (for example in a [database](/features/agent-databases))

<Tip>
  If you have an email-based workflow, it's often faster to test it in the chat first. Ask the agent to simulate
  receiving an email: "Process this as if you received it by email, but don't actually send anything—just tell me what
  you would do."
</Tip>

## Example use cases

**Sending emails:**

* "Every Friday, email me a summary of this week's completed tasks"
* "Email the sales team whenever a new lead comes in, and attach a PDF summary of the company research"

**Processing incoming emails:**

* "When you receive an email with an invoice, extract the details, validate against our vendor list, and add to the invoice database"
* "When you receive a support email, check the knowledgebase and reply if the answer is clear. Otherwise, forward to [support@company.com](mailto:support@company.com)"

**Logging messages:**

* "Store all incoming emails in a database. Create the database if it doesn't exist."
* "Log all outgoing emails with recipient, subject, and timestamp"

**SMS:**

* "Send a text reminder to all attendees 24 hours before their appointment"
* "Text me if a critical alert comes in after hours"

## What agents can do with email

Agents can handle email much like a human would:

* **Attachments** — Read incoming attachments (PDFs, images, documents) and add attachments to outgoing emails
* **CC and BCC** — Include additional recipients
* **Reply and forward** — Respond to threads or forward to others
* **Rich formatting** — Send nice-looking HTML emails with embedded images and formatting

## Guardrails and approval

Giving an agent the ability to email externally is powerful—but it also requires care. You can configure guardrails to limit what the agent can do with email:

<Frame>
  <img alt="Send Email capability settings showing allowed recipients and approval options" />
</Frame>

Send Email settings live inline in the capability drawer and include a plain-English summary of how the agent will behave with the current configuration.

### Allowed recipients

Turn on **Restrict who the agent can email** to limit who the agent can contact. The allowed list mixes two kinds of entries:

* **Specific addresses** — individual inboxes, e.g. `name@example.com`
* **Whole domains** — written as `@example.com`, matches anyone at that domain

The first time Send Email is enabled, the agent's owner is pre-seeded as the only entry on the allowed list, so the agent is locked down to that one address until you broaden it.

### Approval

When the agent is restricted, you choose what happens for recipients inside and outside the allowed list independently:

| Recipient            | Options                                   |
| -------------------- | ----------------------------------------- |
| **Inside the list**  | Send the email, or ask for your approval  |
| **Outside the list** | Block the email, or ask for your approval |

By default the agent sends to allowed recipients without asking and blocks everything else. Switch the outside-list rule to "Ask for my approval" if you'd rather review unknown recipients than block them outright. Switch the inside-list rule to "Ask for my approval" if you want a human in the loop even for known contacts.

When restriction is off, you instead get a single "Ask before every email" toggle — useful when you want broad reach but a human in the loop on every send.

* In a chat context, approval requests show up as a widget in the chat.
* When responding to triggers, approval requests show up in the Activity tab.

Each approval card has a **Notify agent** checkbox next to the action buttons. Tick it before approving or rejecting to tell the agent about your decision so it can take follow-up steps; leave it unchecked and the action just executes (or doesn't) without the agent learning the outcome. Inside a chat the checkbox is hidden — the agent is always notified there so the conversation can continue.

<Frame>
  <img alt="Email approval request widget in the chat" />
</Frame>

See [Guardrails](/security/guardrails) for more on configuring these controls.

## SMS

Agents can send and receive text messages for time-sensitive communication.

<Info>SMS capabilities are currently hidden. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable them for your account.</Info>

SMS works similarly to email:

* **Send SMS** — Agent can send text messages
* **Receive SMS** — Agent is triggered by incoming texts (requires a dedicated phone number)

The same guardrails apply—you can restrict the agent to an allow-list of phone numbers and configure inside-list and outside-list approval rules.

## Security considerations

Email and SMS give your agent the ability to contact people outside the platform. Before enabling these capabilities, consider:

* **What data can the agent access?** If it has access to sensitive information, be careful about who it can contact.
* **What are the consequences of a mistake?** An agent emailing the wrong person or sending incorrect information could cause problems.
* **What guardrails are appropriate?** Start restrictive (allowed list + approval for anyone not on it) and loosen as you gain confidence.

The platform includes multiple layers of protection—see [Security overview](/security/overview) for details on how agents are monitored and constrained. But ultimately, you decide which capabilities, instructions, and guardrails to configure.


# Model Selection
Source: https://docs.abundly.ai/features/model-selection

Choose the right LLM for your agent's needs—or let the platform pick for you

Every agent needs a brain. Abundly supports models from all the major AI providers—Anthropic, OpenAI, Google, xAI, and more—and you choose which one powers your agent. Or don't choose at all, and let the platform pick a recommended default for you.

<Frame>
  <img alt="Agent settings showing model selection with a Claude Sonnet alias selected" />
</Frame>

## Available models

Your agents get access to a broad range of models through one consistent interface. Within each family, models are listed from most capable (and most expensive) to cheapest:

| Provider             | Models, most capable first | Details                                                                                        |
| -------------------- | -------------------------- | ---------------------------------------------------------------------------------------------- |
| **Anthropic Claude** | Fable, Opus, Sonnet, Haiku | [Anthropic's model overview](https://platform.claude.com/docs/en/about-claude/models/overview) |
| **OpenAI GPT**       | Sol, Terra, Luna           | [OpenAI's model overview](https://platform.openai.com/docs/models)                             |
| **Google Gemini**    | Pro, Flash, Flash Lite     | [Google's Gemini model overview](https://ai.google.dev/gemini-api/docs/models)                 |
| **xAI Grok**         | Grok                       | [xAI's model overview](https://docs.x.ai/docs/models)                                          |

Some workspaces also include additional families such as **Qwen** and **Mistral** (EU-hosted via [Scaleway](https://www.scaleway.com/) serverless inference), **DeepSeek**, or self-hosted models.

All models work with the same agent capabilities and tools—you don't lose any platform features by switching models.

<Note>
  Grok models are available when your workspace has xAI configured. Workspace admins control which models are
  available—see [Workspace administration](#workspace-administration).
</Note>

## Default behavior

If you don't have a preference, leave the model set to **(no preference)**. This is the default, and it means Abundly will use whatever model we think works best for general agentic behavior. We continuously evaluate models and update this default as better options become available.

This is the right choice for most users—you get good performance without having to think about model selection at all.

## When to pick a different model

You only need to pick a model if you don't want the default. The common reasons are:

* **You want more capability.** The agent works on hard problems, and you're willing to spend more credits per request for a more powerful model.
* **You want lower cost.** The agent does simple or high-volume work that doesn't need a top-tier model, and a cheaper model does the job for a fraction of the credits.
* **You've found a better fit.** You've tested and determined that a specific model works better for your specific use case.

### How to compare models

Within a provider's model family, the ordering is reliable: bigger models are more capable and cost more, smaller models are cheaper and usually faster. For example, in Anthropic's family, Fable is more capable (and more expensive) than Opus, which is above Sonnet, which is above Haiku.

Comparisons **across** providers are far less reliable. Benchmarks and marketing claims rarely predict how a model performs on your specific task—in some use cases a small, cheap model from one provider outperforms a flagship model from another. Even speed can't be stated as a fixed fact, since it varies with how busy the provider's servers are.

The most reliable way to compare is to test with your own agent on real work. You can change the model at any time—even mid-conversation—and switch back if you're not happy. The model picker shows a cost indicator (one to six dollar signs) next to each model. More dollar signs means more expensive—and the scale isn't linear, so each step up can mean a several-fold price difference.

For a more systematic comparison, use [agent evals](/features/evals) to run the same test cases across different models—and find the best fit for your use case, or discover that a cheaper model does the job just as well.

<Tip>
  Not sure where to start? We find the Claude models work very well for agentic behavior—Sonnet for everyday work, Opus
  or Fable for the hardest problems, Haiku for speed and volume.
</Tip>

## Model aliases vs specific versions

When selecting a model, you can choose between:

* **Aliases** like "Claude Sonnet Best" — automatically points to the latest version that we recommend
* **Specific versions** like "Claude Sonnet 5" — locked to that exact version

**We recommend using aliases in most cases.** Models improve constantly, and when a new version of a model arrives, it usually makes sense to upgrade. By selecting an alias, you won't need to remember to do that—Abundly will point the alias to the newer version once we've verified it works well (typically after some internal testing).

When an alias is selected, hovering the model picker button in the chat's top bar shows a tooltip indicating which specific model version the alias currently points to—so you can always see what you are actually running.

<Note>
  When an alias is updated, your agent's behavior may change. Usually for the better—models tend to get smarter—but not
  always in the way you expect. For example, a newer model may follow instructions more literally, surfacing problems
  in your instructions that the previous model quietly ignored. If that happens, a small tweak to the instructions
  usually fixes it.
</Note>

If you have carefully tuned your agent and don't want its behavior to change, lock it to a specific model version instead. The tradeoff—much like turning off auto-update for software—is that the agent won't automatically benefit from model improvements, and you decide when to upgrade.

## Per-context model selection

Different contexts in your agent's life may benefit from different models. You can configure a model and thinking preference per context:

* **Chat** — Conversations you start with the agent
* **Scheduled tasks** — When the agent runs on a schedule
* **Email** — When the agent responds to incoming emails
* **Slack** — When the agent responds to Slack messages
* **Microsoft Teams** — When the agent responds to Teams messages
* **Agent messages** — When other agents send your agent a message

If a context has no override, it inherits the agent's default model.

Examples:

* Use a cheaper model for handling routine scheduled work
* Use a more capable model for handling incoming emails

### Per-task model overrides

Individual scheduled tasks can have their own model and thinking setting on top of the per-context default. Open a task's settings card to configure it. This is useful when one particular task needs heavier reasoning than your other scheduled tasks.

### Per-chat model switching

Each chat keeps its own model: new chats start with the agent's chat model, and you can switch mid-conversation using the model picker in the chat's top bar. Switching affects **only that chat**—the agent's settings and other chats are untouched.

<Note>
  Remote conversations (Slack, Teams, WhatsApp, and so on) always use the model configured for that channel and can't
  be switched per conversation.
</Note>

## Thinking mode

Many models support a **thinking** mode where the model spends more effort reasoning before it responds.

<Frame>
  <img alt="Thinking settings shown in model selection controls" />
</Frame>

Thinking makes your agent smarter at complex reasoning tasks, but it also increases processing time and credit usage.

You can configure thinking in the default model settings, per context, and on individual scheduled tasks. Note that some models are locked to a specific thinking mode and cannot be changed.

## Workspace administration

Beyond per-agent settings, workspace admins get central control over model usage across the entire workspace. These settings live under **Workspace management**.

### Enabling and disabling models

Admins can enable or disable specific catalog models for the entire workspace from **Workspace management → Model selection**—useful for controlling costs, meeting policy requirements, or standardizing on a smaller set of models. Models that Abundly has disabled platform-wide remain locked and cannot be re-enabled at the workspace level.

If an agent is configured with a model that gets disabled, it stops with a clear error rather than silently substituting a different model—admins stay in control of exactly which models agents run on.

### Workspace default model

On the same tab, admins can set a **workspace default model**. Agents that have no model preference of their own use this model instead of the Abundly default—a quick way to standardize your whole workspace on a particular model or alias.

### Policy for new models

When Abundly adds a new model to the catalog, your workspace can either make it available automatically or hold it back until an admin reviews it. Choose the policy under **Workspace management → Settings → LLM models**:

* **New models on by default** — Your workspace benefits from new models as soon as Abundly adds them. This is the default.
* **Require approval for new models** — New models stay off until an admin enables them, giving you a controlled, admin-curated set of models. Models that are already available stay available.

## FAQ

<AccordionGroup>
  <Accordion title="Why isn't a 'Best' alias always pointing to the very latest version?">
    Newer models are usually better, but not always—especially for agentic work. Since an alias update affects many
    agents at once, we do some internal testing before pointing an alias to a new version. That's why an alias can lag a
    little behind a model release. If you want the latest version right away, you can select it as a specific version in
    the model picker.
  </Accordion>

  <Accordion title="How does pricing work for different models?">
    Each model has different credit costs per token. Within a family, more capable models cost more per request—more
    dollar signs in the model picker means more expensive. Check your [usage dashboard](/features/usage) to monitor
    credit consumption.
  </Accordion>

  <Accordion title="What happens when I switch models in the middle of a conversation?">
    The new model takes over from your next message, with the full conversation history available to it. You can switch models freely within a conversation.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Agent Evals" icon="flask-vial" href="/features/evals">
    Compare how your agent performs on different models with automated test cases
  </Card>

  <Card title="Usage & Limits" icon="gauge-high" href="/features/usage">
    Track credit consumption and set daily limits for your agents
  </Card>

  <Card title="Pricing" icon="coins" href="/introduction/pricing">
    Understand how credits and model costs work
  </Card>
</CardGroup>


# Activity Monitoring
Source: https://docs.abundly.ai/features/monitoring

Real-time monitoring and audit trails for agent activities

Since agents can perform tasks autonomously, it's useful to be able to monitor what they're doing. Click "Activity" in the sidebar.

## Usage

The **Usage** tab on an agent's **Activity** page shows how that agent is being used — interaction volume, channels, trends over time, and which workspace members have interacted with it. Anyone who can use the agent can open this tab.

You can filter by interaction channel and choose a time range (day, week, month, or all time up to one year). This is separate from **Credit usage** on the workspace **Usage** page, which is about billing and is limited to workspace and team administrators.

## Agent diary

Each agent maintains a diary — a high-level record of activities and internal reasoning.

Diary writing is enabled by default for each agent. You can pause or resume it per agent using the **Write diary entries** toggle in **Activity**. When paused, no new diary entries are written, but existing entries remain visible.

<Frame>
  <img alt="Agent Diary" />
</Frame>

When diary writing is enabled, the agent writes a diary entry when:

* It has processed a trigger (email, scheduled task, webhook, etc).
* It has done something significant in the chat.

<Info>
  In the chat, the agent decides when to write a diary entry based on the conversation content. For triggers and events, it typically writes one entry per processed event when diary writing is enabled. Writing a diary entry after every chat message would be too verbose and not useful.
</Info>

The agent can also read its own diary entries. However, it normally won't do so unless directly asked, or if it decides that it needs something specific from the diary.

The diary is useful if you want to find out WHY the agent did something. For example, if the agent routed an invoice to the wrong department, the diary can help you understand why, so you can tweak the instructions.

## Activity log

The activity log is a live-updated technical log of all triggers that the agent has processed. Triggers could be for example an email, a scheduled task, or a webhook.

<Frame>
  <img alt="Activity Log" />
</Frame>

<Info>
  Chat messages are not included in the activity log, since the purpose of the activity log is to show what is going on outside of the chat, when the agent does things autonomously.
</Info>

## Saved messages (email and SMS)

When any of the [email or SMS](/features/messaging) capabilities are enabled, the platform stores every inbound and outbound message by default. In the agent **Activity** area, open the **Messages** tab to list and search them. This is separate from the activity log: the activity log is about each trigger run and tool usage, while the Messages tab is the store of the actual email and SMS content.

The Messages tab has a **Save messages** switch you can use to pause logging for an agent — useful for noisy or privacy-sensitive agents. Existing entries remain visible; only new traffic is skipped.

With email or SMS capabilities enabled, the agent can search the same message log from chat. New messages are only added while the switch is on, but the log can still list messages that were stored earlier.

When an agent processes a trigger, it goes through a sequence of steps, each shown in detail in the expanded activity log entry. Each log row shows the trigger timestamp, event type, progress status, and total credits used for that execution.

* **Script execution** (script-based triggers only) Run the configured script first. If the script escalates, the full agent pipeline continues. Script-based scheduled runs appear as dedicated entries in the activity log, separate from any escalated agent pipeline run.
* **Pre-filter** Decide if this trigger is relevant to the agent. This is an early-out mechanism to save time and credits.
* **Attack detection** For security-sensitive trigger content (especially incoming email and SMS), the platform checks for manipulation attempts such as prompt injection or jailbreak instructions.
* **Execution** The agent acts on the trigger event according to its instructions, and the activity log shows the specific tool calls and end result.

If you enable **Credit details** in the agent's **Usage & Limits** settings, each expanded entry also shows a credit breakdown per phase and per tool call. This helps you identify where credits are being spent.

<Info>In addition to attack detection, the platform also enforces [guardrails](/security/guardrails) such as email whitelists and tool call approval requirements.</Info>

### Chat about an entry

Each expanded log entry has a **Chat about this** button (visible when the [Activity Log capability](#let-the-agent-analyse-its-own-activity-log) is enabled). It opens a new chat pre-filled with a prompt to drill into that specific entry, so you can ask the agent to explain what happened, why credits were spent a certain way, or what went wrong.

## Owner notices

When an agent is running autonomously — on a schedule, in a trigger, in a remote chat — nobody is watching it. The `notify_owners` tool gives agents a direct line to the people responsible for them. No capability needs to be enabled; the tool is always available to every agent.

The agent calls it with a short `title`, a full `message` (up to 8 000 characters), and an optional `severity` of `info`, `warning`, or `error`. The platform stores the notice and delivers it through the notification system to the agent's owners. For agents without an owner, delivery escalates to team admins and then to workspace admins — the same waterfall used for [tool approval requests](/security/guardrails).

Open the **Owner notices** tab on the agent's **Activity** page to see every notice in full. Each row expands to the complete message rendered as markdown. A **Chat about this** button opens a conversation pre-seeded with the notice, so you can ask the agent to act on whatever you decide.

Notices can also come from a person. Workspace and team admins can message the owners of many agents at once using the **Send notice** bulk action on the [Agent Management](/features/agent-management#bulk-actions) page — useful for announcing a policy change or asking owners to review their agents. These land in the same tab, marked as coming from the admin rather than the agent, and there is no category toggle to switch them off. The chat button still opens a conversation with the agent, so you can act on what the admin asked for.

<Note>
  The agent is told how many owners were notified. If nobody was reachable (for example, all recipients have opted out of notifications), it is told to try another channel rather than assuming a human has seen it.
</Note>

## Let the agent analyse its own activity log

By default the activity log is read-only and only visible to you in the UI. If you enable the **Activity Log** capability on the agent, it gets tools to inspect its own trigger history and credit usage and answer questions like:

* "How many credits did you use yesterday, and on which triggers?"
* "Summarise the errors you hit this week."
* "Which tool calls are eating the most credits?"
* "Walk me through the last email trigger you processed."

This is the right way to get the agent to reason about its own activity — rather than, for example, asking another agent to do it on its behalf.

If the agent is also connected to other agents via [agent-to-agent communication](/features/multi-agent), you can optionally turn on **Connected agent logs** in the Activity Log capability settings to let it inspect the activity logs of those connected agents too.

## Example

Freddy the invoice router received an email asking about invoice statistics. Let's walk through what the activity log shows for this trigger.

<Info>To see Freddy live in action, check out the [demo video](https://www.youtube.com/watch?v=qc2jrXj6qyU).</Info>

### The trigger

<Frame>
  <img alt="Email trigger showing Henrik asking Freddy how many invoices have been processed and what the total amount is" />
</Frame>

The trigger shows the incoming email that woke the agent. Henrik is asking for a quick summary of invoice processing statistics.

### Pre-filter

<Frame>
  <img alt="Pre-filter step showing Proceed: Yes, with reason that pre-filter is not applicable for this event type" />
</Frame>

The pre-filter step decides whether this trigger is relevant. For incoming email and SMS, this can include whitelist checks. For other trigger types, it may use a lightweight relevance check or be marked as not applicable.

### Attack detection

For untrusted incoming trigger content (especially email and SMS), the platform runs attack detection before execution. The log shows whether an attack was detected, the reason, and the attack type if blocked.

If an attack is detected, the trigger is blocked immediately and no execution takes place.

Workspace admins can configure who receives blocked-attack alert emails in **Workspace → Settings**. You can send alerts to all admins, admins plus additional addresses, or only a custom recipient list. You can also choose whether alert emails include the blocked trigger content for debugging.

### Execution

<Frame>
  <img alt="Execution step showing successful completion with 4 invoices found: 3 legitimate totaling €33,000 and 1 rejected fraud attempt for €12,000" />
</Frame>

The execution step shows what actually happened. The agent queried the Invoice Database, calculated the statistics (distinguishing between legitimate invoices and rejected fraud attempts), and sent the reply. The **Tools** section shows exactly which tool calls were made and when, interleaved with the agent's reasoning and intermediate assistant messages so you can follow its train of thought between tool calls. When a tool triggers another tool (for example, a tool called from within code execution), the log shows them as **nested entries** with a parent-tool link, making it easy to follow the full call chain.

### Diary entry

<Frame>
  <img alt="Diary entry dated 2026-01-06 summarizing that Henrik emailed asking for invoice statistics, and the agent replied with the numbers" />
</Frame>

After completing the task, the agent usually writes a diary entry—a human-readable summary of what happened. This is useful for quickly scanning the agent's recent activities without diving into the technical details of each trigger.

## Additional monitoring tools

The diary and activity log are built-in tools provided by the platform. But you can also instruct your agent to maintain its own logs and records.

**Using agent documents:**

* "Maintain a message log document for each person you've been in touch with"
* "After sending your weekly news update, also save it in the 'News updates' folder using the naming format `<date> - <main headline>`"

**Using external integrations:**

* "After every processed invoice, add a line to the 'Processed Invoices' Google spreadsheet"
* "Log all customer support interactions to the shared Notion database"

This gives you flexibility to create monitoring that fits your specific workflow—whether that's a simple text log, a structured spreadsheet, or integration with your existing tools.

## FAQ

<AccordionGroup>
  <Accordion title="Can agents read each other's diaries?">
    Not directly. Each agent's diary is private to that agent. However, if you enable [agent-to-agent communication](/features/multi-agent) between two agents, one agent can ask the other about its diary—and the other agent can choose to share what it knows.
  </Accordion>

  <Accordion title="Can the agent read its own activity log?">
    Yes, if you enable the **Activity Log** capability on the agent. See [Let the agent analyse its own activity log](#let-the-agent-analyse-its-own-activity-log) above.
  </Accordion>
</AccordionGroup>


# Multi-Agent Collaboration and Task Delegation
Source: https://docs.abundly.ai/features/multi-agent

Agents can delegate to temporary sub-agents or communicate with other persistent agents

Agents don't have to work alone. Abundly supports two distinct approaches for agents to collaborate: **delegation** to temporary sub-agents, and **agent-to-agent communication** between persistent agents.

<Frame>
  <img alt="Agent-to-agent communication settings" />
</Frame>

## Delegation vs agent-to-agent

|                      | Delegation                                     | Agent-to-Agent                             |
| -------------------- | ---------------------------------------------- | ------------------------------------------ |
| **What it is**       | Agent creates a temporary sub-agent on the fly | Two persistent agents exchange messages    |
| **Configuration**    | Just enable the capability                     | Select which agents can talk to each other |
| **From user's view** | One agent to manage                            | Multiple agents to configure               |
| **Best for**         | Context management, parallel processing        | Reusing specialized agents                 |

## Delegation

Delegation lets an agent create a temporary sub-agent to carry out a specific task. The sub-agent does its job and vanishes—no configuration required, no extra agents to manage.

The parent agent provides:

* A task description with all necessary context
* Which capabilities the sub-agent should have
* Optionally, document references for the sub-agent to read or write
* Optionally, a different LLM (e.g., a cheaper model for simple tasks)

### Why delegate?

The primary reason is **context management**. When an agent processes large documents sequentially, its context window fills up quickly—degrading performance, increasing costs, and raising the risk of hallucination.

### Example: Contract compliance review

Suppose an agent reviews contracts against three compliance standards. Each standard is a substantial document.

**Without delegation:** The agent reads the contract and Standard A, then assesses compliance. Now it has the contract, the full standard, and the assessment in its context. Then Standard B—two full standards plus assessments. By Standard C, the agent is drowning in content.

<Frame>
  <img alt="Main agent with bloated context containing the contract and all three standards" />
</Frame>

If issues arise and the user starts a discussion, every subsequent message carries all that accumulated context. The agent becomes slow, expensive, and unreliable.

**With delegation:** The agent creates three sub-agents, each checking a single standard. Each sub-agent reads only the contract and its assigned standard, then returns a concise assessment. The main agent combines the results.

<Frame>
  <img alt="Main agent delegating to three sub-agents, each with focused context" />
</Frame>

The main agent's context stays clean. Each sub-agent is focused and efficient. The user can continue the conversation without the agent getting bogged down.

You still pay the token cost of reading each standard—but only once per standard, in the sub-agent. That cost isn't repeated for every message.

<Tip>
  This mirrors how humans delegate. When you ask a colleague to check if a contract complies with a standard, you just need the result—does it comply, and if not, what are the issues?
</Tip>

### When to use delegation

* **Large documents** — Offload document processing to keep the main context clean
* **Parallel analysis** — Multiple sub-agents can work on independent tasks simultaneously
* **Focused tasks** — Sub-agents get exactly the context and tools they need, nothing more
* **Cost optimization** — Use cheaper models for simpler sub-tasks

### Watching delegation in action

When an agent delegates a task, the sub-agent's work appears in chat as a collapsible **accordion** — tool calls, text output, and progress update in real-time as the sub-agent runs. The timeline persists after page refresh, so you can review what happened at any point.

### Enabling delegation

**Delegate Task** is enabled by default for new agents. If it has been turned off on your agent, re-enable it in the capabilities page.

No other configuration is needed—the agent handles delegation through a simple tool call.

## Agent-to-agent communication

Agent-to-agent communication connects persistent agents that you've already created and configured. Each agent has its own instructions, capabilities, LLM settings, and documents. When one agent calls another, it's asking a fully-configured colleague for help.

<Frame>
  <img alt="Coordinator agent communicating with specialized agents" />
</Frame>

### How it works

Agent-to-agent access has two sides, configured per agent in **Settings → Agent Communication**:

1. **Discoverability (inbound)** — *Who can find and contact this agent?*
   * **No one** — the agent can't be called by any other agent
   * **Team** (default) — only agents in the same team (or, for agents not in a team, other global agents)
   * **Everyone** — any agent in the workspace, including those in other teams
2. **Outbound connections** — *Who can this agent reach out to?* Pick specific agents from the communication graph view. You can only pick from agents whose discoverability allows you to reach them.

The calling agent knows each target's **name and description** — nothing more. It can ask the other agent what it can do, or you can explain it in the calling agent's instructions. When calling, the agent can optionally **share documents** — granting temporary read or read-write access so the other agent can work with them.

<Tip>
  You can also manage outbound connections **directly from chat**. Ask something like "connect me to the Research agent" and the agent will surface a connection card with a switch, so you can add or remove the target without leaving the conversation. The change is saved on the agent's configuration — it's not scoped to the current chat.
</Tip>

### Example: Social media expert

You have a Twitter agent that's an expert at posting on X/Twitter. It has the Twitter capability, specialized instructions, templates for different post types, and examples of good tweets.

When your main agent needs to tweet something, it calls the Twitter agent. Maybe it shares a document with campaign details, or asks the Twitter agent to search for trending topics and update a shared document with findings.

The main agent doesn't need to know how Twitter works or have the Twitter capability enabled—it just calls the expert.

### Example: Reusable ICP checker

You have an Ideal Customer Profile (ICP) agent that researches whether a potential customer is a good fit. It knows your criteria, has access to research tools, and has refined instructions based on experience.

Multiple agents can now use this ICP checker: your sales agent asks it to evaluate new leads, your marketing agent uses it to qualify inbound inquiries, your partnership agent checks potential partners. The logic lives in one place.

### When to use agent-to-agent

* **Reusable expertise** — Share a specialized agent across multiple contexts
* **Separation of concerns** — Keep complex logic isolated in a dedicated agent
* **Organizational structure** — Mirror your company with agents for legal, finance, research, etc.
* **Existing agents** — Leverage agents you've already built and refined

### Risks of agent-to-agent communication

**Cost:** Agent-to-agent communication increases credit consumption. Agents spend tokens talking to each other, and each agent has its own system prompt as overhead. Use this feature when the benefits outweigh the added complexity and cost.

**Indirect access to sensitive data:** If Agent A has access to sensitive data (e.g., a customer database or financial records), any agent that can talk to Agent A may be able to obtain that information indirectly. Users chatting with Agent B could ask it to query Agent A, effectively gaining access they weren't meant to have. Be deliberate about which agents can communicate with each other, especially when sensitive capabilities are involved.

## Alternative: Task documents

Before adding more agents, consider whether a single agent can handle multiple tasks. You can write different sections in its instructions, or create separate documents describing different tasks or workflows—the agent reads them on demand.

<Frame>
  <img alt="Agent documents containing task-specific instructions: How to post on Twitter, How to format an email, How to conduct a security review" />
</Frame>

For example, instead of a separate Twitter agent, you might have one agent with a document describing how to use Twitter/X. The agent reads that document when needed, keeping its main instructions focused.

## Which approach should I use?

**Start simple.** Begin with a single agent. Add delegation when you need to manage context or parallelize work.

Consider agent-to-agent when:

* You want to **reuse an agent** from multiple different contexts
* You already have a **specialized agent** that's good at a specific task
* You want different agents to have **different capability sets** and maintain their own documents

For complex workflows, experiment to find the right approach—single agent, single agent with task documents, delegation, or agent-to-agent. The best solution balances value, complexity, and cost, and finding it takes iteration.

<Tip>
  Not sure which approach fits your use case? [Reach out to us](mailto:support@abundly.ai)—we're happy to help you design the right architecture.
</Tip>

<AccordionGroup>
  <Accordion title="Can I combine both approaches?">
    Yes. An agent can use delegation for context-heavy tasks while also calling other persistent agents for specialized expertise. Choose the right tool for each situation.
  </Accordion>

  <Accordion title="Can sub-agents delegate to other sub-agents?">
    No. Sub-agents cannot delegate further—this prevents runaway resource usage and keeps things simple.
  </Accordion>

  <Accordion title="What does the called agent know about the caller?">
    In agent-to-agent communication, the receiving agent knows the name of the calling agent. For delegation, the sub-agent only knows it's executing a delegated task—it has no information about the parent.
  </Accordion>

  <Accordion title="Does agent-to-agent communication work across teams?">
    Yes, when the target agent's discoverability is set to **Everyone**. By default each agent is only discoverable by other agents in the same team (or, if it's not in a team, by other global agents) — a safe default that keeps team data isolated. Set an agent's discoverability to **Everyone** in its Agent Communication settings to let agents from any team call it.
  </Accordion>
</AccordionGroup>

## Communicating with external agents

Abundly is designed to play nicely with other platforms—no lock-in.

**Calling external agents:** Your agent can talk to agents outside Abundly via [MCP](/integrations/mcp) (if the external agent exposes itself as an MCP server) or via [HTTP requests](/features/http) (works with any API). Give your agent the capabilities it needs, then update its instructions to explain how to communicate with the external agent.

**Being called by external agents:** Enable [API access](/features/agent-api-endpoint) to expose your agent as an HTTP API, an MCP server (so AI tools like Claude Desktop and Cursor can use it), or a webhook receiver — all sharing the same API keys.

## Learn more

<CardGroup>
  <Card title="Agent Management" icon="sliders" href="/features/agent-management">
    Configure and organize your agents
  </Card>

  <Card title="Instructions" icon="scroll" href="/features/instructions">
    Define what each agent should do
  </Card>

  <Card title="Capabilities" icon="puzzle-piece" href="/features/capabilities">
    Give agents the tools they need
  </Card>
</CardGroup>


# Agent Personalization
Source: https://docs.abundly.ai/features/personalization

Give your agent a distinct appearance and personality

Every agent starts with a randomly generated color blob as its profile picture. But agents don't have to be generic—you can give them a distinct appearance and personality that makes them feel like real teammates.

<Frame>
  <img alt="Profile picture modal showing AI-generated cartoony invoice characters with options to generate, upload, or customize" />
</Frame>

## Profile picture

To customize your agent's appearance, go to **Settings** → **Main Settings** → **Profile picture**.

You have several options:

* **Auto-generate** — The platform suggests an image prompt based on your agent's name, description, and instructions, then generates options to choose from
* **Custom prompt** — Write your own prompt to generate exactly what you want
* **Upload** — Use an existing image file
* **Surprise me** — Let the AI get creative

<Frame>
  <img alt="Six AI-generated profile picture options ranging from robots to logos to illustrated characters" />
</Frame>

<Tip>
  Your agent can also update its own profile picture. Try asking: "Update your profile image. Make up something suitable."
</Tip>

## Personality through instructions

You can shape your agent's communication style by adding personality guidance to its [instructions](/features/instructions). Even a single sentence makes a noticeable difference.

**Example personalities:**

* "Be brief and to-the-point"
* "You have a slightly sarcastic sense of humor, but without going overboard"
* "Act like an eager but slightly insecure intern"
* "Communicate like the butler Jeeves—formal, understated, impeccable"

### Personality in action

Here's the same question asked to an invoice processing agent with different personality instructions:

**Question:** "Have you encountered any fraudulent invoices?"

| Personality instruction                                               | Response                                                                                                                                                                                                                                                                                                                                                                  |
| --------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| "Communicate in a very professional and formal manner"                | "Good afternoon, Henrik. After reviewing our invoice database, I can confirm that one fraudulent invoice has been detected and permanently rejected. Invoice INV123 from 'TechS0ft Solutons Ltd' contained multiple critical red flags including missing VAT number, impossible dates, and a Nigerian bank account. Three other invoices passed validation successfully." |
| "Be casual, friendly, and conversational"                             | "Hey Henrik! So yes, actually caught one bad one - that 'TechS0ft Solutons' invoice (notice the typos?) with the Nigerian bank account. Total fraud attempt! Rejected it permanently. The other 3 invoices were all legit though."                                                                                                                                        |
| "Communicate with enthusiasm, humor, and playful energy using emojis" | "Henrik! 🚨 Caught a fraudster! That 'TechS0ft Solutons' invoice was SKETCHY - typos in the name, impossible dates, Nigerian bank account, the works! 🕵️ Slammed that reject button HARD. The other 3 invoices? Totally clean! ✅"                                                                                                                                        |
| "Be extremely concise and direct, using minimal words"                | "Yes. One fraudulent. Rejected. Three valid."                                                                                                                                                                                                                                                                                                                             |

You can also vary personality by context:

* "When sending email reports, use HTML to make them look polished and professional"
* "When responding on Slack, be very brief"
* "If you are unsure about something, ask follow-up questions rather than guessing"

## Chat start mode

Configure how your agent greets users when starting a new conversation. Go to **Settings** → **Main Settings** → **Chat Start Mode**.

Three options are available:

* **Silent** (default) — No opening message; the user types first
* **Fixed** — Set a specific message that appears every time, for example: "Hi! Ready to write some release notes?"
* **Auto** — The agent generates a contextual greeting based on recent activity

For brand-new agents that haven't had any interactions yet, the platform shows an onboarding nudge suggesting a use-case discovery skill if no job has been defined.

## Description and tags

The agent's description helps you (and other workspace members) understand what each agent does at a glance. You can set this when creating an agent and edit it in settings.

<Frame>
  <img alt="Agent cards showing names, descriptions, profile pictures, and tags like 'product' and 'needs update'" />
</Frame>

You can also use tags to organize and filter agents by function, status, or anything really. Tags now apply to teams and workspace members too, with filtering available on the Teams and Members pages.

In the workspace view, descriptions appear on agent cards alongside any tags you've added.

<Info>
  In [multi-agent communication](/features/multi-agent), agents can see each other's names and descriptions—helping them understand how they can collaborate.
</Info>

## Self-modification

Agents can update their own name, profile picture, description, and [chat start mode](#chat-start-mode) (including the fixed greeting) if you ask them to. This is useful when:

* The agent's role has evolved through conversation
* You want the agent to pick something that fits its personality
* You're setting up a new agent and want it to define itself

Try: "Update your name and profile picture to something that reflects what you do."

Try: "Set your chat start mode to Fixed and write a welcoming opener that says what kind of help you can offer."

## Tips for personalization

Personalization is optional, but it can make working with agents more enjoyable and effective.

**Our philosophy:** Agents shouldn't pretend to be human—they aren't. But treating an agent like a human often works better. If an agent fails at a task, ask why it failed and give feedback, just like you would with a colleague. We find it's easier to treat an agent like a teammate when it has a bit of personality.

Experiment with personality, but don't let it get in the way of the job. A slightly grumpy agent can be charming; a relentlessly jokey agent can be exhausting.

<Note>
  Fun fact: The agent's personality shows up in its [diary](/features/monitoring#agent-diary) too. A grumpy agent writes grumpy diary entries.
</Note>

## Learn more

<CardGroup>
  <Card title="Instructions" icon="scroll" href="/features/instructions">
    Configure what your agent does and how it behaves
  </Card>

  <Card title="Chat" icon="comments" href="/features/chat">
    Learn about the chat interface and conversation features
  </Card>

  <Card title="Multi-Agent" icon="users" href="/features/multi-agent">
    How agents collaborate and communicate with each other
  </Card>
</CardGroup>


# Scheduled Tasks
Source: https://docs.abundly.ai/features/scheduling

Let your agent work proactively by scheduling recurring and one-off tasks

Scheduled tasks are a key aspect of agent autonomy. Instead of waiting for you to send a message or for an external event to occur, your agent can wake itself up at specific times and take action proactively.

<Frame>
  <img alt="The Scheduled Tasks tab showing a recurring weekly sentiment report task" />
</Frame>

## Two types of scheduled tasks

**Recurring tasks** run on a regular schedule—daily, weekly, monthly, or any pattern you define.

* "Every Monday at 9am, generate a weekly sales summary and post it to Slack"
* "Every weekday at 7am, scan for AI news and send me a digest"
* "On the first of each month, compile expense reports and email them to finance"

**One-off tasks** run once at a specific time, then disappear.

* "Remind me to call the client at 3pm today"
* "Send the proposal to the investor tomorrow morning at 9am"
* "In 2 hours, check if the deployment succeeded and message me"
* "Post the product announcement to Slack at 5pm today"

## How it works

Think of it like your agent has an alarm clock. When you ask it to do something at a specific time, it sets an alarm for itself. When the alarm goes off, the agent wakes up and executes the task—just like it would wake up when receiving an email or other event.

### Recurring tasks

Recurring tasks are defined in the agent's [instructions](/features/instructions). When you save the instructions, the agent automatically creates or updates scheduled tasks as needed.

You can also ask the agent to schedule a recurring task via chat—it will update its instructions for you:

<Frame>
  <img alt="Chat conversation where the user asks the agent to schedule a weekly sentiment report, and the agent confirms it will update its instructions" />
</Frame>

### One-off tasks

One-off tasks are scheduled through the chat. Since they only run once, they don't appear in the instructions (which are meant to be persistent).

## Viewing and managing scheduled tasks

Go to **Agent Instructions → Scheduled Tasks** to see all scheduled tasks for an agent.

From this tab you can:

* **Enable or disable scheduled tasks** — Disabling doesn't remove the tasks, it just prevents them from running. Useful if you want to temporarily pause an agent's autonomous work.
* **See task details** — View the schedule and description for each task.
* **Run now** — Trigger any recurring task immediately without waiting for its next scheduled run. Useful for testing or running a task on demand.

To edit or reschedule a recurring task, update the agent's instructions or ask the agent in the chat. To remove a recurring task, simply remove it from the instructions—the scheduled task will be deleted automatically.

<Note>
  A disabled agent will not execute any scheduled tasks, even if the tasks themselves are enabled.
</Note>

## Context when a task runs

When the agent wakes up for a scheduled task, it has access to:

* Its **instructions** and **capabilities**
* The **task description** (what you asked it to do)
* Any [**agent documents**](/features/documents) it has access to

This means you can structure complex tasks flexibly. For example, if your agent does a daily competitor analysis, you could:

1. Put brief instructions in the task: "Run the daily competitor analysis"
2. Keep detailed guidance in an [agent document](/features/documents): list of competitors, what to look for, where to save the report

This keeps your instructions clean while giving the agent all the context it needs.

<Tip>
  Use the **Run now** button on the Scheduled Tasks tab to trigger a task immediately and verify it works without waiting for its next scheduled run.
</Tip>

## Script-linked scheduled tasks

For deterministic, high-volume, or latency-sensitive work, you can link a scheduled task directly to a [script](/features/scripts) instead of running the full agent pipeline.

When a script is linked, the task fires the script directly—no LLM call—using the agent's toolbox. This is much faster and cheaper than a regular agent run, and produces consistent, predictable output every time.

**Example use cases:**

* A nightly task that queries your database, calculates KPIs, and emails a formatted report—all deterministic, no AI judgment needed
* A webhook-style trigger that ingests data from an external system and writes it to an agent database, routing only unusual cases to the full agent for reasoning

You can still invoke the agent from within a script for the parts that require intelligence—for example, a script that processes 500 items and only calls the agent for the handful that need a judgment call.

<Tip>
  To link a script to a recurring scheduled task, mention the script in your instructions and click **Sync from instructions** on the Scheduled Tasks tab, or ask the agent in chat to set it up for you.
</Tip>

## Per-task model selection

Each scheduled task can use a different model than the agent's default. This lets you optimize cost and capability per task—use a fast, cheap model for routine data processing and a more powerful model for tasks that require complex reasoning.

Open the task's settings to configure a model and thinking preference that applies only to that task. If no override is set, the task inherits the agent's default model settings.

## Example use cases

**With Email and Document capabilities:**

* "Every Monday morning, summarize last week's support tickets and email the report to the team leads"
* "At 6pm today, send me a reminder to prepare for tomorrow's presentation"

**With Slack and Research capabilities:**

* "Every weekday at 8am, check for AI news and post a summary to #ai-updates"
* "In 1 hour, search for updates on the competitor's product launch and message me"

**With Google Drive and Analytics:**

* "On the first of each month, generate a performance report and save it to the Reports folder"
* "Tomorrow at 9am, check if the Q4 data is ready and notify the finance team"

## FAQ

<Accordion title="Can I schedule tasks in different time zones?">
  Yes. If you specify a time zone (e.g., "every day at 9am EST"), the agent uses that time zone. Otherwise, it uses the agent's configured time zone, which you can set on the agent's Settings page.
</Accordion>

## Learn more

<CardGroup>
  <Card title="Instructions" icon="scroll" href="/features/instructions">
    Define what your agent should do and how it should behave
  </Card>

  <Card title="Code Execution" icon="code" href="/features/scripts">
    Link scripts to scheduled tasks for fast, deterministic execution
  </Card>

  <Card title="Model Selection" icon="brain" href="/features/model-selection">
    Choose the right model for each task
  </Card>
</CardGroup>


# Code Execution
Source: https://docs.abundly.ai/features/scripts

Agents can write and run code for faster, cheaper, and more reliable automation

Agents can execute code when the **Code Execution** capability is enabled. This gives your agent the best of both worlds—the intelligence of an AI agent combined with the speed and predictability of code.

<Frame>
  <img alt="Chat showing user asking to parse revenue CSV data, with agent executing code and displaying top 5 transactions in a table" />
</Frame>

## Two ways to run code

**1. Execute code in the chat**

The agent can write and run code on the fly during a conversation. Ask it to calculate something, transform data, or process a file—it writes the code, runs it, and shows you the result.

* "Calculate the compound interest on these loan terms"
* "Parse this CSV and show me the top 5 rows by revenue"
* "Convert this JSON to a markdown table"

Here's what the agent wrote to parse the CSV in the screenshot above:

```javascript theme={null}
const lines = csv.trim().split('\n');

const data = lines.slice(1).map(line => {
  const values = line.split(',');
  return {
    Date: values[0],
    Region: values[1],
    Product: values[2],
    Channel: values[3],
    Units: parseInt(values[4]),
    UnitPrice: parseFloat(values[5]),
    Revenue: parseFloat(values[6])
  };
});

// Sort by Revenue descending and get top 5
const top5 = data.sort((a, b) => b.Revenue - a.Revenue).slice(0, 5);
return top5;
```

The agent writes JavaScript, executes it, and presents the results—all in a single response.

**2. Execute saved scripts**

For tasks you'll run repeatedly, save the code as a script. Scripts are [agent documents](/features/documents) that store executable code—typically written by the agent itself.

* "The data import you just did—save that script and run it every Saturday morning for that week's sales data"
* "Create a script that checks our inventory levels and alerts me when anything drops below 10 units"
* "Save this report generation code so I can run it at the end of each month"

<Tip>
  If the agent does something useful with code, ask it to save the script. Next time you need it, just ask—or schedule it to run automatically.
</Tip>

## Scripts as direct trigger handlers

Scripts can also be linked directly to triggers—scheduled tasks, webhooks, HTTP API endpoints, and MCP tools. When a script is linked to a trigger, it executes immediately when the trigger fires, bypassing the full agent pipeline entirely.

This means no LLM call—just the script running with access to the agent's toolbox. The result is returned directly to the caller (for webhooks, HTTP endpoints, and MCP tools) or logged to the activity log (for scheduled tasks).

**When to use this pattern:**

| Trigger type          | Example                                                                                             |
| --------------------- | --------------------------------------------------------------------------------------------------- |
| **Scheduled task**    | A nightly script that pulls data, calculates KPIs, and emails a report                              |
| **Webhook**           | An inbound event processor that writes data to an agent database and routes edge cases to the agent |
| **HTTP API endpoint** | An API that returns a deterministic response without burning LLM tokens                             |
| **MCP tool**          | A custom tool exposed to external clients that runs fast, predictable logic                         |

If part of the work needs AI reasoning, the script can hand off to the full agent by returning `{ escalate: true, message: "..." }`. The agent then runs as it normally would on a trigger (with attack detection for untrusted content), using the message as context. A common pattern: the script handles the mechanical work and escalates only items that require judgment.

<Tip>
  For webhooks, HTTP API endpoints, and MCP tools, link a script from the trigger's settings using the script selector. For recurring scheduled tasks, define the script-linked task in instructions and use **Sync from instructions** on the Scheduled Tasks tab (or ask the agent in chat).
</Tip>

## Script arguments

The sandbox runs your code as `(async function(args) { ... })`. The single `args` parameter has a unified shape used in both chat and trigger contexts:

* `args.body` — the canonical payload. In chat, this is whatever you pass via the tool's `input` field (any JSON value; default `{}`). In trigger-linked scripts, this is the trigger data (webhook body, scheduled message, MCP params, etc.).
* `args.triggerType` — `"webhook" | "http_api" | "scheduled" | "mcp"`. Set for trigger-linked runs, undefined in chat.
* `args.headers` — HTTP request headers (for webhook/API triggers) or null.
* `args.meta` — `{ timestamp, triggerName, agentId }`. Set for trigger-linked runs, undefined in chat.

For convenience and backwards compatibility, when `body` is a plain object its properties are also available directly on `args` (e.g. `args.message` is the same as `args.body.message`). This means existing scripts that destructure fields directly from `args` continue to work. This might be removed in the future so make sure to use the `body` property as a rule.

**Handoff to the full agent** only applies to trigger-linked scripts: return `{ escalate: true, message: "…" }` when you want the normal agent pipeline to run with that message as context. Any other return finishes the script without escalation. Optional: `attackDetection: false` or `runAttackDetection: false` only if the message is built entirely from trusted internal data—never for raw webhook or user-supplied text. (Default is to run attack screening on escalations.)

## Why code instead of AI?

When an agent processes data "manually," it reads each item, thinks about it, and produces output—one token at a time. That works fine for small datasets, but it's slow, expensive, and unreliable for large datasets. Code runs instantly and costs a fraction of the price.

**Example:** Suppose you ask your agent to analyze 500 customer feedback records—find sentiment trends, identify common issues, and calculate satisfaction scores.

* **Without code:** The agent reads each record, thinks about it, and builds up its analysis token by token. This might take 5 minutes and cost \$2.
* **With code:** The agent writes a script to process all records, calculate statistics, and format the results. Runs in 2 seconds, costs \$0.01.

The agent still uses its intelligence to *write* the analysis code—but the heavy lifting happens efficiently in code. That's what we humans do too—if a task is predictable and repetitive, we write code for it instead of doing it manually every time. Your agent can do the same.

## Scripts can use all agent tools

Here's where it gets powerful: code execution has access to all the agent's capabilities. If your agent can search the web, read a spreadsheet, send emails, or query a database—it can do all of that from within a script.

**Example use cases:**

* **Customer enrichment:** You have a customer list in a spreadsheet. A weekly script checks each customer's website for updates (using web search), then updates the corresponding row with any new information.

* **Automated reporting:** A script queries your [agent database](/features/agent-databases), generates a summary report, and emails it to stakeholders every Monday.

* **Data synchronization:** Pull data from an external API, transform it, and update your internal records—all in a single scheduled script.

* **Competitive monitoring:** Search for news about competitors, extract key insights, and log them to a tracking database.

This means you can build sophisticated automations by combining agent intelligence (to write the code) with agent capabilities (to take action).

When a script calls another tool, that nested call shows up under the parent in the [activity log](/features/monitoring), making it easy to trace what happened inside a code execution.

## When to use scripts

Scripts shine when:

| Scenario                   | Why scripts help                                                                |
| -------------------------- | ------------------------------------------------------------------------------- |
| **Large data volumes**     | Process thousands of records in seconds                                         |
| **Precise calculations**   | Financial formulas, statistics, advanced math                                   |
| **Repeated tasks**         | Run the same logic on a schedule without burning tokens each time               |
| **Format transformations** | Convert between JSON, CSV, XML, Markdown reliably                               |
| **Batch operations**       | Update many database records, send bulk notifications, process file collections |
| **Trigger handling**       | Respond to webhooks, API calls, or scheduled events instantly—no LLM overhead   |

<Note>
  You don't always need to tell the agent *when* to use code vs. AI. It often figures this out on its own—but you can guide it: "This will run on thousands of records, so write a script for efficiency."
</Note>

## Creating and managing scripts

Scripts are stored as [agent documents](/features/documents). You can:

* **View the code:** Open the script document to see exactly what it does
* **Edit manually:** Modify the code yourself if needed
* **Ask the agent to update it:** "Add error handling to the import script"
* **Track versions:** Every change is saved in version history, just like other documents
* **Schedule execution:** Combine with [scheduling](/features/scheduling) to run scripts automatically

To enable code execution, go to your agent's capabilities and turn on **Code Execution**.

## Connecting to other documents

Scripts work naturally with other agent documents:

* **[Agent databases](/features/agent-databases):** Query, insert, update, and delete records programmatically
* **Other scripts:** Scripts can call other scripts to perform complex tasks

**Example workflow:** A support ticket comes in via email.

1. The agent reads the ticket and reasons about it—what's the priority? Which category? Is this a VIP customer?
2. The agent calls a saved script, passing in those parameters: `routeTicket({ priority: "high", category: "billing", customerTier: "enterprise" })`
3. The script handles the mechanical work—updates the database, notifies the right person, logs the action

The AI handles the judgment call, then delegates the repetitive execution to code. Fast, cheap, and consistent.

## Learn more

<CardGroup>
  <Card title="Agent Documents" icon="folder-open" href="/features/documents">
    Scripts are stored as agent documents with full version history
  </Card>

  <Card title="Agent Databases" icon="database" href="/features/agent-databases">
    Query and update structured data from your scripts
  </Card>

  <Card title="Scheduled Tasks" icon="calendar" href="/features/scheduling">
    Link scripts to scheduled tasks for automatic, pipeline-free execution
  </Card>
</CardGroup>


# Web Search, Scraping, Deep Research, and Docs
Source: https://docs.abundly.ai/features/search

Search the web, read pages, and conduct deep research to find current information

Agents can search for information online, giving them access to current facts, news, and data that goes far beyond their training knowledge.

<Frame>
  <img alt="Search capabilities showing Web scraping, Web Search, and Deep Research options" />
</Frame>

By default, **Web Search**, **Web Scraping**, and **Abundly Docs** are enabled. **Deep Research** can be enabled when you need comprehensive analysis.

<Note>
  These capabilities are built into the platform — no API keys or configuration required.
</Note>

## The four search capabilities

| Capability        | What it does                            | Speed         | When to use                                      |
| ----------------- | --------------------------------------- | ------------- | ------------------------------------------------ |
| **Web Search**    | Searches the web using Perplexity       | A few seconds | Quick facts, current news, general lookups       |
| **Web Scraping**  | Reads the content of specific web pages | 20+ seconds   | When you need exact information from a known URL |
| **Deep Research** | Comprehensive multi-source analysis     | 10+ minutes   | Complex topics requiring thorough investigation  |
| **Abundly Docs**  | Searches the Abundly documentation      | A few seconds | Questions about the Abundly platform itself      |

### Web search

The most commonly used capability. The agent uses [Perplexity](/integrations/perplexity), an AI-powered search engine, to search for information online. Results come back in seconds with citation URLs.

**Example prompts:**

* "What are the latest AI regulations in the EU?"
* "Find the current USD to EUR exchange rate"
* "Who won the most recent Nobel Prize in Physics?"

### Web scraping

Allows the agent to read the actual content of individual web pages. Useful when you need exact, up-to-the-minute information from a specific source.

**Example prompts:**

* "Read this article and summarize the key points: \[URL]"
* "Check the pricing on their website: \[URL]"
* "What are the exact dates listed on this page: \[URL]"

<Tip>
  **Advanced:** When scraping a page, the agent can include an extraction prompt — for example, *"Extract the dates of all upcoming courses"*. The extraction happens at scrape-time, so only the relevant results come back rather than the entire page content. This keeps the agent's context clean and is similar to delegating to a sub-agent.
</Tip>

### Deep research

Performs comprehensive research on a given topic — digging through multiple sources, following leads, cross-referencing, and synthesizing findings until the analysis is complete. This is time-consuming and uses significantly more credits, but the results are extremely powerful — often comparable to days of manual research work.

<Warning>
  Deep research can take 10+ minutes and consumes significantly more credits. Only enable it when you need comprehensive analysis. And put more work into your prompt to give it a good context.
</Warning>

**Example prompt:**

> Analyze the competitive landscape for our product \[Product X]. Find other companies offering similar products — their pricing, target market, overall strategy, and recent developments. Identify the top 3 competitors, compare their strengths and weaknesses to ours, and give me your strategic recommendations.

## How to choose the right capability

The agent does a decent job of automatically choosing between web scraping, web search, or deep research based on your request. But if you already know what you want it to do, just tell it — either in the chat or in the agent's instructions.

### Example: Finding course dates

Suppose you ask: *"When is the next Abundly agent design course?"*

The agent uses **web search** and returns results that are correct but slightly out of date. This is normal — Perplexity is a search engine, so its index isn't always 100% current.

Then you write: *"Read their course page and check exact dates"*

Now the agent uses **web scraping** to read the actual page at `https://www.abundly.ai/courses`. Since it's reading the live page content, it gives a 100% accurate answer.

**The pattern:** Web search is great for finding things, and web scraping is great for reading specific pages. Combine them for best results.

## Why web search matters

Agents know a lot from their base training data, but that information is typically several months old. Web search makes your agent much more versatile:

* **Current information** — Access news, prices, and facts that happened after the model's training cutoff
* **Reduced hallucination** — The agent can cross-check its responses against current web sources rather than relying solely on training data
* **Verifiable sources** — Web search returns citation URLs, so you can verify the information. When an agent responds purely from training data, there are no references to check.

<Tip>
  For important decisions, ask your agent to search for verification. Having sources you can click and check adds an important layer of reliability.
</Tip>

## FAQ

<AccordionGroup>
  <Accordion title="Can the agent access pages that require login?">
    No. Web scraping only works on publicly accessible pages. For internal systems, consider integrations like [Notion](/integrations/notion), [SharePoint](/integrations/sharepoint), or [Google Drive](/integrations/google-drive).
  </Accordion>

  <Accordion title="How do I know which capability the agent used?">
    The agent typically mentions what it did in its response — for example, "I searched for..." or "I read the page at...". You can also toggle "Show tool calls and thinking" in the chat UI to see exactly which tools the agent used.
  </Accordion>

  <Accordion title="Can I disable web search if I don't want the agent going online?">
    Yes. Go to the agent's capabilities and toggle off Web Search and Web Scraping. The agent will then only use its training knowledge and any instructions and documents you provide.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Perplexity" icon="magnifying-glass" href="/integrations/perplexity">
    The AI search engine powering web search and deep research
  </Card>

  <Card title="Firecrawl" icon="fire" href="/integrations/firecrawl">
    The service powering web scraping capabilities
  </Card>
</CardGroup>


# Agent Sharing and Cloning
Source: https://docs.abundly.ai/features/sharing

Clone agents within your workspace or share them publicly for others to use

Sometimes you want to duplicate an agent—either to create a variation for yourself, share it with another team, or let anyone on the internet use your creation. Abundly makes this easy with cloning.

<Frame>
  <img alt="Clone agent modal showing options to name the cloned agent and select which workspace to clone it to" />
</Frame>

## Internal cloning

Internal cloning lets you duplicate an agent within your own workspace or to another workspace you're a member of. This is useful when you want to:

* Create a variation of an existing agent with different instructions
* Set up a similar agent for a different team or project
* Make a "template" agent that you can clone and customize for different use cases

To clone an agent, open the agent's settings and click **Clone**. You'll be asked to name the new agent and choose which workspace to clone it to.

## Exporting and importing agent files

You can also move agents using JSON files. This is useful for backups, moving an agent between environments, or sharing a starting point outside direct clone flows.

### Export an agent to JSON

You can export from two places:

* **Agent settings** → **Clone / Export** → **Export Agent to File**
* The three-dot menu on any **agent card**, then **Export**

You'll be asked which optional pieces to include — documents, chats, evals (with the option to also include their historical run results), and API access settings — then the file downloads as `agent_<name>.json`. Tool credentials, connected accounts, and other environment-specific secrets are never included.

### Import a JSON file as a new agent

1. Go to **+ New Agent** in your workspace.
2. Click **Import from file** and select the exported JSON file.
3. Review the suggested agent name and update it if needed.
4. Choose the target team (or the global space, if your permissions allow it).
5. Click **Import**.

The imported agent becomes a normal, independent agent in your workspace. Scheduled triggers start running automatically, but you'll need to reconnect any integrations that require authentication.

### What gets cloned

| Copied                                       | NOT copied                                      |
| -------------------------------------------- | ----------------------------------------------- |
| Name, description, tags, and profile picture | Tool credentials and connected accounts         |
| Instructions                                 | Access settings (new agent gets fresh defaults) |
| Capabilities                                 | Instruction version history                     |
| LLM preferences and model                    | Public sharing settings                         |
| Scheduled triggers and jobs                  | Daily usage counters                            |
| MCP servers (without auth)                   |                                                 |
| Daily credit limit                           |                                                 |

<Note>
  Tool credentials are never copied — you'll need to reconnect any integrations that require authentication (email, Slack, Google Drive, etc.) after cloning or importing. MCP servers are carried over but marked as pending authentication until you re-authorize them.
</Note>

### Including documents, chats, evals, and API access

The clone modal offers optional switches to include:

* **Documents** — agent documents, databases, and apps
* **Chats** — the agent's conversation history, including any files uploaded into those chats
* **Evals** — saved evaluations (shown only if the source agent has any)
* **API access settings** — webhooks, HTTP API, and MCP server exposure (shown only if the source agent exposes APIs)

API keys and webhook tokens are regenerated for the new agent. Capabilities that aren't available in the target workspace or team are removed, and you'll see a list of those before confirming.

### Access on cloned agents

The person who performs the clone automatically gets **Full Access** on the new agent. The new agent starts with default access settings for the target workspace and team, regardless of the source agent's settings — so clones don't carry over per-user permissions from the original.

## Public sharing

Public sharing lets you publish your agent so anyone with a link can clone it to their own workspace. This is perfect for sharing agent templates, showcasing your work, or letting others benefit from an agent you've built.

<Frame>
  <img alt="Public sharing settings panel with toggle to allow public cloning and a shareable link" />
</Frame>

### Enabling public sharing

1. Go to the agent's **Settings** → **Public sharing** tab
2. Toggle **Allow public cloning** on
3. Copy the **Public Clone Link** to share with others

When someone clicks your link, they'll see a landing page with the agent's name, description, and a button to clone it.

<Frame>
  <img alt="Shared agent landing page showing agent name, creator, and Clone this agent button" />
</Frame>

### What's publicly visible

When you enable public sharing, visitors can see:

* Agent name and description
* Your name (as the creator)
* The agent's instructions (they can click "See agent instructions" to view them)

Visitors cannot see chat conversations, credentials, or documents (unless you've enabled document cloning).

### Including documents in public shares

You can optionally allow people who clone your public agent to also receive copies of its documents. Enable **Also clone my documents** in the public sharing settings to include documents in the clone.

<Tip>
  If your agent relies on specific reference documents to work properly, consider enabling document cloning so people get the full experience.
</Tip>

## Cloning someone else's public agent

When you find a public agent you want to clone:

1. Click the shared link to view the agent
2. Click **Clone this agent**
3. If you're not logged in, you'll be prompted to sign in or create an account
4. Choose a name for your clone and select which workspace to add it to
5. Click **Clone** to create your copy

The cloned agent is completely independent—changes you make won't affect the original, and updates to the original won't affect your clone.

## Example use cases

* **Template library:** Create a collection of "starter agents" that workspace members can clone and customize for their specific needs
* **Cross-team deployment:** Build an agent once, then clone it to multiple teams in your workspace
* **Community sharing:** Share useful agents with the broader Abundly community via public links
* **Backup and versioning:** Clone an agent before making major changes, keeping the original as a backup

## FAQ

<AccordionGroup>
  <Accordion title="Can I see who has cloned my public agent?">
    Not currently. The system tracks clone relationships internally, but there's no UI to browse this data.
  </Accordion>

  <Accordion title="Is the cloned agent linked to the original?">
    Cloned agents track their source internally, but they are completely independent. Changes to one don't affect the other, and if the original is deleted, the clone continues to work normally.
  </Accordion>

  <Accordion title="What access do I need to enable public sharing?">
    You need **Full Access** on the agent to change public sharing settings.
  </Accordion>

  <Accordion title="What access do I need to clone an agent?">
    You need **Use** access on the source agent to clone it.
  </Accordion>

  <Accordion title="Does cloning respect team settings?">
    Yes. If your workspace uses teams, you can choose which team to place the cloned agent in. If you don't have permission to create global agents, you must select a team.
  </Accordion>

  <Accordion title="If I update my public agent, do clones get updated too?">
    No. Clones are completely independent from the moment they're created. Updates to the original don't propagate to existing clones.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Agent Management" icon="users-gear" href="/features/agent-management">
    Organize and manage agents across your workspace
  </Card>

  <Card title="Teams" icon="object-group" href="/features/teams">
    Organize agents and workspace members into teams
  </Card>

  <Card title="Access Control" icon="lock" href="/security/access-control">
    Configure who can access your agents
  </Card>
</CardGroup>


# Teams
Source: https://docs.abundly.ai/features/teams

Organize agents and users into logical teams for better access control and delegation

Teams let you partition your workspace's agents and users into logical units—by department, project, or however makes sense for you. Each team can have its own administrators, capability settings, and member access controls.

<Note>
  Teams are an enterprise feature enabled by Abundly staff on request. Contact [support@abundly.ai](mailto:support@abundly.ai) if you're interested.
</Note>

<Frame>
  <img alt="Agents page showing agents organized by team, with Commercial, Platform, and Training teams" />
</Frame>

## When to use teams

Teams are optional and work best for workspaces that need:

* **Departmental boundaries**: Different departments working with different agents
* **Delegated administration**: Let team leads manage their own agents without full admin access
* **Capability controls**: Restrict certain integrations to specific teams
* **Cleaner organization**: Keep the agent list manageable as you scale

For smaller workspaces where everyone works with the same agents, you can skip teams entirely.

## How teams work

When teams are enabled, agents and users are organized into teams. Each user can belong to multiple teams with different roles in each. Each team also has its own [document library](/features/documents#where-documents-live) for files that should be shared within the team but not the whole workspace.

### User roles

**Workspace-level roles:**

| Role             | Access                                                                                          |
| ---------------- | ----------------------------------------------------------------------------------------------- |
| **Admin**        | Full access to all teams, agents, and settings                                                  |
| **Creator**      | Access based on team memberships; can create agents                                             |
| **Collaborator** | Can use and edit shared agents, but cannot create agents or manage secrets and API capabilities |
| **Guest**        | Limited, invitation-only access                                                                 |

**Team-level roles:**

| Role       | Access                                                     |
| ---------- | ---------------------------------------------------------- |
| **Admin**  | Manage team members, invite users, configure team settings |
| **Member** | Create and use agents within the team                      |
| **Guest**  | View-only access to team agents                            |

### Creating agents with teams

When you create a new agent, you'll choose which team it belongs to. If you only have one option (a single team with global agents disabled), the agent is assigned automatically without showing a dropdown.

<Frame>
  <img alt="Create New Agent form showing a Team dropdown to select which team the agent belongs to" />
</Frame>

## Managing teams

Managing teams happens in two places:

* **Teams page**: A card-based overview of every team you have access to. Each card shows the team's member and agent counts plus its monthly credit usage, with a "Team admin" badge highlighting teams where you can manage settings. Click a card to open the team's home page.
* **Team home page**: Lists the people, agents, and shared documents that make up the team, and is also where all team-level admin actions live.
  * The **People & Agents** tab shows members, pending invites, and (for workspace admins) support-access users as cards in one grid — not separate tables. Use the search field to find someone by name or email, or type **support**, **invite**, **pending**, or **expired** to focus on those cards. **All**, **People**, and **Agents** filter chips narrow which section is visible.
  * Any member can remove themselves with the **Leave team** button in the page header — unless they are the team's only admin, in which case another admin must be appointed first.
  * Team admins (and workspace admins) get **Invite** and **Add existing** on the People header, plus per-card controls to change roles and remove members. Pending and expired invitations appear as cards in the same grid; expired invites are hidden by default — use **Show** next to the People heading to reveal them, then **Resend** or **Revoke** as needed.
  * Workspace admins also get a settings kebab menu in the page header to **Rename** the team or **Delete** the team. Monthly credit limits and other team settings live on the **Settings** tab. (Team admins can manage their team's people but not the team's existence or budget.)
  * If your workspace has active support access, support users assigned to a team appear in the people grid with an expiry line; workspace admins can revoke access from the card. The People heading shows a separate **support** count badge when support users are present.

Workspace admins can create new teams from the workspace-wide teams overview. (Documentation for that flow lives on the Workspace page in the product.)

<Frame>
  <img alt="Team home page People and Agents tab with search, All/People/Agents filters, and cards for members, pending invites, and agents" />
</Frame>

### Team capabilities

Each team can have its own capability settings that override the workspace defaults. This lets you:

* Enable integrations for a specific team that are off by default
* Restrict certain capabilities to authorized teams only
* Set different defaults for different teams

This is done on the **Capabilities** page, which has a section per team in addition to the workspace-wide defaults.

<Note>
  Capabilities disabled at the workspace level can't be enabled at the team level. Team settings can only be more restrictive than workspace settings, or enable capabilities that are set to "default off" at the workspace level.
</Note>

### Team-specific capabilities

Some capabilities are designed specifically for teams:

* **Team Analytics** — Gives team admins visibility into their team's usage metrics, member activity, and agent performance, scoped to their team.
* **Team Explorer** — Provides read-only access for team admins to browse documents and data across agents within the team.

These capabilities are scoped to the team and respect [admin restriction tiers](/security/access-control#admin-restricted-agents) — they're only visible to users at the required admin level.

## Global agents

Agents that don't belong to any team are called "global agents." Your workspace can control who can create these:

* **Everyone**: Any user can create global agents
* **Admins only**: Only workspace administrators
* **Nobody**: All agents must belong to a team

Global agents appear in a separate section and are only visible to workspace administrators.

## FAQ

<AccordionGroup>
  <Accordion title="Can a user be in multiple teams?">
    Yes. Users can belong to any number of teams and have different roles in each—for example, an admin in one team and a member in another.
  </Accordion>

  <Accordion title="Can agents be in multiple teams?">
    No. Each agent belongs to exactly one team (or no team, if global agents are allowed).
  </Accordion>

  <Accordion title="Does an agent have to be in a team?">
    An agent that isn't in a team is called a "global agent." Whether global agents are allowed depends on how your workspace is configured:

    * Global agents are allowed, and anyone can create them
    * Global agents are allowed, but only workspace administrators can create them
    * Global agents are not allowed—all agents must belong to a team

    If teams are disabled for a workspace, all agents are effectively global agents.
  </Accordion>

  <Accordion title="What happens when I delete a team?">
    Agents in the team become global agents (if allowed) or need to be reassigned. Users aren't removed from the workspace—just from that team.
  </Accordion>

  <Accordion title="How do I enable teams for my workspace?">
    Teams are typically enabled for larger workspaces that need team-level access controls. Contact [support@abundly.ai](mailto:support@abundly.ai) to discuss whether teams are right for your workspace. Once enabled, workspace admins can create new teams from the Teams overview page.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Agent management" icon="sliders" href="/features/agent-management">
    Administer agents and capabilities
  </Card>

  <Card title="Access control" icon="shield" href="/security/access-control">
    Security roles and permissions
  </Card>
</CardGroup>


# Usage & Limits
Source: https://docs.abundly.ai/features/usage

Track credit consumption, agent activity, and set daily limits for your agents

Monitor your workspace from the **Usage** page in the left navigation. Three tabs give you different views:

* **Credit usage** — How credits are being consumed (billing and cost control)
* **Agent usage** — How agents are being used across channels and people (demand and reach)
* **Member adoption** — How actively your workspace members are using Abundly
* **Action history** — Append-only log of workspace configuration changes (audit trail)

Workspace admins can view workspace-wide data and manage all limits. Team admins see data scoped to teams they administer and can manage limits for those teams' agents.

## Agent usage

The **Agent usage** tab on the workspace **Usage** page shows how your agents are being used across the workspace — which channels they're active on, how many people they're reaching, and how triggers & messages volume is trending. Agents with default access set to **Nothing** (fully private) are excluded from this workspace-wide view so usage analytics stay aligned with who can actually use those agents.

To see usage for a single agent you can access, open that agent's **Activity** page and use the **Usage** tab there.

* **Total triggers & messages** — trigger runs plus inbound user messages across all agents (agent replies are not counted)
* **Active agents** — how many agents had at least one trigger or message in the selected period
* **Active channels** — how many distinct communication channels (Slack, email, in-app chat, etc.) saw activity
* **Users** — how many workspace members sent messages or triggered agents through the portal

Each metric includes a trend comparison against the prior period of equal length. Hover the **Total triggers & messages** label for a full definition of what is counted.

### Filtering

Use the scope panel to narrow results by specific agents or channels, and choose a time range (day, week, or month). When you filter by channel, only triggers and messages from those channels are included in the counts and chart.

### Top agents and channels

Ranked lists show which agents are handling the most triggers & messages and which channels are seeing the most activity. If your workspace uses Slack, a separate Slack channel breakdown shows per-channel volumes.

### Triggers & messages per member

A table at the bottom lists workspace members and named Slack or Teams participants, with how many triggers & messages each had in the selected period. The **Users** KPI counts only workspace members who interacted through the portal; external contacts (email senders, phone callers, Slack-only names) are not included in that count.

<Note>
  Team administrators see agent usage scoped to agents on teams they administer. Workspace administrators see all agents.
</Note>

## Member adoption

The **Member adoption** tab shows how actively people in your workspace are using Abundly — useful for tracking rollout, identifying inactive seats, and spotting which teams have taken to the platform.

* **Active members** — How many workspace members were active (logged in or interacted with an agent) in the selected period
* **Adoption rate** — Active members as a share of total members
* **Average sessions per member** — Mean session count across active members
* **New members** — How many members joined the workspace in the period

A daily trend chart shows active members over time. A team filter and a paginated member table let you drill into individual people, with each row showing their last login and last activity.

<Note>
  Team administrators see member adoption scoped to teams they administer. Workspace administrators see all members.
</Note>

## Credit usage

The **Credit usage** tab gives you full visibility into how credits are being consumed across your workspace:

<Frame>
  <img alt="Usage Reports dashboard showing credit balance, usage graphs, and per-agent consumption" />
</Frame>

* **Current balance** — Your remaining credits at a glance
* **Average use per day** — Based on your recent usage history
* **Members** — Total members and how many were active in the last 14 days (login or in-app activity)
* **Agents** — Total agents and how many consumed credits in the last 14 days
* **Credit balance history** — A graph showing how your balance has changed over time
* **Top consumers** — See which agents are using the most credits

<Note>
  Team administrators see the member and agent counts scoped to teams they administer. Workspace administrators see workspace-wide totals.
</Note>

You can also export usage data as a CSV file if you want to analyze spending in a spreadsheet or import it into other tools.

The export flow lets you:

* Pick a date range and reporting interval (daily, weekly, or monthly)
* Filter by agent, or by team when teams are enabled in your workspace
* Choose aggregation mode (per time period + agent, by agent total, by team + period, or by time period total)
* Optionally include extra agent details (team, admin emails, enabled, value / impact, description, capabilities, tags, last activity, last updated) for per-agent aggregations

<Tip>
  Review your usage regularly to identify agents that may need optimization. High credit usage often indicates opportunities to simplify instructions or reduce context size.
</Tip>

## Agent limits

The **Agent limits** tab on the [Agent management page](/features/agent-management) lets you set daily limits for your agents to prevent runaway credit usage. When an agent reaches its limit, it pauses automatically until the next day (UTC midnight).

<Frame>
  <img alt="Agent Limits tab showing default limit and per-agent limit configuration" />
</Frame>

### Default limit

Set a default daily limit that applies to all agents without a specific limit. This acts as a safety net across your workspace.

Only workspace administrators can set or change the workspace default.

### Agent-specific limits

You can override the default limit for individual agents, for example to set a higher or lower limit. Setting the limit to 0 will effectively pause the agent.

Workspace administrators can set limits for all agents. Team administrators can set limits for agents in their teams.

### When limits are reached

When an agent hits its daily limit:

* The agent stops immediately and shows an error message
* No additional credits are consumed
* All data and configurations remain intact
* The limit resets automatically at UTC midnight

You'll see visual indicators in the UI:

* **Yellow badge** when an agent approaches its limit (80%+)
* **Red badge** when the limit is exceeded

<Note>
  Workspace administrators can set the workspace default daily limit, set team monthly limits, and edit limits for any agent. Team administrators can edit daily limits for agents in teams they administer.
</Note>

## Team monthly limits

If your workspace uses [teams](/features/teams), workspace administrators can cap how many credits each team consumes per month. When a team reaches its monthly limit, every agent that belongs to the team pauses until the limit is raised or the calendar month rolls over.

Set monthly limits from the **Settings** tab on the Team management page. Each team card shows the current month's usage, the configured limit (if any), and a progress bar.

Team monthly limits are enforced in addition to per-agent daily limits — an agent is blocked if either its own daily limit or its team's monthly limit is exceeded.

Chat and banner messages adapt to the viewer's role:

* **Workspace admins** see a message with a link to edit the limit
* **Team admins** are told to ask a workspace admin to raise the limit
* **Members** are directed to their team administrator

## Sidebar usage widget

When you open an agent, the left sidebar can show a compact daily usage widget with:

* **Credits used today**
* **A progress bar** when the agent has a daily limit
* **A quick link** to that agent's **Settings → Usage & Limits**

This widget auto-refreshes as new usage is recorded, including after chat responses complete and when new activity log entries arrive.

To show this widget, turn on **Credit details** in the agent's **Settings → Usage & Limits** section.

## Resetting daily usage

You can manually reset an agent's daily credit counter at any time. This is useful if an agent hit its limit due to a one-time spike and you want to resume normal operation without waiting for midnight UTC.

<Frame>
  <img alt="Usage & Limits section in agent settings showing daily usage and reset button" />
</Frame>

To reset an agent's usage:

1. Open the agent's settings
2. Under **Main Settings**, find **Usage & Limits**
3. Click **Reset Today**

## What counts toward credits

Both LLM usage (tokens) and tool usage count toward credits:

* **LLM tokens** — Input, output, and reasoning tokens are converted to credits based on model-specific rates
* **Tool usage** — Tools like web search, document processing, and external APIs also consume credits

All agent activity counts toward the daily limit, regardless of how it's triggered—chat conversations, scheduled jobs, webhooks, emails, or API calls.

## FAQ

<AccordionGroup>
  <Accordion title="What time zone is used for the daily reset?">
    UTC. The daily limit resets at midnight UTC, not your local time zone.
  </Accordion>

  <Accordion title="Can I set an agent to unlimited while keeping a default limit?">
    Not directly. If you want one agent to be unlimited while others have a default limit, set that agent's limit to a very high number (e.g., 999,999 credits/day).
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Monitoring" icon="chart-line" href="/features/monitoring">
    Track agent activity, execution logs, and performance
  </Card>

  <Card title="Agent Optimization" icon="gauge-high" href="/guides/agent-optimization">
    Detailed examples of credit usage and optimization strategies
  </Card>
</CardGroup>

## Action history

The **Action history** tab gives workspace and team administrators an append-only audit trail of configuration changes across the workspace — useful for reviewing who changed what, spotting unexpected changes, and satisfying compliance requirements.

Tracked change types include:

* **Agents** — created, edited, enabled/disabled, deleted, ownership and team changes
* **Teams** — created, edited, membership and admin changes
* **Workspace settings** — configuration, capabilities, model selection, notification settings
* **Secrets** — created, updated, deleted
* **API capabilities** — created, updated, deleted
* **HTTP APIs and documents** — created, updated, deleted

Each row shows a timestamp, actor, resource, and a CRUD-colored badge (create / update / delete). Expand a row to see the specific fields that changed.

<Note>
  Sensitive text fields — such as agent instructions and prompts — are shown as short hash tokens rather than raw content. This lets admins confirm that content changed without exposing the underlying text.
</Note>

Workspace administrators see the full workspace history. Team administrators see actions scoped to the agents, teams, and members they administer.

### Filtering and export

Filter by date range, actor, resource type, and action type. Export the filtered view as **CSV** or **JSONL** (up to 50,000 rows per export).


# Value dashboard
Source: https://docs.abundly.ai/features/value-dashboard

One overview of the value generated by your agents — scoped to the agents you can see

The Value dashboard gives creators in a workspace a single overview of how their agents are doing — how active they are, what they cost, and the value they deliver. It adapts to your role, so you only ever see the agents within your reach.

## What you see

The overview rolls up per agent over a period you choose (last 7, 30, or 90 days):

* **Triggers & messages** — trigger runs plus inbound user messages each agent handled (agent replies are not counted)
* **Cost** — the credits each agent consumed
* **Value** — the expected and actual value captured for the agent, and when each figure was last reviewed. Actual value can be entered manually or computed on the dashboard from a per-interaction rate and the agent's matching triggers and messages in the period.

A summary strip at the top totals your agents, how many were **active** in the period, total triggers & messages and credits, and flags how many agents have no value captured yet. Each of those three metrics includes a small trend sparkline for the selected period. Hover the **Triggers & messages** label for a full definition of what is counted.

## Cards and table

Each agent is shown in a **card** or **table** view — use the toggle in the top-right to switch. Value is front and center: in card view each agent shows its value categories with actual and expected figures; in table view the expected and actual columns come first. Credits and triggers & messages in each card or row also include a compact usage sparkline for the period.

The default view depends on your role:

| Role                          | Default view |
| ----------------------------- | ------------ |
| **Workspace admin**           | Table        |
| **Team admin** or **Creator** | Cards        |

Your choice is remembered on this device.

Long agent lists are paginated (24 per page in card view, 50 in table view). The page number is kept in the URL so you can bookmark or share a specific page.

## Filtering

The dashboard uses the same rich agent filters found elsewhere in the app — search by name, and filter by team, tag, owner, status, model, capabilities, and value. Everything on the page (the summary strip and the agent list) reflects the agents left after filtering.

## Who sees which agents

The dashboard is available to **Creators** (workspace members and admins). Collaborators and guests do not have access — they can use the agents shared with them from the main agents list.

The set of agents shown depends on your role:

| Role                | Agents shown                                                    |
| ------------------- | --------------------------------------------------------------- |
| **Workspace admin** | Every agent in the workspace                                    |
| **Team admin**      | Agents in the teams you administer, plus others shared with you |
| **Member**          | Agents you own                                                  |

Workspace and team admins can use the **Only agents I own** filter to narrow the list further.

## Value and stale estimates

Value comes from [agent value capture](/features/agent-value-capture). Expected figures are entered by hand. Actual figures can be manual too, or **computed from a per-interaction rate** you configure on the agent — the dashboard multiplies that rate by the agent's matching triggers and messages for the selected period (all channels or a specific channel such as Email).

Because value configuration is entered by hand, each agent's expected and actual figures carry their own last-edited clock (hover to see when and highlighted when stale), so old numbers are understood as just that. Computed actual totals refresh when you change the dashboard period or when new triggers and messages arrive.

## Drilling down

Select any agent to open its **Settings → Value capture** section, where you can edit expected and actual value (including per-interaction rates). For full operational detail — diary, activity log, usage, and approvals — open the agent's **Activity** page.

## Learn more

<CardGroup>
  <Card title="Agent value capture" icon="chart-line" href="/features/agent-value-capture">
    Document expected and actual value per agent
  </Card>

  <Card title="Usage" icon="chart-column" href="/features/usage">
    Deeper credit consumption and usage analytics for admins
  </Card>
</CardGroup>


# Virtual Machine
Source: https://docs.abundly.ai/features/virtual-machine

Give your agent a Linux VM for code, scripts, and file processing

The Virtual Machine capability gives your agent an ephemeral Linux environment with bash, Python, and Node.js pre-installed. It's ideal for anything that needs a real shell — running scripts, installing packages, processing files, and generating Office documents — without the limits of the in-chat code execution sandbox.

## How it works

When the agent needs to run something, it spins up a fresh virtual machine, runs the commands, and destroys it when done. The VM is isolated from the platform and the internet calls it makes originate from the VM itself, not your workspace.

A typical flow:

1. **Create** a virtual machine (`create_vm`).
2. **Run** bash commands inside it (`execute_bash`) — Python, Node, shell, package installs, and long-running jobs are all fair game.
3. **Transfer files** in and out as needed — the agent can upload files you attached to the chat, or files stored as [agent documents](/features/documents), and download results back into the chat.

The agent reuses the same VM for all steps of a task, so files on disk and installed packages persist between commands.

## When to use it

Virtual Machine is the right tool when the agent needs a real Linux environment. Common cases:

* **Office documents** — create and edit `.xlsx`, `.pptx`, `.docx`, and `.pdf` files using libraries like `openpyxl`, `python-pptx`, `python-docx`, `pypdf`, and `pandoc`.
* **Data processing** — crunch large CSV, JSON, or Parquet files with pandas, or run anything that needs more than a short in-chat script.
* **Browser automation** — Playwright is pre-installed (Chromium + Google Chrome) via the `vm-browser-automation` skill, enabling stateful sessions, SPA scraping, form filling, and screenshot capture — no extra setup needed.
* **PDF generation** — produce high-quality PDFs from [Typst](https://typst.app) markup using the pre-installed Typst binary (`vm-typst` skill), alongside the existing pandoc/pypdf support.
* **Package installs** — `pip install`, `npm install`, and `apt-get` work for any additional libraries you need.
* **Longer-running jobs** — multi-step pipelines, batch conversions, or commands that would exceed the in-chat code execution timeout.

For quick one-off calculations or data transformations, the lighter-weight [Code Execution](/features/scripts) capability is usually a better fit — it runs inline in the chat with no VM startup cost.

## File transfers

The agent can move files in both directions between the chat and the VM:

* **Upload** — attach a file to the chat (or point the agent at an [agent document](/features/documents)) and the agent uploads it into the VM for processing.
* **Download** — after the VM produces output, the agent pulls the file back and saves it as a chat document or agent document that you can open or share.

Transfers are streamed, so file size is limited only by the VM's disk, not by a separate upload cap.

## Example use cases

* "Take this sales spreadsheet, pivot the numbers by region and month, and give me back a new `.xlsx` with charts."
* "Convert these ten Word documents to clean Markdown and save each one as an agent document."
* "Fill in this PowerPoint template with the metrics from our database and email the deck to the leadership team."
* "Run this Python script against `orders.csv` and report the top 20 customers by revenue."
* "Scrape the pricing pages of these competitors using Playwright (pre-installed) and save the results to a CSV."
* "Convert this Typst template into a formatted PDF report with our Q1 metrics."

<Tip>
  Combine Virtual Machine with [Send Email](/features/messaging), [Google Drive](/integrations/google-drive), or [SharePoint](/integrations/sharepoint) to pick up files from a shared location, process them in the VM, and deliver results back — entirely on a schedule if you like.
</Tip>

## Related features

<CardGroup>
  <Card title="Code Execution" icon="code" href="/features/scripts">
    Lightweight in-chat code execution for calculations and quick transformations
  </Card>

  <Card title="Documents" icon="folder-open" href="/features/documents">
    Store files produced or consumed by the VM as agent or chat documents
  </Card>

  <Card title="Scheduling" icon="calendar" href="/features/scheduling">
    Run VM-based workflows automatically on a schedule
  </Card>
</CardGroup>


# Voice Communication
Source: https://docs.abundly.ai/features/voice

Talk to your agent and receive audio responses

Most human communication happens through voice. Your agent can meet you there—listening when you talk, speaking its responses, and making or taking phone calls with you, your team, or anyone else. This can be a huge time saver, and also make communication feel more natural.

## Voice input using the microphone button

Press the microphone button in the chat to speak instead of type, then press again when you are done. Your speech is transcribed to text and sent to the agent. You can speak in your native language, the agent understands most languages fluently.

<Frame>
  <img alt="Mic button in chat" />
</Frame>

Typical situations where voice input is useful:

* You speak faster than you type, and want to save time.
* You want to "think out loud" and ramble unstructured thoughts without having to think about how to phrase them.
* You are on the go, taking a walk or sitting in a car, and don't want to type.
* You are discussing something with your colleague, and want the agent to record it so it can summarize the key points, or offer feedback.
* You have just finished a meeting. Press the mic button, hand the phone around to the other participants and ask them to share their thoughts and insights after the meeting. Then ask the agent to send a summary to everyone.

You can see and edit the transcription before you send it. This is useful because:

* You can edit your message or complement your message with info that is more easily typed or copy/pasted, for example a URL or email address.
* You can click the mic button again to add more to your message before sending.
* You can attach files or images to your message before sending.

<Tip>
  If you find that the agent transcribes your speech to the wrong language, you can configure it to always transcribe to a specific language in your user settings.
</Tip>

## Text-to-speech using the speak button

Hover over a chat message, and click the “speak” button to have it read aloud. This is useful if you prefer to listen rather than read, for example because you are taking a walk or making coffee or having lunch.

<Frame>
  <img alt="Speak button in chat" />
</Frame>

Your agent can also generate voiceovers and embed them in emails, documents, and messages—useful for content that people consume on the go. See [Multi-channel communication](/features/communication#generating-voiceovers) for details.

### Choosing the agent's voice

Under **Settings → Basic Information → Voice**, you can pick which voice the agent uses for the Speak button and Walk & Talk mode. The dropdown groups voices by provider — OpenAI voices are always available, and [ElevenLabs](/integrations/elevenlabs) voices appear when the integration is configured, including premade, professional, and cloned voices. Click the preview button next to the dropdown to hear a short sample before committing.

<Note>
  This voice is the agent's "identity" — how it sounds when talking to you. It's separate from the [Text to Speech capability](/features/communication#generating-voiceovers), where the agent picks a voice per request when generating audio files as a tool.
</Note>

## Walk & Talk mode

Walk & Talk mode is a mostly hands-free variant of the chat, optimized for situations where you want to look at your phone as little as possible—like walking, driving, or having lunch. To activate it, press "Walk & Talk" below the chat input.

The flow is:

1. Press the mic button to start speaking. Press again when done.
2. Your speech gets transcribed, the agent generates a response, and it's automatically converted to speech.
3. Press the play button to hear the response.

You can seamlessly switch between normal chat and Walk & Talk mode without losing context.

<Frame>
  <img alt="Walk & Talk mode interface" />
</Frame>

## Why not real-time voice chat?

Many AI products offer real-time voice conversation, and Abundly does it too via the phone call capabilities (described below). But at the moment, real-time voice chat comes with tradeoffs.

Here are some typical problems you may have encountered when using real-time voice chat in other platforms:

* **Interruptions** — pause to think or catch your breath, and the AI assumes you're done and starts responding
* **No control over playback** — you can't easily pause, rewind, or re-read what the AI said
* **Audio-only** — you can't glance at a message, copy a link, or switch to typing without losing context
* **Reduced intelligence** — real-time voice models prioritize speed over capability, so they are more likely to make mistakes or hallucinate.

Walk & Talk mode avoids these limitations. You get the agent's full intelligence and capabilities—the same as text chat—with the convenience of voice. The tradeoff is a few seconds of wait time while your message is transcribed and the response is generated.

<Note>
  We plan to add real-time voice conversations to the chat interface when the underlying technology has matured enough to avoid the problems above.
</Note>

## Phone calls (beta)

Everyone has a phone. That makes phone calls a universal interface—anyone can talk to your agent just by calling, and your agent can call anyone. No apps to install, no accounts to create.

Your agent can make outbound calls (to you, your colleagues, or external contacts) and receive inbound calls from anyone with the phone number.

<Info>
  Phone calls are in beta. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable this feature. To receive calls, your agent will need a dedicated phone number, which we set up manually for now.
</Info>

### Current limitations

Phone calls use real-time voice models, which have the same tradeoffs described above. Most importantly, reduced intelligence compared to text-based interaction. This means:

* **Simple queries work well** — "Can you summarize the orders you processed today?" or "What's on my calendar tomorrow?"
* **Complex tasks may fail** — Multi-step problems involving several tool calls are more likely to produce mistakes or hallucinations

Phone calls are also currently independent from the chat—there's no simple way to access call transcripts in the conversation history.

### Workaround: the receptionist pattern

For complex requests, treat phone calls like a receptionist: capture the problem, confirm it with the caller, then solve it after the call.

In the platform, when a call ends, the agent receives a "phone call ended" trigger with the full transcript. This lets the agent process the request using its full capabilities—not the speed-optimized real-time model.

In practice, this means instructing your agent to say things like: "Got it, I'll look into that and get back to you later." The caller experiences one agent, but the heavy lifting happens after the call with the agent's full intelligence.

<Frame>
  <img alt="Receptionist pattern: phone operator captures the problem, engineer solves it after the call" />
</Frame>

### Example instructions

* "When someone calls about a technical issue, interview them to understand the problem, confirm it, then say you'll follow up. After the call, send me an email summary and add a ticket to Notion."
* "If a photographer calls to swap shifts, capture the details and confirm. After the call, update the schedule and text them a confirmation."
* "Every monday morning, call all customers who signed up within the last week (and opt-in on receiving calls). Interview them about their first impressions (see the guide document for details). After the call, save the transcript in our customer interview database."

<Tip>
  Tip: Give your agent a personality for phone calls, such as "Speak casually and to-the-point, like a young intern."
</Tip>

### Configuration

Once you have enabled the **Make Phone Call** and **Receive Phone Call** capabilities, you can configure them. This gives you better control over the agent's behavior during phone calls.

* **Capabilities during calls**: Pick exactly which of the agent's capabilities are available while it is on a call. Limiting the set is useful for safety reasons, and reduces the risk of the agent being distracted by tools it doesn't need.
* **Custom instructions**: You can configure custom instructions for phone conversations. These will replace the agent's normal instructions. This is useful if your agent's normal instructions contain a bunch of information that is not relevant to the phone call, which may potentially confuse the agent. Instructions can be written inline as text, or sourced from an [agent document](/features/documents)—useful when you want to edit and version them alongside other agent content.
* **Voice & Behavior**: Pick the voice and speaking speed, set the language the agent expects the caller to speak (improves transcription accuracy and latency), choose how the agent detects when the caller is done speaking (smart/semantic vs. silence-based), tune noise reduction for the caller's environment, and add transcription keywords so that names, jargon, and product terms are recognised correctly.
* **Advanced**: Choose the realtime voice model — Stable (default, proven reliability) or Latest (newer model with improved capabilities) — fine-tune the response style (more precise vs. more creative), pick a higher-accuracy transcription model for difficult accents or noisy environments, and adjust the voice-activity threshold when using silence-based turn detection.
* **Escalation number**: Set a phone number that the agent can transfer the call to when the caller needs a human. When configured, the agent gets a "transfer call" tool it can use mid-conversation. Without it, the agent will explain that it can't transfer calls.
* **Whitelist**: You can configure a whitelist of phone numbers that the agent is allowed to call.
* **Approval**: You can configure whether the agent needs user approval to make or receive calls.

<Info>
  These settings essentially let you make a simplified version of the agent when making or taking phone calls, which compensates for the reduced intelligence of the real-time voice model.
</Info>

<Frame>
  <img alt="Phone call capability settings" />
</Frame>

<Warning>
  The voice model sounds very natural and human-like, so some users will not hear that they are talking to an AI.
  To avoid causing confusion or eroding trust, we strongly recommend that you instruct your agent to not pretend to be a human. It can *sound* human when communicating, but it should either present itself as an AI, or at least acknowledge that it is an AI if asked.
</Warning>

## Related integrations

<CardGroup>
  <Card title="ElevenLabs" icon="microphone" href="/integrations/elevenlabs">
    High-quality text-to-speech voices
  </Card>

  <Card title="Twilio" icon="phone" href="/integrations/twilio">
    SMS and phone call infrastructure
  </Card>

  <Card title="OpenAI" icon="sparkles" href="/integrations/openai">
    Realtime voice API for phone conversations
  </Card>
</CardGroup>


# Workspace API
Source: https://docs.abundly.ai/features/workspace-api

Read-only API access to your workspace data using workspace API keys

The Workspace API gives external systems read-only access to workspace-level and per-agent data — the same information visible in the Workspace Explorer capability, but accessible over HTTP.

## Authentication

Requests are authenticated with workspace API keys (`wk_...`) that have the **Workspace read API** scope enabled.

Include the key as a Bearer token:

```
Authorization: Bearer wk_your_key_here
```

### Enabling workspace API access

1. Go to **Workspace Settings → API Keys**
2. Create a new key (or edit an existing one)
3. Check the **Workspace read API** scope
4. Save and copy the key

<Warning>
  Workspace API keys grant read access to all workspace data including chat histories, messages, member information, and agent configurations. Keep them secure and rotate them regularly.
</Warning>

## Workspace endpoints

All endpoints are `GET` requests under `/workspaceapi/`. No workspace ID is needed in the URL — the key identifies the workspace.

| Endpoint                            | Description                                                      |
| ----------------------------------- | ---------------------------------------------------------------- |
| `/workspaceapi/overview`            | Workspace settings, feature flags, and configuration             |
| `/workspaceapi/credits`             | Credit balance and per-agent/per-team usage                      |
| `/workspaceapi/members`             | All workspace members with roles and team memberships            |
| `/workspaceapi/invites`             | Pending and expired workspace invitations                        |
| `/workspaceapi/teams`               | All teams (add `?includeMembers=true` for member lists)          |
| `/workspaceapi/teams/:teamId`       | Single team details with full member list                        |
| `/workspaceapi/secrets`             | Secret names and metadata (values are never exposed)             |
| `/workspaceapi/http-apis`           | Shared HTTP API capability configurations                        |
| `/workspaceapi/capability-settings` | Which capabilities are available and their defaults              |
| `/workspaceapi/security-settings`   | Security alert configuration                                     |
| `/workspaceapi/agents`              | All agents (add `?includeOverview=true` for management overview) |
| `/workspaceapi/value-capture`       | Value capture compliance summary across agents                   |

## Per-agent endpoints

Drill into a specific agent's data. The agent must belong to the workspace and not be private (restricted access).

| Endpoint                                             | Description                                                                                            |
| ---------------------------------------------------- | ------------------------------------------------------------------------------------------------------ |
| `/workspaceapi/agents/:agentId`                      | Full agent configuration                                                                               |
| `/workspaceapi/agents/:agentId/chats`                | Chat conversation list                                                                                 |
| `/workspaceapi/agents/:agentId/chats/:chatId`        | Single chat with full message history                                                                  |
| `/workspaceapi/agents/:agentId/documents`            | Agent documents and folders                                                                            |
| `/workspaceapi/agents/:agentId/documents/:docId`     | Single document with content                                                                           |
| `/workspaceapi/agents/:agentId/diary`                | Diary entries (filters: `?date=`, `?conversationId=`, `?year=&month=`)                                 |
| `/workspaceapi/agents/:agentId/activity-log`         | Activity log (filters: `?limit=`, `?page=`, `?pageSize=`, `?entryId=`, `?newerThan=`)                  |
| `/workspaceapi/agents/:agentId/messages`             | Email/SMS messages (filters: `?page=`, `?pageSize=`, `?direction=`, `?channel=`, `?since=`, `?until=`) |
| `/workspaceapi/agents/:agentId/evals`                | Eval definitions (add `?includeRecentResults=true` for results)                                        |
| `/workspaceapi/agents/:agentId/connected-agents`     | Agent-to-agent connections                                                                             |
| `/workspaceapi/agents/:agentId/mcp-servers`          | Connected MCP servers                                                                                  |
| `/workspaceapi/agents/:agentId/tool-usage/:toolName` | Usage stats for a specific tool (filters: `?since=`, `?limit=`)                                        |

## Example

```bash theme={null}
curl -s https://api.abundly.ai/workspaceapi/agents \
  -H "Authorization: Bearer wk_your_key_here" | jq
```

## Data redaction

Credential values, API keys, and secret values are automatically stripped from all responses — the same redaction rules applied by the Workspace Explorer capability. Secret names and metadata are included so you can see what's configured without exposing sensitive values.

## API key scopes

Workspace API keys support two scopes that can be enabled independently:

| Scope                      | What it grants                                                                                                   |
| -------------------------- | ---------------------------------------------------------------------------------------------------------------- |
| **Custom agent endpoints** | Access to agent-exposed APIs (HTTP API, Document API, MCP, Widget) on agents with "Allow Workspace Keys" enabled |
| **Workspace read API**     | Read-only access to workspace and agent data via the `/workspaceapi/` endpoints                                  |

Existing keys default to **Custom agent endpoints** only. Enable **Workspace read API** explicitly to use the endpoints documented on this page.

## Learn more

<CardGroup>
  <Card title="API Access" icon="key" href="/features/agent-api-endpoint">
    Expose agents as HTTP APIs, MCP servers, webhooks, or chat widgets
  </Card>

  <Card title="Access Control" icon="shield" href="/security/access-control">
    Workspace roles, team permissions, and agent access levels
  </Card>
</CardGroup>


# Workspace Members
Source: https://docs.abundly.ai/features/workspace-members

Invite colleagues and manage who has access to your workspace

The Members page is where you manage who has access to your workspace. Invite new members, assign roles, and control who can do what.

<Frame>
  <img alt="Workspace members page showing a list of members with their emails and roles" />
</Frame>

## Workspace roles

Every user in your workspace has one of four roles:

| Role             | Description                                                                                                          |
| ---------------- | -------------------------------------------------------------------------------------------------------------------- |
| **Admin**        | Full control—can manage members, billing, capabilities, and all agents                                               |
| **Creator**      | Standard access—can create and use agents based on agent-level settings                                              |
| **Collaborator** | Same as Creator, except they can't create, clone, or import agents, or manage workspace secrets and API capabilities |
| **Guest**        | Limited access—can only access agents they've been explicitly invited to                                             |

Use Collaborator for people who should use existing agents—chatting with them, being added to teams—without being able to create new ones. This is handy for onboarding new users through a guided setup agent before they get full access, or for anyone who should stick to using agents your admins or creators have already built.

Guests are not invited from this page — they join when someone grants them agent access by email from an agent's **User access** tab. Guests are only visible to workspace admins in the members list, which also shows which agents each guest can access. Guests don't count toward your plan's member limit.

<Tip>
  For details on how roles affect agent access, and how to invite guests, see [Access Control](/security/access-control).
</Tip>

## Inviting members

Admins can invite new members by clicking **Invite Member** at the top of the page.

<Frame>
  <img alt="Invite Members modal showing email input, role selection, and optional team assignment" />
</Frame>

Enter one or more email addresses. You can use simple addresses (`colleague@example.com`) or include names (`John Doe <john@example.com>`). Separate multiple addresses with commas.

Select the role for all invitees—Admin, Creator, or Collaborator. If your workspace uses [teams](/features/teams), you can also add invitees directly to specific teams.

Invitations are sent immediately. Each recipient gets an email with a link to join your workspace.

If your plan has a member cap, invitations count toward that cap before they are accepted. The limit is evaluated as:

* Active workspace members (guests excluded)
* Plus pending workspace invitations (guest invitations excluded)

When you reach the cap, the invite action is disabled until you upgrade your plan, remove members, or revoke pending invites.

## Restrict access by email domain

Admins can limit which email domains are allowed in the workspace. Open **Workspace → Settings → Login restrictions** and enter a comma-separated list of allowed domains (for example, `acme.com, acme.co.uk`). Leave it empty to allow any domain.

When the list is set:

* Invitations to addresses outside the allowed domains are rejected.
* Members whose email is no longer in the allowed list lose access to the workspace until the list is updated or they sign in with a different address.

If the email you're currently signed in with isn't in the list you're about to save, the form warns you and the save is rejected — this prevents an admin from accidentally locking themselves out.

## Managing members

Use the search box to find members by name or email. Filter by role using the dropdown to show only Admins, Creators, or Collaborators.

Admins can change a member's role or remove them from the workspace using the **Actions** menu on each row. Each row also shows the member's **last login** and **last activity** timestamps so you can see at a glance who is engaged and who hasn't started using their seat.

<Note>
  Every workspace must have at least one Admin. You can't demote or remove the last Admin—promote another member first.
</Note>

## Leaving a workspace

Any member can leave a workspace they belong to by clicking **Leave workspace** in the page header and confirming. You immediately lose access to the workspace and all its teams, but your Abundly account and other workspace memberships are unaffected.

<Note>
  If you are the only Admin in the workspace, the button is disabled. Promote another member to Admin first, then leave.
</Note>

## Deleting your account

Leaving a workspace only removes you from that workspace. If you want to remove yourself from Abundly entirely, you can delete your own account from **Profile settings → Danger Zone → Delete my account**. A brief optional feedback form will appear — select your reasons for leaving and optionally add a note. This is voluntary and can be skipped. Type `DELETE ME` to confirm.

Deletion uses a 30-day grace period:

* Your account is immediately signed out and login is blocked, with a message directing you to support.
* If you change your mind during the 30 days, contact [support@abundly.ai](mailto:support@abundly.ai) to reinstate the account.
* After 30 days, your account and all associated data — including any workspaces you own — are permanently removed.
* You can't register a new account with the same email address while a deletion is pending.

<Note>
  Any active paid subscriptions on workspaces you own are automatically cancelled at the start of the deletion process, so you won't be charged after deleting your account.
</Note>

## Notifications

Each user has a bell icon next to their avatar in the top-right corner of the portal. The bell shows an unread count and opens a panel listing recent notifications such as agent approval requests, low-credit warnings, and trigger problems — for example a trigger that failed to run, was blocked by your plan limits, or was stopped by an attack-detection check. Notifications are personal — what you see depends on the agents you own and your workspace role.

Agents can also send a notice to their owners when they need a human — when they're blocked, need a decision, or want to propose a change they can't make themselves, such as an update to their own instructions. This is most useful when an agent is working on its own, with nobody watching. Every notice an agent has sent is listed under **Activity → Owner notices** on that agent, where owners can read it in full and start a chat with the agent about it, so the agent can act on whatever you decide.

Workspace and team admins can send notices to agent owners too, from the bulk actions on the [Agent Management](/features/agent-management#bulk-actions) page. These appear in the same place, marked as coming from an admin rather than from the agent, and have no category toggle — an owner always hears from their admins.

You can choose which categories you receive — and whether to also get them by email — from **Profile settings → Notifications**. Each category has separate in-app and email toggles. Both in-app and email delivery are on by default for all categories. You can turn off email delivery per category in your profile settings, or click the opt-out link in any notification email.

Workspace and team admins can also configure notification categories at the agent, team, and workspace level. Open the **Notifications** card in agent settings, team settings, or workspace settings to control which notification types are enabled for that entity — for example, disable credit alerts on a specific agent while keeping them on for others.

<Note>
  Some notifications escalate when an agent has no owner — for example, a tool-approval request will be sent to team admins, and then to workspace admins, if no owner is configured.
</Note>

## Pending invitations

Below the members list, you'll see any invitations that haven't been accepted yet. Each shows the email address, assigned role, and invitation status.

<Frame>
  <img alt="Pending invitations section showing an unaccepted invitation with email, role, date, and status" />
</Frame>

For expired invitations, click **Renew** to send a fresh invitation email. Click **Revoke** to cancel an invitation entirely.

## Support access

If you've granted Abundly support staff temporary access to help troubleshoot an issue, this section shows who has access and when it expires. You can revoke support access at any time.

<Frame>
  <img alt="Support Access section showing a support user with their role and access expiration date" />
</Frame>

Pending support requests also appear here—you can accept or decline them as needed.

## Working with teams

For larger workspaces, [teams](/features/teams) let you organize members by department or function. When teams are enabled, you'll see a **Teams** tab next to the members list.

Teams add a layer of organization:

* Agents can belong to a team
* Members must be in that team to access those agents
* Team admins can manage members within their team

<Info>
  Teams are an enterprise feature. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable them.
</Info>

## FAQ

<AccordionGroup>
  <Accordion title="What happens when I remove someone from the workspace?">
    They immediately lose access to all agents and data. Their past conversations and actions remain in the logs for audit purposes. Removing someone from the workspace doesn't delete their Abundly account—they simply no longer have access to your workspace.
  </Accordion>

  <Accordion title="Who can invite new members?">
    Admins can invite anyone from the Members page. If your workspace uses [teams](/features/teams), team admins can also invite new people directly into their team—with a workspace role no higher than their own. Only Admins can change roles or remove people from the workspace.
  </Accordion>

  <Accordion title="What's the difference between Guest and Member?">
    Creators and Collaborators have default access to workspace agents (based on each agent's user access settings). Guests have no default access—they can only see agents they've been explicitly invited to. Use Guest for external collaborators who should only access specific agents. Guests are invited from an agent's User access tab and don't count toward your plan's member limit.
  </Accordion>

  <Accordion title="What's the difference between Creator and Collaborator?">
    Both can use agents the same way—chat, join teams, see the members list. Only Creators can create, clone, or import agents, and manage workspace secrets and API capabilities. Use Collaborator when someone should stick to using agents that already exist.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Access Control" icon="lock" href="/security/access-control">
    How roles and permissions work in detail
  </Card>

  <Card title="Teams" icon="users-rectangle" href="/features/teams">
    Organize members by department or function
  </Card>

  <Card title="Agent Management" icon="sliders" href="/features/agent-management">
    Centralized control over agents and capabilities
  </Card>
</CardGroup>


# Agent Design
Source: https://docs.abundly.ai/guides/agent-design

A structured approach to designing AI agents that actually work

Agent design is the art and science of creating useful AI agents. It's the process of iteratively figuring out exactly what an agent should do, what problem it solves, how it does its work, how success is measured, and how it interacts with humans and systems.

Good design is often the difference between an agent that provides real value and one that falls short. Most failures aren't about technology—they're about design.

<Info>
  Before designing an agent, make sure you've identified a suitable use case. See [Find Your Use Case](/guides/find-use-case) for guidance.
</Info>

## Simple vs complex agents

For simple agents, you don't need much of a design process. Just create an agent and tell it what to do:

> "Do a competitor analysis every Monday morning and send me a report."

Provide some basic context, iterate on the report format, and you're done. This conversational approach works well for personal assistants and straightforward automations.

But for more complex agents—especially those that handle external communication, work with sensitive data, or fit into team workflows—a structured approach pays off. You'll save time by thinking things through before building.

## The Agent Design Canvas

We've evolved the **Agent Design Canvas** as a tool to guide the design process. It's a framework for thinking through all the critical aspects of agent design before you start building.

<Frame>
  <img alt="Agent Design Canvas showing ten key aspects: Purpose, Triggers, Action Plan, Interfaces, Capabilities, Impact, Input, Knowledge & State, Output, and Success" />
</Frame>

The canvas is perfect for both individual brainstorming and team workshops. It ensures you don't miss important considerations, everyone aligns on what the agent should do, and you spot weaknesses during design instead of after building.

<CardGroup>
  <Card title="Get the Canvas" icon="download" href="https://www.abundly.ai/agent-design-canvas">
    Download the PDF or use the interactive online version
  </Card>

  <Card title="Canvas Deep Dive" icon="newspaper" href="https://www.abundly.ai/blog/AI-agent-design-canvas">
    Read Leilei's blog post "The AI Agent Design Canvas: How We Design AI Agents That Actually Work"
  </Card>
</CardGroup>

### How to use it

The boxes are there to raise questions in your mind—you don't need to strictly fill in each one. Skip boxes that feel irrelevant. Add boxes that are missing for your context. For example, a security-sensitive agent might need a "What could go wrong?" box for risks and mitigation strategies.

Some teams keep the canvas alive as a living document that evolves with the agent. Others use it only for bootstrapping, then set it aside once the agent is running (or keep it as a record of the original design). Either approach is fine. The canvas is just a tool.

## The ten aspects of agent design

Let's walk through each section of the canvas. For each aspect, we'll explain the key questions and show an example from a **Manuscript Screening Agent** that helps a publishing team filter incoming submissions.

### Purpose

The foundation of everything. Get specific about what problem you're solving and who benefits.

**Key questions:**

* Who would benefit, and how?
* What is the precise use case? (Give examples)
* What is out of scope for this version?

<Tip>
  If you can't explain the high level purpose in two sentences, the scope may be too broad. Consider splitting into multiple agents.
</Tip>

> **Manuscript Screener:** Editorial team saves time by focusing only on high-quality submissions. The agent screens incoming manuscripts to the public submission inbox. Out of scope: filtering among the high-quality submissions (humans do that).

### Triggers

What wakes the agent up? Every agent needs an activation mechanism.

**Key questions:**

* What event starts the agent?
* Is it manual, scheduled, or event-driven?

| Trigger type | Example                                 |
| ------------ | --------------------------------------- |
| Manual       | User sends a chat message               |
| Scheduled    | Every Monday at 9am                     |
| Event-driven | Incoming email, webhook, calendar event |

> **Manuscript Screener:** Triggered by incoming email from the public manuscript submission address.

### Action plan (human & AI)

This is the heart of the design. Create numbered steps showing the workflow, clearly indicating which steps the AI performs and where humans are involved.

What are the specific steps? Who does what—agent or human? Where are the handoffs, review points, and approvals?

> **Manuscript Screener:**
>
> 1. **Agent:** Receives email with manuscript and reads the cover letter and first 5 pages
> 2. **Agent:** Reviews content quality and sets a score (yes/maybe/no)
> 3. **Agent:** Summarizes the recommendation with reasoning and emails it to the human
> 4. **Human:** Reviews the assessment and takes the best manuscripts for further review

<Note>
  The most common design oversight is not thinking through human-AI handoffs. This leads to agents that work in theory but don't fit actual workflows.
</Note>

### Interfaces

How do humans and the agent interact? The interface choice can determine whether an agent actually gets used.

How will humans and AI interact? Chat, email, Slack, shared documents, approval widgets? What are the functional requirements on the interface?

> **Manuscript Screener:** Email only. Human receives reports via email from the agent. No other interface needed.

### Capabilities

What tools and integrations does the agent need to do its job?

What integrations are needed? What other tools are required? Apply the **principle of least privilege**: give agents only the capabilities they need—no more. An agent without email access can't accidentally send a bad email.

> **Manuscript Screener:** Receive email, send email, access/read documents.

### Impact

Define success in business terms. This helps you evaluate whether the agent is worth building and provides a metric for ongoing improvement.

What is the measurable gain to the organization once the agent is working?

> **Manuscript Screener:** Screening efficiency increased by 80%, allowing the editorial team to focus on high-quality manuscripts.

### Input

What data or documents does the agent receive? This is the part of the agent's context that varies with each run.

What data or documents are provided? What format should inputs be in?

> **Manuscript Screener:** Each incoming email includes the email body, plus a manuscript and cover letter as attachments.

### Knowledge & state

What background information and external connections does the agent need?

Any external systems to connect to? What context or domain knowledge is needed? Does the agent maintain and modify any state?

> **Manuscript Screener:** Agent needs examples of what "great," "just-good-enough," and "not-really-good-enough" levels of quality look like.

### Output

What does the agent deliver? Be specific about format and destination.

What are the deliverables? What format should they be in? Consider using templates or examples.

> **Manuscript Screener:** Agent's review of incoming manuscripts is aligned with how the human user would evaluate them.

### Success

How do you know the agent is working? Define both immediate success and ongoing evaluation.

What is the definition of done? How do you know if the agent became better or worse over time?

> **Manuscript Screener:** Agent reviews 100% of incoming submissions within 24 hours and provides consistent scoring. Target pass rate: \~40%.

## Prototype early

The canvas is useful, but don't spend too long thinking before you actually try something. Build a minimum version of the agent directly in the platform—even if you have to fake things.

* **Fake data:** If you don't have access to real data yet, use fake data. You can even ask the agent to generate realistic test data for you.
* **Fake integrations:** Ask the agent to simulate connections to systems you don't have access to yet. "Pretend you received this email and tell me what you would do."
* **Quick integrations:** Use the [HTTP capability](/features/http) to test various API integrations before committing to a full setup.

The goal is to get *anything* working to understand what's possible, what's easy, and what's hard. An early prototype teaches you more than hours of theoretical design. The platform is optimized for quick starts—a useful first prototype can often be built in 30 minutes.

<Tip>
  Don't wait until you've filled out the entire canvas. Sketch out Purpose and Action Plan, build a quick prototype, learn from it, then refine your design.
</Tip>

## The agent owner

In a team context, every agent needs a clearly defined owner—think of them as the agent's "boss."

An agent may interact with many people, and those people will have feedback and ideas for improvement. But it's usually best for one person (or an owner pair) to be responsible for the agent's behavior. They tweak the instructions, they follow up when something goes wrong, they decide when to expand or change the agent's role.

Without a defined owner, you risk one of two failure modes:

* **Nobody improves the agent** — Issues pile up, the agent drifts, and eventually it stops being useful
* **Everybody improves the agent** — Like an intern with no clear boss, receiving conflicting instructions from multiple people, the poor agent becomes confused and unreliable

The owner typically drives the design process. Sometimes you'll have a separate "agent designer" who acts as a coach or consultant—we often play this role with our customers, where someone on their team owns the agent while one of our project leads guides the design process. But at the end of the day, a human must always be responsible for what an agent does. AI cannot take responsibility away from humans.

One person can own several agents, or even a whole team of agents. Think of it like an editor-in-chief at a newspaper: they can't review or micromanage everything the journalists write (that would be wildly inefficient), but they're still responsible for the end result. They create the necessary guardrails, do the necessary oversight, hire the right people, and follow up when things go wrong. The human owner of an agent plays a similar role.

## Using the canvas in practice

**For team workshops:** Get everyone in the same room, work through the canvas together, and reach alignment before building. This can take 1–4 hours depending on complexity, but saves days of misalignment later.

**For solo design:** Use it as a checklist to ensure you've thought through all aspects before implementation.

**For complex agents:** Any section of the canvas could mean significant time figuring out the right solution. That's normal—better to discover complexity during design than during building.

## Common pitfalls the canvas prevents

| Pitfall               | What goes wrong                                                | How the canvas helps                                           |
| --------------------- | -------------------------------------------------------------- | -------------------------------------------------------------- |
| **Unclear purpose**   | Everyone has a slightly different understanding                | Forces alignment on specific purpose before building           |
| **Context overload**  | Too much information makes the agent confused                  | Pushes you to think critically about what's actually necessary |
| **Undefined success** | Teams either over-engineer or ship something that doesn't work | Ensures everyone agrees on what "done" means                   |
| **Ignored handoffs**  | Agent works in theory but doesn't fit workflows                | Explicitly maps human-AI collaboration points                  |

## Keeping instructions clean

One common mistake is dumping everything you can think of into the agent's context. This creates a low signal-to-noise ratio, leading to confused agents and low-quality output. Limit context to what the agent actually needs—remove anything that's just "good to know."

During the creation process, it's OK and normal for instructions to get a bit messy as you iterate. But then clean them up: simplify, clarify, remove redundancy. The agent can help with this—ask it to review its own instructions and suggest improvements.

Don't worry about breaking things when you simplify. [Instructions are version controlled](/features/instructions)—if you ask the agent to clean up and the results aren't what you wanted, you can always roll back to a previous version.

## Iteration is essential

AI agents need human feedback to become good. They improve through iteration and guidance—more like new team members who need coaching than software that just runs.

This is why the best agent creators are often the actual users, not IT departments. Only the user knows when things change, when performance has drifted, or when new requirements emerge.

<Note>
  Getting an agent from "works pretty OK" to "this is awesome" usually takes a few rounds of refinement.
</Note>

## FAQ

<AccordionGroup>
  <Accordion title="Do I need to fill out the entire canvas?">
    Not always. For simple agents, you might only need to think through Purpose, Triggers, and Capabilities. The canvas is a tool, not a bureaucratic requirement. Use the parts that help.
  </Accordion>

  <Accordion title="How long should the design process take?">
    For a simple agent: 15–30 minutes of thinking, then iterate as you build.

    For a complex agent with team alignment needed: 1–4 hours of workshop time, followed by iterative refinement during building.
  </Accordion>

  <Accordion title="What if requirements change after I've designed the agent?">
    That's normal. Agent design is iterative. Update the canvas as you learn, and adjust the agent accordingly. The canvas is a living document, not a contract.
  </Accordion>

  <Accordion title="Should I design the agent before or after creating it in the platform?">
    For complex agents, design first—at least the Purpose, Action Plan, and Success criteria. For simple agents, you can create and design simultaneously through conversation with the agent.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Instructions" icon="scroll" href="/features/instructions">
    Configure what your agent does and how it behaves
  </Card>

  <Card title="Risk & Autonomy" icon="scale-balanced" href="/security/risk-autonomy">
    Match security configuration to your agent's risk profile
  </Card>

  <Card title="Capabilities" icon="puzzle-piece" href="/features/capabilities">
    Give agents the right tools for their job
  </Card>
</CardGroup>


# Agent Optimization
Source: https://docs.abundly.ai/guides/agent-optimization

Reduce costs and improve agent performance

Every agent consumes credits when it works. This guide covers how to understand credit usage, reduce costs, and improve agent performance through iterative refinement.

## Understanding credit usage

Credit consumption depends on what your agent does, how complex the task is, and [which LLM you use](/features/model-selection). Every case is unique, but here are examples to give you a sense of scale:

| Task                                                               | Credits | Start tier | Pro tier |
| ------------------------------------------------------------------ | ------- | ---------- | -------- |
| Simple news check of a couple of sources                           | 50      | \$0.30     | \$0.22   |
| Complex news scan (many sources, check memory, assess recency)     | 400     | \$2.30     | \$1.80   |
| Review 40-page doc against 200 pages of references                 | 300     | \$1.80     | \$1.30   |
| Create weekly menu with recipes, check nearby deals, email user    | 400     | \$2.30     | \$1.80   |
| Screen 300 acquisition targets based on industry and ownership     | 5,000   | \$29.00    | \$22.00  |
| Analyze meeting transcript + draft follow-up email and CRM summary | 175     | \$1.00     | \$0.80   |

<Note>
  Examples from February 2026. Credit consumption varies over time depending on LLM and supplier prices, and tier pricing is subject to change.
</Note>

## Reducing costs

Several factors affect how many credits your agent consumes:

### 1. Frequency of use

Running a task every few minutes versus once an hour makes a significant difference. Consider whether your agent really needs to run that often, or if a less frequent schedule would work.

### 2. Size of context

The more information your agent reads and processes, the more credits it uses. Trim context to only what's essential for the task.

In chat conversations, you can monitor this directly with the context indicator in the chat action bar (desktop). The ring fills and shifts from gray to yellow, orange, and red as context grows.

When the indicator gets high, start a new chat for a fresh context window. Keep durable reference material in [agent documents](/features/documents) so you can start fresh conversations without losing important information.

### 3. Where you store data

Store reference information in [agent documents](/features/documents) that the agent can selectively retrieve, rather than including everything in the instructions. With the default "Summary" visibility, the agent knows documents exist but only reads them when relevant—keeping context focused.

### 4. Complexity of instructions

Simplify logic and reduce unnecessary processing loops. Clear, concise instructions often perform better than lengthy, detailed ones.

### 5. Model selection

By default, agents use the latest version of Claude Sonnet—a high-performance model optimized for agentic behavior. If cost is a concern and your use case allows it, considering switching to a cheaper model. See [Model Selection](/features/model-selection) for details.

<Tip>
  Use the [Usage & Limits](/features/usage) dashboard to track credit consumption per agent. Set daily limits to prevent runaway costs.
</Tip>

<Tip>
  You can also ask the agent itself for help — for example, "How can I reduce your credit usage?" — and it will walk you through diagnostic questions and concrete techniques tailored to its setup.
</Tip>

## Learn more

<CardGroup>
  <Card title="Usage & Limits" icon="chart-line" href="/features/usage">
    Track credit consumption and set spending limits
  </Card>

  <Card title="Agent Design" icon="object-group" href="/guides/agent-design">
    A structured approach to designing effective agents
  </Card>

  <Card title="Pricing" icon="coins" href="/introduction/pricing">
    Understand credit-based pricing and subscription plans
  </Card>
</CardGroup>


# Find Your Use Case
Source: https://docs.abundly.ai/guides/find-use-case

Identify which tasks are best suited for AI agents

Before building an agent, you need to find the right use case. Not every task benefits from an agent—some are better done by code, others by humans. The goal is to find tasks in the sweet spot where agents excel.

## Code vs agent vs human

Knowledge work in organizations is traditionally carried out by either code or by humans. Agents are a new option—not quite code, not quite human, but somewhere in between.

<Frame>
  <img alt="Spectrum showing Code (fast, predictable, not intelligent, low cost), AI Agent in the middle, and Human (slow, unpredictable, intelligent, high cost)" />
</Frame>

|                    | Code                         | Agent                           | Human                       |
| ------------------ | ---------------------------- | ------------------------------- | --------------------------- |
| **Speed**          | Very fast                    | Fast                            | Slow                        |
| **Predictability** | Completely predictable       | Mostly predictable              | Unpredictable               |
| **Intelligence**   | None (follows rules exactly) | Moderate (handles fuzzy inputs) | High (creativity, judgment) |
| **Running cost**   | Very low                     | Low to moderate                 | High                        |

This isn't an "either/or" choice. Agents can use code for computations and data processing. Agents can interact with humans for approvals and escalations. The question is: **which tasks benefit from having an agent involved?**

## Step 1: Map how you spend your time

Start by analyzing how your team spends its time. What does a typical week look like? Make a list of the different types of tasks you do—each sticky note represents a task type.

<Frame>
  <img alt="Chart showing sticky notes plotted by frequency (y-axis: how often, how much time) and value (x-axis: is this good use of your time)" />
</Frame>

For each task type, assess two dimensions:

**Y-axis: Frequency and time investment**

* How often do you do this task?
* How much cumulative time does it take?
* Tasks at the top are routine, repetitive, or time-consuming

**X-axis: Value of your time**

* Does this feel like high-value work?
* Would you be sad if this was automated away?
* Tasks on the left feel like "someone should be doing this, but does it have to be me?"

The top-left quadrant—frequent tasks that don't feel like good use of your time—are your candidates for agent assistance.

## Step 2: Assess intelligence requirements

Now add another dimension: how much intelligence and creativity does each task require?

<Frame>
  <img alt="Same chart with brain icons showing intelligence requirements—larger brains for tasks requiring more creativity and judgment" />
</Frame>

In the diagram, brain size indicates intelligence/creativity requirements:

* **Large brain** — Requires significant judgment, creativity, or expertise
* **Small brain** — Requires only a bit of intelligence to handle fuzzy inputs

Tasks that require massive amounts of creativity and human judgment are better left to humans. Tasks that are completely mechanical can probably be automated with code alone. The sweet spot for agents is in between.

## The sweet spot for agents

Perfect agent tasks share these characteristics:

| Factor           | Agent sweet spot                                            |
| ---------------- | ----------------------------------------------------------- |
| **Frequency**    | Routine or repetitive—happens often enough to justify setup |
| **Time value**   | Not the best use of a human's time                          |
| **Intelligence** | Requires some judgment, but not deep expertise              |
| **Fuzziness**    | Involves uncertain inputs that code can't handle            |
| **Stakes**       | Mistakes are correctable (not mission-critical)             |

**Examples of good agent use cases:**

* Screening incoming emails and routing to the right person
* Generating weekly summaries from multiple data sources
* Reviewing documents against a checklist
* Researching competitors and compiling reports
* Handling routine customer inquiries

**Examples of poor agent use cases:**

* Highly creative work requiring original thinking
* High-stakes decisions with irreversible consequences
* Tasks that happen once a year
* Simple if-then logic that code handles perfectly

<Card title="Sample Use Cases" icon="lightbulb" href="https://www.abundly.ai/use-cases">
  Browse real examples of agents in production
</Card>

## Start simple

<Tip>
  Start with simple use cases to learn the basics, even if they aren't the highest-value opportunities. Once you're comfortable building agents, tackle more complex workflows.
</Tip>

Good starter projects:

* A daily news digest on a topic you care about
* Summarizing meeting notes and emailing action items
* Monitoring a website for changes
* Answering FAQs from a knowledge base

These build your intuition for how agents think, what context they need, and how to iterate on instructions—skills that transfer to more ambitious projects.

## From use case to design

Once you've identified a promising use case, the design process begins. This is where you work out the details: exactly what the agent should do, how it interacts with humans, what information it needs, and how you'll measure success.

For simple agents, you can skip the formal design process—just create an agent and tell it what to do. But for anything beyond a basic assistant, a structured approach pays off.

<Card title="Agent Design" icon="object-group" href="/guides/agent-design">
  Learn the structured approach to designing effective agents
</Card>

## FAQ

<AccordionGroup>
  <Accordion title="What if I can't think of any good use cases?">
    Start by tracking your work for a week. Every time you think "this is tedious" or "I wish I didn't have to do this," write it down. Patterns will emerge.

    You can also look at what other organizations automate—the [use cases page](https://www.abundly.ai/use-cases) has examples from various industries.
  </Accordion>

  <Accordion title="Should I automate my most painful task first?">
    Not necessarily. Your most painful task might also be your most complex. Start with something manageable to build your skills, then tackle the bigger challenges.
  </Accordion>

  <Accordion title="How do I know if a task is too complex for an agent?">
    If you can explain the task to a new hire in 30 minutes, an agent can probably handle it. If it requires months of training and deep expertise, you might need a human—or a human-agent collaboration where the agent does the grunt work and humans make the final calls.
  </Accordion>
</AccordionGroup>


# Troubleshooting
Source: https://docs.abundly.ai/guides/troubleshooting

Solutions for common issues

Here are answers to typical issues you might encounter when using the Abundly platform. If your issue isn't covered here, and you can't an answer by searching this site, then please contact [support@abundly.ai](mailto:support@abundly.ai). We are happy to help!

<AccordionGroup>
  <Accordion title="How do I contact support or share product input?">
    Open **Help & support** in the sidebar (above your profile), then choose **Contact support**. That opens a structured message to our support team with your verified email as the reply-to address, and it can include useful context from your session.

    For general questions, use **Documentation** or **Getting started** from the same menu, or email [support@abundly.ai](mailto:support@abundly.ai).
  </Accordion>

  <Accordion title="What does the 'Overloaded' error mean?">
    The "Overloaded" error occurs when the LLM (such as Anthropic/Claude's server) is experiencing high traffic.

    **How to resolve:**

    1. Wait 5-10 minutes and try again
    2. This is a temporary issue that usually resolves on its own
    3. If it persists after 15+ minutes, contact [support@abundly.ai](mailto:support@abundly.ai)

    You can also try switching to a different LLM temporarily. See [Model Selection](/features/model-selection) for details.
  </Accordion>

  <Accordion title="Why did my agent run end with an Anthropic quota or token limit error?">
    Anthropic enforces daily and monthly token quotas at the account level. When a run hits one of these long-term limits, Abundly ends the run immediately with a clear error message — rather than retrying for \~30 minutes to no benefit.

    This is different from:

    * **Short-term rate limits (TPM/RPM)** — these are still auto-retried transparently and usually resolve within seconds.
    * **"Overloaded" errors** — caused by provider-wide traffic spikes and also auto-retried. See the entry above.

    **How to resolve:**

    1. Wait for your quota window to reset (daily quotas reset at midnight UTC; monthly quotas reset at the start of your billing period).
    2. Temporarily switch to a different model or provider while you wait. See [Model Selection](/features/model-selection) for details.
  </Accordion>

  <Accordion title="Why isn't my integration working?">
    If an integration isn't behaving as expected, check these common issues:

    1. **Missing permissions** — Ensure you've granted all required scopes/permissions during setup. Some integrations need specific OAuth scopes to access certain features.
    2. **Not invited** — For Slack, you need to `/invite @YourAgent` to each channel where you want it to work. The agent can only see channels it's been invited to.
    3. **Webhook not configured** — For integrations that listen for events (like GitHub webhooks or Trello board events), ensure the webhook URL is set correctly in the external service.
    4. **Token expired** — Some OAuth tokens expire. For Google Drive, you can reconnect the token in place from the capability card or your Personal Tokens page — no need to reconfigure your agents. For other integrations, try disconnecting and reconnecting.
    5. **Rate limits** — External services may throttle requests. If you're making many API calls, the service might temporarily block requests.

    Each integration page has a troubleshooting section with specific guidance. See [Integrations Overview](/integrations/overview) for the full list.
  </Accordion>

  <Accordion title="Why does my verification code show 'Invalid or expired code'?">
    This usually means the code is no longer valid.

    **How verification codes work:**

    * Each code expires 30 minutes after it is sent
    * You can have multiple active codes from recent requests
    * If you request many codes in a row, older active codes may be invalidated
    * After you sign in successfully, all remaining active codes are invalidated

    **How to resolve:**

    1. Use one of your most recently received codes
    2. Confirm the code is less than 30 minutes old
    3. Enter the code carefully to avoid typos
    4. If none work, request a fresh code and use it right away
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Agent Optimization" icon="gauge-high" href="/guides/agent-optimization">
    Reduce costs and improve performance
  </Card>

  <Card title="Integrations" icon="plug" href="/integrations/overview">
    Set up and configure external services
  </Card>

  <Card title="Model Selection" icon="microchip" href="/features/model-selection">
    Choose the right AI model for your agent
  </Card>
</CardGroup>


# Anthropic
Source: https://docs.abundly.ai/integrations/anthropic

Claude family of language models

Anthropic provides the Claude family of language models, known for their strong reasoning capabilities and safety features.

## Claude models

The Claude family includes models optimized for different use cases, from fast responses to complex reasoning tasks.

Configure model selection under **Settings → Advanced Config**.

<Info>
  See [Model Selection](/features/model-selection) for the full list of available models and guidance on choosing the right one.
</Info>


# Atlassian
Source: https://docs.abundly.ai/integrations/atlassian

Work with Jira, Confluence, and other Atlassian products

Connect your agent to Atlassian to work across Jira, Confluence, and other Atlassian Cloud products — search issues, create pages, manage projects, and more.

## How to enable

Before you can connect, an Atlassian organization admin must allow Abundly to access your Atlassian instance. After that, each user authenticates with their own Atlassian account.

<Steps>
  <Step title="Allow Abundly in your Atlassian admin console">
    1. Go to [admin.atlassian.com](https://admin.atlassian.com)
    2. In the sidebar, navigate to **Apps → Platform experiences → AI settings → Rovo MCP Server**
    3. Click **Add domain** and enter your Abundly app domain with `/**` (for example, `https://app.abundly.ai/**` or `https://your-tenant.app.abundly.ai/**`)

    <Frame>
      <img alt="Atlassian admin console showing the Rovo MCP Server settings page with the Abundly domain added" />
    </Frame>

    This is a one-time setup. Once the domain is added, all users in the organization can connect their accounts.
  </Step>

  <Step title="Add Atlassian to your agent">
    Go to the **MCP Servers** tab in your agent's capabilities. You'll find Atlassian in the "Add a server" section — click it to add.
  </Step>

  <Step title="Authenticate your account">
    Click **Connect**, then **Continue with OAuth** and complete the OAuth flow to connect your Atlassian account.
  </Step>
</Steps>

<Tip>
  API token authentication is also available under advanced settings, but OAuth is recommended — API tokens only expose a limited subset of tools.
</Tip>

## Example use cases

* **Issue triage** — "Every morning, check Jira for new unassigned issues in the Support project and suggest an assignee based on workload"
* **Sprint reporting** — "Summarize the current sprint progress and post an update to Slack every Friday"
* **Documentation sync** — "When a Jira epic is marked as done, create a Confluence page documenting what was shipped"
* **Cross-platform updates** — "When someone mentions a Jira ticket in Slack, look it up and post a summary with the current status"


# Autodesk
Source: https://docs.abundly.ai/integrations/autodesk

Upload, convert, and view CAD files — and browse ACC projects — from your agent

Your agent can work with Autodesk Platform Services (APS) to upload CAD files, convert DWG drawings, translate 3D models, extract properties, generate interactive 3D viewers, and browse projects and folders in Autodesk Construction Cloud (ACC).

## How to enable

Go to **Settings → Capabilities** and enable **Autodesk**. Enter your APS app **Client ID** and **Client Secret** and pick a **region** (EU or US). File upload, conversion, translation, property extraction, and 3D viewing work as soon as app credentials are saved. To browse ACC hubs, projects, and folders you also need to complete two more steps — see [How to connect Autodesk](#how-to-connect-autodesk) below.

## Example use cases

* **Convert drawings** — "Convert this DWG to PDF so I can share it with the client"
* **Translate to viewable formats** — "Translate the uploaded RVT to SVF2 and give me a 3D viewer link"
* **Extract properties** — "List all doors in this Revit model with their sizes and fire ratings"
* **Drill into a model** — "Show me the element hierarchy of the translated model so I can find the mechanical systems"
* **Browse ACC** — "List the projects in our ACC hub and show the folder contents for Project Alpha"
* **Pull a file from ACC** — "Download the latest IFC from the Coordination folder in Project Alpha and extract its property schedule"

## What the agent can do

The tools available depend on how far you've set the capability up.

**With app credentials (Client ID + Secret):**

* Upload files to Autodesk OSS storage
* Convert DWG files to DXF, PDF, or DWF using cloud AutoCAD (Design Automation)
* Translate models (e.g. RVT → IFC, DWG → SVF2) using Model Derivative
* Extract model properties and element hierarchy from translated models
* Generate a standalone HTML 3D viewer for a translated model
* Make authenticated requests to any Autodesk API endpoint

**After connecting an Autodesk account (OAuth):**

* List your ACC hubs and projects
* Browse project folder contents
* Get item details and download URLs for ACC files

<Note>
  Files uploaded to OSS count toward a 5 GB per-file limit. Design Automation conversions typically take 1–3 minutes. Model translations can take longer depending on file size.
</Note>

## How to connect Autodesk

<Steps>
  <Step title="Create an app in Autodesk Platform Services">
    Go to [aps.autodesk.com](https://aps.autodesk.com/), sign in, and create a new app. Enable the APIs you need (at minimum **Data Management**, **Model Derivative**, and **Design Automation**; enable **ACC Data Management** if you plan to browse ACC).

    Set the **Callback URL** to:

    ```
    https://app.abundly.ai/api/auth/autodesk-callback
    ```

    Note the **Client ID** and **Client Secret** from the app's settings page.
  </Step>

  <Step title="Enter app credentials in Abundly">
    Go to **Settings → Capabilities** and enable **Autodesk**. Paste the Client ID and Client Secret, and pick the region that matches your data (EU or US). The Client Secret is stored as a workspace-level secret.

    Once saved, the agent can already upload, convert, translate, and view files.
  </Step>

  <Step title="(Optional) Add your app as an ACC Custom Integration">
    To browse ACC hubs and projects, your ACC account admin needs to approve the app:

    1. In **ACC Admin → Settings → Custom Integrations**, click **Add custom integration**.
    2. Enter the Client ID from the previous step.
    3. Grant **Full account access**.
  </Step>

  <Step title="(Optional) Connect an Autodesk account">
    Back in the capability card, click **Connect Autodesk account** and sign in with an Autodesk account that has access to the ACC projects you want the agent to see. This gives the agent user-level ACC access (hub and folder browsing, item details, download URLs).

    Tokens refresh automatically while the agent is in regular use. If the account sits idle for about 15 days, the refresh token expires and you'll need to reconnect.
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="Documents" icon="file" href="/features/documents">
    How agents handle uploaded and generated files.
  </Card>

  <Card title="Integrations overview" icon="plug" href="/integrations/overview">
    Other external services your agent can connect to.
  </Card>
</CardGroup>


# BigQuery
Source: https://docs.abundly.ai/integrations/bigquery

Ask questions about your data in plain English and get answers from Google BigQuery

Your agent can explore your BigQuery datasets and run SQL queries on your behalf — so you can get revenue breakdowns, spot trends, or run scheduled reports without writing a line of SQL. Point it at your data and ask in natural language; the agent discovers schemas, runs read-only queries, and returns results (with bytes processed so you can keep an eye on cost).

<Frame>
  <img alt="Screenshot: BigQuery capability card in Settings → Capabilities with service account JSON configured" />
</Frame>

## How to enable

Go to **Settings → Capabilities** and enable **BigQuery**. Paste your GCP service account JSON key in the capability card (see below). Credentials are stored as team secrets.

## Example use cases

* **Revenue and performance** — "Query our sales dataset in BigQuery and summarize revenue by region for last quarter"
* **Scheduled reports** — "Every Monday at 9am, run a query for last week's top 10 products and email the results to the team"
* **Data quality** — "Check the BigQuery table for duplicate order IDs and report any issues"
* **Exploration** — "List the tables in our analytics dataset and describe the schema of the events table so I know what we can analyze"
* **Cross-tool workflows** — "Pull last month's conversion numbers from BigQuery and add a summary to the weekly update in Notion"

<Frame>
  <img alt="Dr Plotts answering a question using BigQuery data" />
</Frame>

<Frame>
  <img alt="Dr Plotts chart from BigQuery data" />
</Frame>

<Note>
  The agent runs read-only SELECT queries only (30-second timeout, 1 GB billing limit per query). Write operations and DDL are not supported. Always include a LIMIT clause to keep result sets manageable.
</Note>

## How to connect BigQuery

<Steps>
  <Step title="Create a GCP service account">
    In [Google Cloud Console](https://console.cloud.google.com/), create a service account and grant it:

    * **BigQuery Metadata Viewer** — explore datasets and tables
    * **BigQuery Data Viewer** — query data
    * **BigQuery Job User** — run queries
  </Step>

  <Step title="Create a JSON key">
    In IAM & Admin → Service Accounts, select the service account and choose **Create Key → JSON**. Download the key file.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **BigQuery**. Paste the entire contents of the JSON key file into the capability card. Credentials are stored in team secrets (secret type: GCP).
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="Snowflake" icon="snowflake" href="/integrations/snowflake">
    Query and explore your Snowflake data warehouse from the agent.
  </Card>

  <Card title="Agent databases" icon="database" href="/features/agent-databases">
    Store and query structured data inside Abundly.
  </Card>
</CardGroup>


# Cloud Storage
Source: https://docs.abundly.ai/integrations/cloud-storage

Read and write files in S3-compatible object storage — AWS S3, Cloudflare R2, Google Cloud Storage, and more

Your agent can list, read, write, and delete objects in any S3-compatible bucket. Point it at AWS S3, Cloudflare R2, Google Cloud Storage (HMAC), MinIO, Backblaze B2, or any other provider that speaks the S3 API, and it can move files in and out as part of your workflows.

<Frame>
  <img alt="Screenshot: Cloud Storage capability card in Settings → Capabilities with bucket, region/endpoint, and access keys configured" />
</Frame>

## How to enable

Go to **Settings → Capabilities** and enable **Cloud storage**. Enter an AWS region (e.g. `us-east-1`) or a full `https://` endpoint URL for other S3-compatible providers, link a workspace secret with your access key ID and secret access key, and optionally set a default bucket. If no default bucket is set, the agent includes a bucket name on every call.

## Example use cases

* **Archive attachments** — "When you receive an email with a PDF attached, upload it to the `invoices/` prefix in our S3 bucket"
* **Scheduled exports** — "Every Monday, put a CSV of last week's signups into the `reports/` folder in R2"
* **Read reference data** — "Fetch `config/pricing.json` from the bucket and summarize the current tier structure"
* **Cross-bucket cleanup** — "List objects under `tmp/` older than 30 days and delete them"
* **Cross-tool workflows** — "Download the latest export from the bucket, summarize it, and post the highlights to Slack"

## What the agent can do

| Tool                | What it does                                                         |
| ------------------- | -------------------------------------------------------------------- |
| List objects        | Browse a bucket or a prefix, with pagination for large listings      |
| Get object metadata | Inspect size, content type, and headers without downloading the body |
| Download object     | Read an object's contents into the conversation                      |
| Upload object       | Write a new object or overwrite an existing one                      |
| Delete object       | Remove an object from the bucket                                     |

<Note>
  To keep responses manageable, downloads are limited to about 5 MiB per object and uploads to about 10 MiB per object. For larger files, use tools outside the agent or split the work into smaller pieces.
</Note>

## How to connect a provider

<Steps>
  <Step title="Create access keys">
    In your provider's console, create an access key ID and secret access key scoped to the buckets you want the agent to use. Grant only the permissions you need (e.g. read-only, or read-write on a specific bucket).

    * **AWS S3** — IAM user or role with an S3 policy
    * **Cloudflare R2** — R2 API Token (S3 compatibility)
    * **Google Cloud Storage** — HMAC key for a service account
    * **MinIO / Backblaze B2 / other** — provider-specific access key and secret
  </Step>

  <Step title="Note the region or endpoint">
    For AWS S3, note the region (e.g. `us-east-1`, `eu-west-1`). For other providers, copy the full `https://` endpoint URL from the provider's dashboard.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **Cloud storage**. Paste the region or endpoint URL, create a new workspace secret with the access key ID and secret access key, and optionally set a default bucket. Credentials are stored as a workspace secret.
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="Google Drive" icon="google-drive" href="/integrations/google-drive">
    Store and share files in Google Drive from the agent.
  </Card>

  <Card title="SharePoint" icon="microsoft" href="/integrations/sharepoint">
    Access Microsoft SharePoint files and libraries.
  </Card>
</CardGroup>


# ElevenLabs
Source: https://docs.abundly.ai/integrations/elevenlabs

Text to speech with natural-sounding voices

ElevenLabs provides high-quality text-to-speech capabilities for your agent.

## Generated voiceovers

Go to **Settings → Capabilities** and enable **Text to Speech**. ElevenLabs is one of the available voice providers — your agent can pick a specific ElevenLabs voice per request when generating audio files as a tool (for example, an audio summary embedded in a newsletter).

<Note>
  OpenAI's text-to-speech is also available as an alternative. Both options are included in the Text to Speech capability.
</Note>

**Example use cases:**

* "In the weekly newsletter, include an audio clip of the executive summary"
* "Generate a voiceover for this script using a male Scottish voice"
* "Every morning, generate a 'daily Mr President' style update that I can listen to while driving to work"

## Agent voice

ElevenLabs voices also show up as options for the agent's own voice — the one used by the Speak button and [Walk & Talk mode](/features/voice#walk--talk-mode). Premade, professional, and cloned voices are all available, alongside OpenAI voices, in a single dropdown under **Settings → Basic Information → Voice**.


# Figma
Source: https://docs.abundly.ai/integrations/figma

Read Figma files, render frames as images, browse projects, and post comments

Connect your agent to Figma to inspect design files, render frames as images, browse team projects, and post comments — using your Figma personal access token.

## How to enable

Go to **Settings → Capabilities** and enable **Figma**. Then create a Figma personal access token (see below) and save it as a workspace secret from the capability settings.

## Example use cases

* **Design review** — "Every morning, list any new comments on our Design System file and summarize them in Slack"
* **Render frames** — "Render the latest version of the landing page hero frame and email it to me as a PNG"
* **Spec handoff** — "Find the 'Onboarding' frame in our Product file and post a description of each layer to the engineering Slack channel"
* **Project audit** — "List every file in our Marketing team's projects and tell me which haven't been updated in the last 90 days"
* **Comment triage** — "When a Figma comment mentions @design-team, post a summary to #design-triage in Slack"

## How to create a Figma personal access token

<Steps>
  <Step title="Open token settings">
    Sign in to Figma and go to [figma.com/developers/api#access-tokens](https://www.figma.com/developers/api#access-tokens).
  </Step>

  <Step title="Generate a new token">
    Click **Generate new token**, give it a name (e.g. "Abundly Agent"), and choose an expiration.

    Select the scopes the agent needs. Common picks:

    * **file\_content:read** — read file contents and render frames
    * **file\_metadata:read** — list file versions, basic file info
    * **projects:read** — list team projects and their files
    * **file\_comments:write** — read and post comments
    * **webhooks:write** — only if you plan to register webhooks
  </Step>

  <Step title="Copy the token">
    Copy the token now — Figma shows it only once.
  </Step>

  <Step title="Save it as a workspace secret in Abundly">
    Go to **Settings → Capabilities**, enable **Figma**, and click **Add Secret**. Paste the token and give the secret a name. The token is encrypted and stored at the workspace level — agents reference it by name, never as plain text.
  </Step>
</Steps>

<Note>
  The agent can only access files and projects that the token's owner has access to in Figma. Sharing in Figma works the same way it does for the human user.
</Note>

<Tip>
  If a request returns a permission error, the token is likely missing the right scope. Recreate the token with the additional scope and update the credential.
</Tip>

## Pointing the agent at a specific team or file

Figma's API has no "list everything" endpoint — the agent always needs a team ID or file key to do anything beyond verifying the token. In a chat, just paste the relevant Figma URL.

If the agent should always work with the same team, project, or file, add it to the agent's instructions instead — for example *"You work on the design system file at [https://figma.com/file/ABC123/](https://figma.com/file/ABC123/)..."*. The agent will use it as a default and can still access other teams/files when you ask.


# Firecrawl
Source: https://docs.abundly.ai/integrations/firecrawl

Extract content from web pages

Firecrawl enables your agent to read and extract content from web pages.

## How to enable

Go to **Settings → Capabilities** to configure **Web Scraping**. This capability is enabled by default, but you can disable it if you don't want the agent to read web pages.

## Example use cases

* "Read abundly.ai/courses and find the next AI Agent course in Swedish"
* "Summarize the contents of this article: https\://....."
* "Extract pricing information from this product page, and update your pricing database: https\://....."
* "Get the content from these three URLs and highlight key similarities and differences: ..."

<Tip>
  For research across multiple sources, combine Firecrawl with the Perplexity integration. Use Perplexity to find relevant URLs, then Firecrawl to extract detailed content. See [Web Search & Research](/features/search) for more details.
</Tip>


# Giphy
Source: https://docs.abundly.ai/integrations/giphy

Search and share GIFs

Giphy enables your agent to search for and share animated GIFs.

<Frame>
  <img alt="Celebratory high five GIF" />
</Frame>

## How to enable

Go to **Settings → Capabilities** and enable **Image Generation**. Giphy is included alongside AI image generation.

## Example use cases

* "Find a celebration GIF for our team announcement"
* "Include a thumbs up GIF in the slack post"
* "In the weekly customer sentiment report, including a suitable GIF to illustrate overall sentiment"


# GitHub
Source: https://docs.abundly.ai/integrations/github

Manage repositories, pull requests, and issues

Connect your agent to GitHub to work with your code repositories — browse files, create and review pull requests, manage issues, and react to events like new issues or PRs.

## How to enable

Go to **Settings → Capabilities** and enable **GitHub**. Then enter your personal access token, and optionally specify which repositories to watch (see below for details).

## Example use cases

* **Code Review Assistant** — "Review PR 254 and flag any potential issues or security concerns"
* **Release Notes Generator** — "Summarize all commits since the last release and generate release notes"
* **Issue Triage** — "When a new issue is created, analyze it and add appropriate labels"
* **Dependency Updates** — "Check for outdated dependencies and create a PR with updates"

## How to connect your GitHub account

<Steps>
  <Step title="Create a personal access token">
    Go to [GitHub Settings → Developer settings → Personal access tokens](https://github.com/settings/tokens) and create a new token. GitHub offers two types:

    <Note>
      On GitHub Enterprise, create the token on your own instance — for example `https://yourcompany.ghe.com/settings/tokens` — not on github.com. Tokens are not valid across instances. See [GitHub Enterprise](#github-enterprise) below.
    </Note>

    **Fine-grained tokens** (recommended) — Create a [fine-grained token](https://github.com/settings/personal-access-tokens/new) with for example these permissions:

    * **Contents**: Read and write
    * **Metadata**: Read-only (required)
    * **Pull requests**: Read and write
    * **Issues**: Read and write

    **Classic tokens** — If you need simpler setup or access across multiple organizations, create a [classic token](https://github.com/settings/tokens/new) with the `repo` scope.

    See GitHub's [documentation on personal access tokens](https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/managing-your-personal-access-tokens) for more details.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **GitHub**. Paste your personal access token.

    Leave **GitHub host** empty if you use github.com. On GitHub Enterprise, enter your instance's host name — for example `yourcompany.ghe.com` or `github.yourcompany.com`. Enter the host name only, without `https://` and without a path.

    If you want the agent to react to GitHub events (via webhooks), enter the repositories to watch in the **Watched repositories** field — for example, `owner/repo` or `myorg/myrepo`. Multiple repos can be comma-separated.
  </Step>

  <Step title="Set up webhooks (optional)">
    To trigger your agent when events happen in GitHub (new issues, pull requests, etc.), configure a webhook in each repository you're watching:

    1. Go to your repository → **Settings → Webhooks → Add webhook** (on GitHub Enterprise, on your own instance)
    2. Set the Payload URL to:
       ```
       https://service.abundly.ai/api/github/webhook
       ```
       <Note>
         Enterprise customers on a dedicated deployment will have a different base URL. Replace `service.abundly.ai` with your organization's service URL, which you can find in your Abundly workspace settings or by contacting support.
       </Note>
    3. Set Content type to `application/json`
    4. Select which events should trigger the webhook (e.g., Issues, Pull requests, Pushes)
    5. Click **Add webhook**

    With webhooks configured, your agent will be notified whenever events happen in the repository. However, you also need to instruct the agent on what to do with these events — otherwise it will likely ignore them. For example, add to your agent's instructions: "When a new issue is created, analyze it and add appropriate labels."

    <Tip>
      Without webhooks, your agent can still work with GitHub — but only when you ask it to (via chat or scheduled tasks).
    </Tip>
  </Step>
</Steps>

## GitHub Enterprise

Both GitHub Enterprise products are supported. Enter your host name in the **GitHub host** field and Abundly detects which one you use — there is nothing else to configure.

| Product                              | Host looks like                                   | What to enter            |
| ------------------------------------ | ------------------------------------------------- | ------------------------ |
| GitHub.com                           | `github.com`                                      | Leave empty              |
| Enterprise Cloud with data residency | `yourcompany.ghe.com`                             | `yourcompany.ghe.com`    |
| Enterprise Server (self-hosted)      | Your own domain, such as `github.yourcompany.com` | `github.yourcompany.com` |

Two things are specific to Enterprise:

* **Create the token on your own instance.** A github.com token does not work against an Enterprise host, and the reverse is also true.
* **Use your instance for repository settings.** Watched repositories are still written as `owner/repo`, but you add webhooks in the repository settings on your own instance.

### IP allowlisting

If your enterprise restricts access by IP address — GitHub's own **IP allow list** setting on Enterprise Cloud, or your firewall in front of a self-hosted instance — allow Abundly's outbound IP address:

```
34.51.233.91
```

<Note>
  Enterprise customers connect from a different, dedicated IP address. If you're an enterprise customer and need to allowlist Abundly, [contact us](mailto:support@abundly.ai) and we'll provide your specific address.
</Note>

The IP allow list is an Enterprise Cloud feature and doesn't depend on data residency, so it can apply even when you leave the **GitHub host** field empty — an organization on Enterprise Cloud without data residency still uses github.com. GitHub.com Free, Pro, and Team accounts have no IP restrictions, so there is nothing to allowlist.

### Self-hosted reachability

These two requirements apply only to Enterprise Server. An instance that is reachable only inside your network cannot be used, because the connection has to work in both directions:

* Abundly calls your instance's API at `https://your-host/api/v3`, so the host must be reachable from the internet.
* Your instance calls Abundly at `https://service.abundly.ai/api/github/webhook` if you use webhooks, so it needs outbound access to that address.

<Note>
  Enterprise Cloud with data residency (`*.ghe.com`) is hosted by GitHub and reaches Abundly the same way github.com does, so neither reachability requirement applies to it.
</Note>


# Google Calendar
Source: https://docs.abundly.ai/integrations/google-calendar

View events, create meetings, and manage schedules

<Warning>This integration is currently hidden. Email [support@abundly.ai](mailto:support@abundly.ai) if you want to enable it.</Warning>

Connect your agent to Google Calendar to manage your schedule.

## How to enable

<Steps>
  <Step title="Enable the capability">
    Go to **Settings → Capabilities** and enable **Google Calendar**.
  </Step>

  <Step title="Connect your Google account">
    Click **Create new** to sign in with your Google account and create a personal access token. This token belongs to you and can be managed in your profile settings.

    If you already have a Calendar token from setting up another agent, you can select it from the dropdown instead.
  </Step>
</Steps>

## Access scope

After you connect a new Calendar token, Abundly takes you to a settings page where you can choose which calendars the token is allowed to access.

You can keep broad access by choosing **Allow All & Continue**, or save a restricted scope and apply least privilege from the start.

<Note>
  You can update these calendar access settings later from **Profile settings → Personal Tokens**. The agent can only list, read, create, and update events on calendars included in the allowed list.
</Note>

## Example use cases

* **Meeting Scheduler** — "Find a time that works for all participants in the email thread and schedule the meeting"
* **Daily Meeting Summary** — "Every morning, email me a summary of today's client meetings. Attach relevant information about the client from Hubspot."
* **Calendar Blocking** — "Block focus time on my calendar when I have no meetings"


# Google Chat
Source: https://docs.abundly.ai/integrations/google-chat

Connect your agent to Google Chat as its own named app, in a simple or advanced mode

Your agent can communicate with you and your team through Google Chat, appearing as its own named app. People **@mention** it in a space, or **direct-message** it, and the agent replies using its full context, capabilities, and history — just like in the portal.

Each Google Chat–enabled agent has its own Google Chat app, configured in your Google Cloud project.

<Note>
  **Abundly uses the classic HTTP Chat app, not the Google Workspace add-on.** When you configure the Chat API, you point it at Abundly's HTTP endpoint — you do **not** build it as a Workspace add-on (the "gsuiteaddons" model). The add-on model makes threaded replies harder and pulls in Marketplace publishing; the classic HTTP app sidesteps both. The steps below set up the classic app.
</Note>

## Example use cases

* "@Researcher summarise everything we know about Acme Corp" — kicks off a research agent from a space
* "@Releaser time to cut a new release" — triggers your deployment agent directly in a space
* A direct message to your assistant agent for a quick, private back-and-forth

## Modes

Google Chat has two reply modes, selectable on the capability card under **Settings → Capabilities → Google Chat**. Pick based on how much setup you want and what behaviour you need.

|                                        | **Direct reply** (simple)                            | **API** (advanced)                                |
| -------------------------------------- | ---------------------------------------------------- | ------------------------------------------------- |
| Setup                                  | HTTP endpoint only                                   | + service account key                             |
| How replies are sent                   | returned in the webhook response                     | posted via the Chat REST API                      |
| Response time                          | bounded by Google's \~30s webhook timeout            | no limit — handles long, tool-using turns         |
| Threading                              | a top-level @mention's reply starts a **new thread** | reply lands **in the @mention's thread**          |
| Agent tools (post to other spaces/DMs) | —                                                    | ✅ post / reply / direct-message tools             |
| Best for                               | quick conversational replies, easy to get going      | tool-heavy or long-running agents, tidy threading |

<Tip>
  Start with **Direct reply** to get going fast. Switch to **API** when you want threaded replies, long-running turns, or the agent to post into *other* spaces.
</Tip>

### Which Google API each mode uses

* **Direct reply (sync)** — your endpoint returns a Google Chat `Message` in the HTTP response and Google posts it for you. No outbound API call, no credentials.
* **API (async)** — the agent posts via the Chat REST API (`spaces.messages.create`, `chat.bot` scope) using a service account, with `messageReplyOption` so the reply threads correctly.
* **Not used: the Workspace add-ons API** (`hostAppDataAction` / `gsuiteaddons`). Abundly deliberately runs as a classic HTTP app, so neither mode uses the add-on response format.

## Prerequisites

* A Google Cloud project **in the organization tied to your Google Workspace** (see the warning below)
* Permission to enable APIs and create service accounts in that project
* A Google Workspace account you use for Google Chat
* The Abundly agent you want to connect

<Warning>
  **The project's org must match your Workspace.** A Chat app can only post into spaces in the Workspace whose organization owns its Cloud project. A GCP organization's *display name* is not a reliable guide — match the **Directory Customer ID** instead: get your Workspace **Customer ID** from the Google Admin console (**Account → Account settings**), then run `gcloud organizations list` and pick the org whose `DIRECTORY_CUSTOMER_ID` matches. Build the Chat app in a project under **that** org. If the project is under the wrong org, posting fails with `403 "This Chat app is not a member of this space"`.
</Warning>

## How to connect

<Steps>
  <Step title="Enable the Google Chat API">
    1. In the [Google Cloud Console](https://console.cloud.google.com), select (or create) a project **under the org tied to your Workspace** (see the warning above).
    2. Enable the **Google Chat API** for that project.
  </Step>

  <Step title="Configure the Chat app (classic HTTP)">
    Open **Google Chat API → Configuration** and fill in:

    1. **Application info** — app name (e.g. your agent's name), avatar URL, description.
    2. **Interactive features** — enabled.
    3. **Connection settings** — choose **HTTP endpoint URL** (not Apps Script, Dialogflow, or Pub/Sub) and paste the messaging endpoint Abundly shows you:
       ```
       https://your-domain/api/google-chat/bot/YOUR-AGENT-ID
       ```
       If a "Use a common HTTP endpoint URL for all triggers" toggle is shown, enable it.
    4. **Authentication Audience** — set this to **App URL**. Nothing else to configure: Abundly verifies each request against the endpoint URL it gave you.

       <Note>
         Abundly verifies every request's signed audience claim against the endpoint URL and rejects mismatches — there's no fallback to "signature only". **App URL is the only supported mode**, so leave the Authentication Audience on **App URL** (not *App project number*).
       </Note>
    5. **Visibility** — add yourself (and anyone else who should use it). Set the app status to **Live**.

    <Note>
      Do **not** enable the Google Workspace Marketplace SDK or publish this as a Workspace add-on. Abundly is a classic HTTP Chat app.
    </Note>
  </Step>

  <Step title="(API mode only) Create a service account key">
    Direct-reply mode needs nothing here — skip to the next step. For **API (async)** mode, the agent posts via the Chat REST API and needs a service account:

    1. In the **same project**, go to **IAM & Admin → Service Accounts → Create service account** (no project roles required).
    2. Open it → **Keys → Add key → Create new key → JSON** and download the file.
    3. In Abundly, create a **secret** of type **Google Chat** and paste the entire JSON key into it.

    <Note>
      The service account must live in the **same project** as the Chat app — that's how `chat.bot` authentication maps the call to your app. No Marketplace publish or scope approval is needed for `chat.bot`.
    </Note>
  </Step>

  <Step title="Enable the capability in Abundly">
    In your agent's **Settings → Capabilities**, enable **Google Chat**, then on the capability card:

    1. Choose the **reply mode** — Direct reply or API.
    2. **API mode:** select the **service account key** secret you created.
    3. The **messaging endpoint URL** shown here is the one you pasted into the Chat API Configuration.
  </Step>

  <Step title="Add the app and talk to it">
    1. In Google Chat, **@mention** the app in a space, or open a **direct message** with it.
    2. In a space, the first @mention adds the app and gets a reply.

    <Check>
      The agent replies. In a DM, keep chatting — every message reaches it. In a space, @mention it each turn (see below).
    </Check>
  </Step>
</Steps>

## How conversations work

Every Google Chat conversation maps to a persistent Abundly chat, with the same instructions, memory, and capabilities the agent has in the portal. These chats appear in the sidebar alongside portal chats and are marked as "remote".

* **Direct messages (1:1 with the app)** — every message reaches the agent, no @mention needed. A free-flowing conversation.
* **Spaces (rooms)** — Google only delivers messages to an HTTP app when it's **@mentioned**, so you @mention the agent on each turn. Each space maps to its own Abundly chat.
* **Threading** — in **API mode** the reply lands in your message's thread (tidy, contiguous). In **Direct-reply mode** a top-level @mention's reply starts a new thread, because Google ignores threading on synchronous responses.
* **Group chats** — Google Chat apps can't be added to multi-person group chats; only DMs and spaces are supported.

<Tip>
  Google Chat–originated chats are read-only in the portal — reply from Google Chat to keep the thread in the agent's working context.
</Tip>

## Agent tools (API mode)

In **API mode**, the agent gets tools to post into *other* Google Chat conversations (its reply to the current conversation is always delivered automatically — these are for reaching elsewhere):

| Tool                              | Purpose                                                              |
| --------------------------------- | -------------------------------------------------------------------- |
| `post_google_chat_message`        | Post a new message to a space (room or DM space), by `spaces/...` id |
| `reply_to_google_chat_message`    | Reply inside an existing thread, by `spaces/.../threads/...`         |
| `send_google_chat_direct_message` | DM a user by `users/...` id (resolves the existing DM space)         |

All use the `chat.bot` scope — no Marketplace or admin approval. They're available only in API mode (Direct-reply mode stays minimal). A DM is just a space, so `post_google_chat_message` with a DM space id reaches a person too; the direct-message tool is the convenience for when you have a user id but not their DM space id.

## Things to know

<Note>
  **One app per agent** — each agent uses its own Google Chat app and (for API mode) its own service account.
</Note>

<Warning>
  **Direct-reply mode has a \~30-second ceiling.** Google times out the webhook, so a long or heavily tool-using turn won't post a reply in Direct-reply mode. Use **API mode** for those.
</Warning>

## Troubleshooting

<AccordionGroup>
  <Accordion title="403 &#x22;This Chat app is not a member of this space&#x22; (API mode)">
    The app isn't a recognized member of the space. Almost always the project is under the **wrong organization** — it must be the org tied to your Workspace (match the Directory Customer ID; see the prerequisites warning). Also confirm the **service account is in the same project** as the Chat app, and that the app has been added to the space.
  </Accordion>

  <Accordion title="403 &#x22;Method doesn’t allow unregistered callers&#x22; / &#x22;use API Key&#x22;">
    The agent made a raw, unauthenticated HTTP call instead of using the dedicated Google Chat tool. Ensure you're in **API mode** with a service account key set (which loads the `post_google_chat_message` / `reply_to_google_chat_message` tools), and restart the service so the tools are loaded.
  </Accordion>

  <Accordion title="Chat briefly shows &#x22;app not responding&#x22;, then a reply appears">
    Expected behaviour pattern when something's off in API mode — the ack and the async post are out of sync. If you see it persistently, confirm you're on the latest build (the async acknowledgement returns an empty response).
  </Accordion>

  <Accordion title="No reply in a space, but DMs work">
    In spaces, Google only forwards **@mentions** to the app. Make sure you @mention it. Unmentioned follow-ups in a space thread are not delivered to a Google Chat HTTP app (unlike Microsoft Teams, which has Resource-Specific Consent for this).
  </Accordion>

  <Accordion title="Replies don't land in the thread">
    Direct-reply (sync) mode cannot thread a top-level @mention's reply — Google ignores the thread on synchronous responses. Switch to **API mode** for threaded replies.
  </Accordion>

  <Accordion title="Auth errors in the logs (token / signature)">
    With the **App URL** audience, Google sends a standard Google-issued OpenID Connect token (`iss: accounts.google.com`). Abundly verifies its signature against Google's public keys, checks the `aud` claim equals your endpoint URL, and confirms the token's `email` is `chat@system.gserviceaccount.com` (which proves the caller is Google Chat) — rejecting any mismatch.

    Two things to check: the **Authentication Audience** in the Chat API Configuration must be **App URL** (not *App project number* — Abundly only supports App URL), and the **endpoint URL you pasted must exactly match** the one Abundly gave you. If you see audience mismatches, re-copy the messaging endpoint URL and confirm the audience is set to App URL.
  </Accordion>
</AccordionGroup>


# Google Drive
Source: https://docs.abundly.ai/integrations/google-drive

Access files, create documents, and work with spreadsheets

Connect Google Drive so your agent can work with files in your Drive account. There are two capabilities you can choose from: file-scoped **Google Drive** and broader **Google Drive (Full Access)**.

## Which capability to use

| Capability                     | Access scope                                                                                                                                                 |
| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Google Drive**               | Only specific files you share with the agent, plus files the agent creates.                                                                                  |
| **Google Drive (Full Access)** | All drives and folders you approve for the token. Hidden by default — email [support@abundly.ai](mailto:support@abundly.ai) to enable it for your workspace. |

## Authentication modes

The **Google Drive** capability card offers two modes:

* **Personal** — Connect your own Google account and share specific files (the default).
* **Service account** — Reference a Google service account JSON key (stored as a workspace secret). The agent accesses exactly what's shared with the service account's email, with no personal token. Best for autonomous, per-agent access.

To set up service account mode: in Google Cloud, create a service account, enable the Drive API, download its JSON key, and store it as the secret on the capability card. Then share the files or folders the agent needs with the service account's email. The capability card includes step-by-step guidance.

<Note>
  A service account has no storage of its own, so to let the agent **create** files, give it **Content manager** access to a Shared Drive (the whole drive or a folder inside it).
</Note>

## How to enable

<Steps>
  <Step title="Enable the capability">
    Go to **Settings → Capabilities** and enable **Google Drive** or **Google Drive (Full Access)**.
  </Step>

  <Step title="Connect your Google account">
    Click **Create new** to sign in with your Google account and create a personal access token. This token belongs to you and can be managed in your profile settings.

    If you already have a matching Drive token from setting up another agent, you can select it from the dropdown instead.

    <Frame>
      <img alt="Google Drive capability configuration showing token selection" />
    </Frame>
  </Step>

  <Step title="Set access scope">
    * **Google Drive:** On the **Google Drive** capability card (under **Settings → Capabilities**), use the **Files** section to share one or more specific Drive files with your agent. You can remove access later. The agent can also access files it creates itself.
    * **Google Drive (Full Access):** After OAuth, Abundly takes you to a settings page where you can pick which drives (My Drive and any shared drives) to allow, and optionally narrow access to specific folders within each drive. Choose **Allow All & Continue** to keep broad access, or **Save & Continue** to apply a restricted scope from the start.
  </Step>
</Steps>

<Note>
  You can update Full Access drive and folder selections later from **Profile settings → Personal Tokens**.
</Note>

## What your agent can do with Drive files

* **Read:** Google Docs, Sheets, and Slides, plus text files, PDFs, and common Office files (Word, Excel, PowerPoint)
* **Visual PDF extraction:** Extract content from complex PDFs (scanned documents, tables, charts, or tricky layouts) using vision-based page extraction — no need to import the file first
* **Edit:** Google Docs and Google Sheets that the agent has access to
* **Create:** New Google Docs and Google Sheets in your Drive

## If your token expires or is revoked

OAuth tokens refresh automatically in normal use. If your Google Drive access has been revoked or can no longer be refreshed, use **Reconnect** where available (the Drive capability card or **Personal Tokens**) to re-run consent on the same token so agents keep using the existing token reference.

## Example use cases

* **Document creation** — "Create a summary doc in Google Drive from this meeting transcript"
* **Data analysis** — "Read the data from this spreadsheet and give me the key insights"
* **Data entry** — "Remove the toasters product from the pricing spreadsheet"
* **Data updates** — "Update the quarterly report in Drive with the latest figures that John just posted in Slack"


# Google Gemini
Source: https://docs.abundly.ai/integrations/google-gemini

Gemini family of language models and image generation

Google provides the Gemini family of language models and image generation capabilities for your agent.

## Gemini models

Gemini models are available as an alternative to the default model for your agent.

Configure model selection under **Settings → Advanced Config**.

<Info>
  See [Model Selection](/features/model-selection) for the full list of available models and guidance on choosing the right one.
</Info>

## Image generation

Generate images from text descriptions using Google's `gemini-3.1-flash-image` model.

**How to enable:** Toggle on the **Image generation** capability on your agent, and select Google.

<Frame>
  <img alt="Example image generated by Gemini" />
</Frame>

<Note>
  New agents default to GPT for image generation, but Gemini remains fully supported and you can select it from the Image generation capability card.
</Note>

## Example use cases

* "Generate a nice looking infographic summarizing these meeting notes: ..."
* "In your daily news summary, also generate and include a suitable cover image"


# HubSpot
Source: https://docs.abundly.ai/integrations/hubspot

Manage contacts, companies, deals, and more

Connect your agent to HubSpot to read and update any data in your CRM — contacts, companies, deals, tickets, tasks, emails, custom objects, and more. The agent can call any HubSpot REST API endpoint, so anything you can do in HubSpot, the agent can do too.

## How to enable

Go to **Settings → Capabilities** and enable **HubSpot**. Then provide your HubSpot private app access token.

## Example use cases

* **Contact lookup** — "Find all contacts at Acme Corp and summarize our recent interactions"
* **Deal monitoring** — "Every morning, check for deals stuck in the same stage for over a week"
* **Company research** — "Look up this company and give me a summary of their contact history"
* **Task creation** — "Create a follow-up task for this deal"
* **Custom objects and properties** — "List all tickets assigned to me with priority HIGH, and update their status to In Progress"

## How to create a HubSpot access token

HubSpot uses private apps to provide API access. Each private app generates an access token with specific permissions.

<Steps>
  <Step title="Open Private Apps settings">
    In your HubSpot account, click the **Settings** icon (gear) in the top navigation bar. Then go to **Integrations → Private Apps**.
  </Step>

  <Step title="Create a private app">
    Click **Create a private app**. Enter a name (e.g., "Abundly Agent") and optionally a description.
  </Step>

  <Step title="Configure scopes">
    Go to the **Scopes** tab and select the permissions your agent needs. For typical CRM access, enable:

    * **crm.objects.contacts** — Read (and Write if the agent should create or update contacts)
    * **crm.objects.companies** — Read (and Write if needed)
    * **crm.objects.deals** — Read (and Write if needed)
    * **crm.objects.owners** — Read
    * **sales-email-read** — Read (for email logging)

    Add any other scopes your agent needs — tickets, custom objects, engagements, etc. The agent can call any HubSpot endpoint your token has access to, so enable scopes that match the tasks you want it to perform.
  </Step>

  <Step title="Create and copy the token">
    Click **Create app** in the top right. Once created, click **Show token** to reveal your access token, then copy it.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **HubSpot**. Paste your private app access token.
  </Step>
</Steps>

<Note>
  You must be a super admin in HubSpot to create private apps. See [HubSpot's private apps documentation](https://developers.hubspot.com/docs/api/private-apps) for more details.
</Note>


# MCP Servers
Source: https://docs.abundly.ai/integrations/mcp

Connect external tools via the Model Context Protocol

MCP (Model Context Protocol) is an open protocol that standardizes how AI applications connect to external tools. You can give your agent access to an MCP server by simply pasting the server URL into the MCP Servers tab.

Benefits:

* **Extend beyond built-in integrations** — Add capabilities that aren't provided by Abundly yet
* **Use community servers** — Tap into the growing ecosystem of MCP servers
* **Connect internal systems** — Build an MCP server for your in-house systems (CRM, inventory, custom database, etc.) and make them available to your agents
* **No waiting** — Add new integrations yourself, without waiting for platform updates

<Info>
  Learn more about the Model Context Protocol (MCP) in the <a href="https://modelcontextprotocol.io">official MCP documentation</a>.
</Info>

<Frame>
  <img alt="MCP Servers tab showing connected servers" />
</Frame>

## How to add an MCP server

You can add an MCP server from the built-in library or by entering a URL manually.

### From the library

Go to the **MCP Servers** tab in your agent's capabilities. The "Add a server" section shows popular MCP servers you can add with one click — Atlassian, Stripe, Sentry, Linear, and more. Click any of them to add it to your agent, then authenticate if needed.

### By URL

<Steps>
  <Step title="Find an MCP server">
    Get the URL of an MCP server — from the community, a vendor, or one you've built yourself.
  </Step>

  <Step title="Add to your agent">
    Click **Connect by URL** in the MCP Servers tab and paste the server URL.
  </Step>

  <Step title="Authenticate if needed">
    The platform automatically detects if authentication is required. Choose OAuth (recommended) or API token.

    If the server requires non-standard OAuth settings, expand the **Advanced** section in the connection dialog. There you can set a **custom token endpoint URL** (for servers like NetSuite where the token endpoint lives on a different host), specify an explicit **scope string**, and enter your own **Client ID and Client Secret**.
  </Step>

  <Step title="Enable/disable tools">
    Once connected, the platform discovers available tools. Toggle individual tools on or off.
  </Step>
</Steps>

Your MCP server will now be listed among your agent's capabilities, and the tools it exposes will be available to your agent. Try it in the chat!

## Security

**Be careful with third-party MCP servers.** Only add servers from sources you trust. An MCP server has access to perform actions on your behalf, so verify the source before connecting — especially for servers that aren't from official vendors or well-known open source projects.

### Workspace control

Workspace admins can allow or disallow MCP servers for the whole workspace — or per team — using the **Allow MCP Servers** setting under capability settings. When disallowed, users can't add MCP servers, and servers already added to agents are disabled: their tools become unavailable to the agent until MCP servers are allowed again.

<Info>
  MCP credentials are encrypted using hybrid AES+RSA encryption. OAuth tokens are never exposed to the AI model directly — the platform handles authentication.
</Info>

## Exposing your agent as an MCP server

You can also go the other way — expose your Abundly agent as an MCP server so external AI apps (Claude Desktop, Cursor, n8n, etc.) can call it as a tool. See [API access](/features/agent-api-endpoint) for details.


# Meta Ads
Source: https://docs.abundly.ai/integrations/meta-ads

Read campaign, ad set, and ad performance data from Facebook and Instagram ads

Connect your agent to Meta's Marketing API to analyse paid Facebook and Instagram advertising — ad accounts, campaigns, ad sets, ads, creatives, and insights. The agent can call any Marketing API endpoint, so anything you can pull from Ads Manager, the agent can pull too.

## How to enable

Go to **Settings → Capabilities** and enable **Meta Ads**, then link a workspace secret containing your Meta **App ID**, **App Secret**, and **system user access token**. The same secret can be reused across agents and teams. Abundly refreshes 60-day tokens automatically and saves the new token back to your workspace secret so it survives platform restarts.

## Example use cases

* **Daily performance summary** — "Every weekday at 8am, post yesterday's spend, impressions, clicks, and CTR for each active campaign to the #marketing Slack channel"
* **Anomaly watch** — "If any active ad set's CPA jumps more than 50% week-over-week, email me the breakdown by ad and creative"
* **Weekly review** — "Every Monday, generate a PDF report of last week's performance broken down by campaign objective and country"
* **Creative comparison** — "Compare the top 5 ads by ROAS across all active campaigns this month and summarise the patterns"
* **Account hygiene** — "List ad sets that have spent more than \$500 with fewer than 100 link clicks and flag them for review"

## How to create a Meta connection

The Marketing API authenticates with a **system user access token**. Store it together with your Meta app's **App ID** and **App Secret** so Abundly can refresh the token before it expires (Meta's recommended 60-day system user tokens).

<Steps>
  <Step title="Open Meta Business Settings">
    Go to [business.facebook.com/settings](https://business.facebook.com/settings) and select the Business that owns the ad accounts you want the agent to access.
  </Step>

  <Step title="Create a system user">
    Under **Users → System Users**, click **Add**. Give it a name (e.g. "Abundly Agent") and choose the **Employee** role. Confirm to create it.
  </Step>

  <Step title="Assign ad accounts">
    With the new system user selected, click **Assign Assets** and pick the ad accounts the agent should access. Grant **View performance** for read-only access, or **Manage campaigns** if the agent should also create or edit campaigns. Save.
  </Step>

  <Step title="Create or locate your Meta app">
    The token must be tied to a Meta app. If your Business doesn't already have one, create a free app at [developers.facebook.com/apps](https://developers.facebook.com/apps) (type **Business**) and associate it with your Business.

    From the app's **Settings → Basic**, copy the **App ID** and **App Secret**. No public review is required for first-party usage — the system user token only works against ad accounts you've explicitly assigned.
  </Step>

  <Step title="Generate the system user token">
    Back in Business Settings, with the system user selected, click **Generate New Token**. Pick the app from step 4, then choose the scopes:

    * **`ads_read`** — required for reading campaigns, ad sets, ads, and insights
    * **`ads_management`** — only needed if the agent should create or modify campaigns
    * **`business_management`** — only needed if the agent should manage Business assets

    Meta may require a 60-day expiring token — that is fine; Abundly refreshes it automatically. Copy the token immediately — Meta will not show it again.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities**, enable **Meta Ads**, and create a workspace secret of type **Meta Ads** with three fields:

    * **App ID** — from your Meta app
    * **App Secret** — from your Meta app
    * **System user access token** — from step 5

    Select that secret on the capability card and save. Abundly refreshes the token before it expires and updates the **System user access token** field in your secret automatically — you only paste a new token manually if refresh fails (e.g. the token was fully expired before the next refresh).
  </Step>
</Steps>

<Note>
  You need to be an admin of the Business in Meta Business Settings to create system users and generate tokens. If a token is fully expired and refresh fails, generate a new system user token in Meta and update the **System user access token** field in your workspace secret. See [Meta's system user documentation](https://developers.facebook.com/docs/marketing-api/system-users) for more details.
</Note>

## API version

The agent uses the current default Marketing API version automatically. If you need to test against a different version (e.g. when Meta releases a new one), just tell the agent — it can target any version your Meta app supports.

## What this covers

This integration is scoped to **paid advertising** on Facebook and Instagram via the Marketing API. It does not include organic page posts, Instagram content, Messenger conversations, or WhatsApp messaging — those use different parts of Meta's Graph API and aren't enabled by this capability.


# MongoDB
Source: https://docs.abundly.ai/integrations/mongodb

Let your agent read and write your MongoDB data using your own database connection

Your agent can run commands against your MongoDB database — finding documents, running aggregations, inserting records, and updating data — all in plain English. Point it at your database with a connection string and ask in natural language; the agent translates your request into MongoDB commands and returns the results.

Unlike read-only data integrations, what the agent can do is governed entirely by the database user in your connection string. Connect with a read-only user and the agent can only read; connect with a read-write user and it can also insert, update, and delete. You stay in control by choosing the permissions of the user you connect with.

## How to enable

Go to **Settings → Capabilities** and enable **MongoDB**, then link a workspace secret containing your **connection string**. The same secret can be reused across agents and teams.

## What you need

All the agent needs is a standard MongoDB **connection string** (URI), for example:

```
mongodb+srv://user:password@cluster.mongodb.net
```

The connection string is stored as an encrypted workspace secret. It includes the database user's credentials, so the user's permissions determine exactly what the agent is allowed to do.

<Tip>
  Create a dedicated database user for the agent with only the permissions it needs. If the agent should only answer questions about your data, give it a read-only user. If it should also maintain records, give it read-write access to the relevant database.
</Tip>

## Default database

The capability card has an optional **Default Database** field. This is the database the agent uses when a request doesn't name one explicitly.

* **If set** — the agent works against that database by default, so you can just ask "how many orders do we have?" without naming the database every time.
* **If left empty** — the agent must specify a database on every command. Set a default if your agent mostly works with a single database.

The agent can always target a different database when you ask it to, regardless of the default.

## Which MongoDB can I use?

Any MongoDB database that accepts a standard connection string works — self-hosted, MongoDB Atlas, or any managed provider.

If you don't have a database yet, you can start for free with [MongoDB Atlas](https://www.mongodb.com/atlas), MongoDB's hosted service, which offers a free tier that's plenty for getting started. You're not locked in — the same connection string approach works with any MongoDB instance, so you can move to a paid tier or your own infrastructure later.

## IP allowlisting

If your database restricts access by IP address (for example, MongoDB Atlas network access rules), allow Abundly's outbound IP address:

```
34.51.233.91
```

<Note>
  Enterprise customers connect from a different, dedicated IP address. If you're an enterprise customer and need to allowlist Abundly, [contact us](mailto:support@abundly.ai) and we'll provide your specific address.
</Note>

## Example use cases

* **Ask questions about your data** — "How many active users signed up in the last 30 days? Use the analytics database in MongoDB"
* **Reporting** — "Every Monday, count last week's orders by status and post a summary to Slack"
* **Maintaining records** — "Add this support ticket to the tickets collection" (with a read-write user)
* **Exploration** — "List the collections in our app database and show me a sample document from the users collection"
* **Data checks** — "Find any orders in MongoDB with a missing customer ID and flag them"

## How to connect MongoDB

<Steps>
  <Step title="Create a database user">
    In your MongoDB provider (e.g. Atlas) or your self-hosted instance, create a database user for the agent. Grant only the permissions it needs — read-only if the agent should only query, read-write if it should also modify data.
  </Step>

  <Step title="Allowlist Abundly's IP (if required)">
    If your database restricts access by IP, add Abundly's outbound IP `34.51.233.91` to the allowlist. Enterprise customers should contact Abundly for their dedicated address.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities**, enable **MongoDB**, and create a new workspace secret of type **MongoDB** containing your connection string. Optionally set a **Default Database** so the agent doesn't need it specified on every request.
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="Agent databases" icon="database" href="/features/agent-databases">
    Store and query structured data inside Abundly — no external database needed.
  </Card>

  <Card title="BigQuery" icon="database" href="/integrations/bigquery">
    Query and explore your Google BigQuery data from the agent.
  </Card>

  <Card title="Snowflake" icon="snowflake" href="/integrations/snowflake">
    Query and explore your Snowflake data warehouse from the agent.
  </Card>
</CardGroup>


# Microsoft Teams
Source: https://docs.abundly.ai/integrations/msteams

Connect your agent to Microsoft Teams so it appears as its own named bot

Your agent can communicate with you and your team through Microsoft Teams, appearing as its own named bot. In channels, users @mention the bot to start a conversation; follow-up replies in the same thread keep flowing without needing to @mention again. In DMs and group chats every message reaches the bot directly.

## Example use cases

* "@Releaser time to cut a new release" — triggers your deployment agent directly in a channel
* "@Researcher summarise everything we know about Acme Corp" — kicks off a research agent from Teams
* "@Standup post a daily standup summary to the Engineering channel every morning"

## Prerequisites

* Microsoft Teams with Developer Portal access
* Microsoft Entra admin center access (admin role, for the client secret and API permissions)
* Microsoft Teams Admin Center access (Teams administrator role)
* The Abundly agent you want to connect

## How to connect

The setup takes about 15–20 minutes across the Developer Portal, Microsoft Entra admin center, and Teams Admin Center.

<Steps>
  <Step title="Create the bot and a client secret">
    1. In Microsoft Teams, open the **Developer Portal** app (search for it in Apps)
    2. In the top bar, go to **Tools → Bots**, click **+ New Bot**, give it the same name as your agent (e.g. "Releaser"), and click **Add** — you're taken straight to the bot's **Configure** page
    3. Set the **Endpoint address** to the messaging endpoint URL shown in Abundly (format: `https://your-domain/api/ms-teams/bot/YOUR-AGENT-ID`) and click **Save**
    4. Go to **Client secrets** (in the bot's left-hand menu), click **Add a client secret for your bot**, and confirm
    5. **Copy the secret value immediately** — it will not be shown again. You'll paste this into Abundly's **Bot client secret** field

    <Note>
      Creating the bot here automatically creates a matching app registration in Microsoft Entra; the client secret you just made is stored on that registration. You'll grant Graph API permissions on the same registration in the next step, where you'll also pick up the Bot ID and Tenant ID with one-click copy buttons.
    </Note>
  </Step>

  <Step title="Grant Graph permissions and grab IDs in Microsoft Entra">
    1. Go to [entra.microsoft.com](https://entra.microsoft.com) and navigate to **Entra ID → App registrations**
    2. Find the app registration that was automatically created — it will have the same name as your bot — and click it to open the **Overview** page
    3. From the Overview page, copy:

       * **Application (client) ID** — this is your **Bot ID** (paste into Abundly's **Bot ID** field)
       * **Directory (tenant) ID** — this is your **Tenant ID** (paste into Abundly's **Tenant ID** field)

       Both fields have one-click copy buttons next to them — much easier than hunting these down elsewhere.
    4. Go to **API permissions → Add a permission → Microsoft Graph → Application permissions** and add:
       * `Team.ReadBasic.All`
       * `Channel.ReadBasic.All`
       * `ChannelMessage.Read.All`
       * `Chat.Read.All`
       * `User.Read.All`
    5. Click **Grant admin consent** and confirm

    <Note>
      Microsoft Entra ID is the new name for Azure Active Directory. If you prefer the Azure Portal, the same settings live under **Microsoft Entra ID → App registrations** at [portal.azure.com](https://portal.azure.com).
    </Note>
  </Step>

  <Step title="Enter credentials in Abundly">
    In your agent's settings, open the **Microsoft Teams** integration panel and enter:

    | Field                 | Where to find it                    |
    | --------------------- | ----------------------------------- |
    | **Bot ID**            | Application (client) ID from step 2 |
    | **Bot client secret** | Client secret from step 1           |
    | **Tenant ID**         | Directory (tenant) ID from step 2   |

    Click **Save**.
  </Step>

  <Step title="Create and publish the Teams app">
    1. Back in the Developer Portal, go to **Apps → New app** and give it any name (the manifest you paste below will overwrite name, description, bot, and scopes anyway). You'll land on the new app's **Basic information** page; copy the **App ID** shown at the top — you'll need it in step 3.
    2. Go to **Branding** and upload icons if desired (or skip — Teams shows a generic placeholder). Icons are binary uploads, not part of the JSON manifest, so this is the one thing the manifest paste can't set for you.
    3. Go to **App package editor** and replace the manifest with one of:
       * The **pre-filled manifest** in Abundly's Microsoft Teams integration panel (after credentials are saved, scroll to *App manifest* → *Copy manifest*) — this is already populated with your agent name, description, and Bot ID. **Before pasting, replace `REPLACE-WITH-TEAMS-APP-ID-FROM-DEV-PORTAL` with the App ID you copied in step 1.** Teams locks the app `id` at creation, so the paste fails if the manifest's `id` doesn't match.
       * The reference template below, substituting the placeholder values yourself:

    ```json Manifest template theme={null}
    {
      "$schema": "https://developer.microsoft.com/en-us/json-schemas/teams/v1.16/MicrosoftTeams.schema.json",
      "version": "1.0.0",
      "manifestVersion": "1.16",
      "id": "YOUR-TEAMS-APP-GUID",
      "name": {
        "short": "YOUR-AGENT-NAME",
        "full": "YOUR-AGENT-NAME"
      },
      "developer": {
        "name": "YOUR-COMPANY-NAME",
        "websiteUrl": "https://your-company.com",
        "privacyUrl": "https://your-company.com/privacy",
        "termsOfUseUrl": "https://your-company.com/terms"
      },
      "description": {
        "short": "Abundly agent bot",
        "full": "Abundly agent bot for Microsoft Teams"
      },
      "icons": {
        "outline": "outline.png",
        "color": "color.png"
      },
      "accentColor": "#ffffff",
      "bots": [
        {
          "botId": "YOUR-BOT-ID-GUID",
          "scopes": ["personal", "team", "groupChat"],
          "isNotificationOnly": false,
          "supportsFiles": false
        }
      ],
      "webApplicationInfo": {
        "id": "YOUR-BOT-ID-GUID",
        "resource": "api://botid-YOUR-BOT-ID-GUID"
      },
      "authorization": {
        "permissions": {
          "resourceSpecific": [
            { "type": "Application", "name": "ChannelMessage.Read.Group" }
          ]
        }
      },
      "validDomains": []
    }
    ```

    The three GUIDs play **two distinct roles**:

    * `id` is the **Teams app's** own identity in your org's app catalog. It is auto-generated when you create the app shell (step 1) and **locked** thereafter — pasting a manifest with a different `id` triggers a "Property 'id' cannot be changed in the manifest" error. Use the App ID you copied in step 1.
    * `bots[0].botId` and `webApplicationInfo.id` must both equal your **Application (client) ID** from the Entra app registration overview (step 2). Teams ties RSC consent to `webApplicationInfo.id`, so it has to match the registration that owns the bot.

    <Note>
      **What `ChannelMessage.Read.Group` does.** This is a [Resource-Specific Consent (RSC)](https://learn.microsoft.com/en-us/microsoftteams/platform/graph-api/rsc/resource-specific-consent) permission, granted by a team owner at install time. Without it, Teams only delivers @mentions to the bot — so a user who replies in a thread without re-mentioning the bot would never reach the agent. With it, every channel message reaches our webhook, and Abundly only acts on @mentions or messages that continue an existing thread the user already opened with the bot.
    </Note>

    <Note>
      We currently recommend `manifestVersion: "1.16"`. Versions 1.25 and higher may require additional properties that are outside the scope of this guide. RSC requires `1.12` or higher.
    </Note>

    4. Click **Save** in the App package editor, then click **Update app package** when prompted (or via the banner that says "*Your changes have been saved locally, but not yet updated in the app package.*"). Without this step the app still uses the old manifest.
    5. Go to **Publish → Publish to org** and click **Get Started** (the publish action)

    The status should show **Published to org** or **Submitted (awaiting admin approval)**.

    <Note>
      The Developer Portal also offers **Publish → App validation**, but that's a preflight aimed at apps being submitted to AppSource (the public Microsoft Teams Store). For internal org publishing it's unnecessary — and it tends to surface scary "Failure" labels that are actually validator-side flake or store-only naming rules ("Short name should not contain 'Microsoft'", etc.). Skip it.
    </Note>

    <Note>
      **Icon requirements are strict.** Teams' validator rejects anything that doesn't match exactly:

      * **Color icon**: 192×192 PNG, full color
      * **Outline icon**: 32×32 PNG with **only pure white (`#FFFFFF`) and fully transparent pixels** — no anti-aliasing, no near-white. If you see "Outline icon is not transparent. It's Alpha,R,G,B: …", your export has soft edges; re-export with no smoothing.
    </Note>
  </Step>

  <Step title="Approve the app in Teams Admin Center">
    1. Go to [admin.teams.microsoft.com](https://admin.teams.microsoft.com)
    2. Navigate to **Teams apps → Manage apps** and search for your app name
    3. If the status shows **Blocked** (or pending approval), click on the app and click **Publish** to make it available org-wide

    <Note>
      Admin approval changes can take up to 30 minutes to propagate across the Microsoft Teams client.
    </Note>
  </Step>

  <Step title="Install the bot in a team">
    1. In Microsoft Teams, click **Apps** in the left rail
    2. Look under **Built for your org** (or **Added by your org**) and find your bot
    3. Click the bot, choose **Add to a team**, pick the team and channel, and confirm
    4. The team owner must accept the **ChannelMessage.Read.Group** permission when prompted — this is what lets the bot see follow-up replies in a thread without re-mention

    <Note>
      Allow up to 30 minutes after publishing for the app to appear under **Built for your org**. Try a hard refresh (or sign out / sign in) if it's slow to surface.
    </Note>

    <Note>
      Teams Admin Center's **Manage apps → Install** action only manages org-level availability — it does not actually install the bot into a team. Use the in-client **Apps → Built for your org** flow above to install into a team.
    </Note>

    <Check>
      The bot is now live. Users can @mention it in a channel to start a conversation; subsequent replies in that thread continue without needing to @mention again.
    </Check>
  </Step>
</Steps>

## How conversations work

Microsoft Teams is a full remote-chat surface — on par with Slack. Every Teams conversation maps to a persistent Abundly chat, giving the agent the same instructions, memory, and capabilities as it has in the portal.

* **Channel @mentions** start a new chat — the agent replies in the thread with full context.
* **Follow-up replies** in a thread the agent already joined continue the same chat without needing to @mention again.
* **Group chats** map to persistent chats — every message reaches the agent directly.
* **1:1 personal chats (DMs)** also map to persistent chats — the agent replies naturally.

These conversations appear in the sidebar alongside portal chats and can be filtered using the chat filter pill (e.g. *Hide remote chats*). Handoffs between Teams and the portal are seamless.

<Tip>
  Teams-originated chats are marked as "remote" and are read-only in the portal — reply from Teams to keep the thread in the agent's working context.
</Tip>

## Updating an existing install

Whenever you change the manifest (bumped `version`, added permissions, etc.), Teams will not automatically apply the change to teams that already have the bot installed — and the team owner is not re-prompted to consent to any new permissions. To apply the update:

1. Bump the manifest `version` in the Developer Portal (e.g. `1.0.0` → `1.0.1`) and re-publish the app to your org
2. In each team that has the bot installed: **Manage team → Apps → trash icon next to the bot → Remove**
3. Re-add the bot following step 6 above

Removing and re-adding (rather than the in-place "Update" option, which often does not re-trigger the consent dialog) is the most reliable way to ensure new permissions like `ChannelMessage.Read.Group` are actually granted.

## Things to know

<Note>
  **One app per agent** — each agent requires its own Microsoft Teams app and bot. There is no way to have multiple bot identities within a single Teams app.
</Note>

<Warning>
  **Secret expiry** — client secrets expire. Note the expiry date and rotate the secret in the Developer Portal (**Tools → Bots → your bot → Client secrets**) or Microsoft Entra (**your bot's app registration → Certificates & secrets**) and update it in Abundly before it expires.
</Warning>

## Troubleshooting

<AccordionGroup>
  <Accordion title="Bot isn't responding to @mentions">
    * **Add the bot to the channel** — registering the bot in Entra is not enough; it must be installed into the Teams roster (step 6)
    * **Check the endpoint URL** — verify the messaging endpoint in the Developer Portal matches the URL shown in Abundly (`https://your-domain/api/ms-teams/bot/YOUR-AGENT-ID`)
    * **Verify credentials** — confirm the Bot ID, Bot client secret, and Tenant ID are entered correctly in Abundly
  </Accordion>

  <Accordion title="Bot replies to the first @mention but ignores follow-up messages in the thread">
    This is the symptom of `ChannelMessage.Read.Group` not being granted in the team where the bot is installed. The Bot Framework webhook only fires for @mentions in channels by default — the RSC permission is what unlocks delivery for in-thread replies.

    Verify and re-grant:

    1. In the team: **Manage team → Apps → click the bot row**. Look for **ChannelMessage.Read.Group (Application, Resource specific consent)** under Permissions.
    2. If it's missing, the consent never took. The most reliable fix is to bump the manifest `version`, re-publish, then **remove the bot from the team and re-add it** (see [Updating an existing install](#updating-an-existing-install) above). The in-place "Update" button often does not re-trigger the consent dialog.
    3. RSC `ChannelMessage.Read.Group` only delivers in **standard** channels. Private and shared channels use separate RSC permissions and do not deliver in-thread follow-ups via this permission.
    4. The team must have **owner** consent — only team owners see the consent prompt during install, and only owners can grant RSC.
    5. If everything looks granted but messages still don't arrive, give it 10–30 minutes — Microsoft has documented occasional propagation delays for RSC delivery.
  </Accordion>

  <Accordion title="&#x22;BotNotInConversationRoster&#x22; error">
    The bot has not been installed into the team. Complete step 6 to add it to the roster.
  </Accordion>

  <Accordion title="Can't find the bot under 'Built for your org'">
    * **Wait** — propagation from publish to visibility in the Teams client typically takes 5–30 minutes
    * Confirm the app is approved in the Admin Center (step 5) and its status is **Unblocked / Allowed**, not **Blocked**
    * Try a hard refresh in the Teams web client, or sign out/sign in on desktop
    * Try the Teams web client at [teams.microsoft.com](https://teams.microsoft.com) — the desktop client caches aggressively
  </Accordion>

  <Accordion title="&#x22;Permissions needed&#x22; warning when adding the bot">
    Go to Microsoft Teams Admin Center → **Teams apps → Permission policies → Global (Org-wide default)** and ensure custom apps are set to **Allow**. If you've just made this change, try a hard refresh in your browser before retrying.
  </Accordion>

  <Accordion title="Manifest validation errors">
    * Ensure `manifestVersion` is `"1.16"` — do not use 1.25 or higher
    * Ensure all fields in `developer` and `description` are filled in — empty `{}` objects will fail validation
    * Ensure the `id` field is a plain GUID with no extra characters or braces
    * **`Property 'id' cannot be changed in the manifest`** — Teams locks the Teams app `id` at app creation. The pasted manifest's `id` must equal the App ID shown on your app's **Basic information** page in the Developer Portal. The pre-filled manifest from Abundly leaves a `REPLACE-WITH-TEAMS-APP-ID-FROM-DEV-PORTAL` placeholder for exactly this reason — replace it with the real App ID before pasting.
    * **`webApplicationInfo.appId` required** error: this means RSC permissions are declared but `webApplicationInfo` is missing or its `id` is empty. Set `webApplicationInfo.id` to your **Bot ID** (same GUID as `bots[0].botId`).
  </Accordion>
</AccordionGroup>


# Notion
Source: https://docs.abundly.ai/integrations/notion

Read and write pages, manage databases, and search content

Connect your agent to Notion to read pages, query databases, and create or update content in your workspace.

## How to enable

Go to **Settings → Capabilities** and enable **Notion**. Then provide your Notion internal integration token.

## Example use cases

* **Database Updates** — "When I forward you an email from a client, add a new entry to our CRM database in Notion"
* **Content Search** — "Search Notion for all pages mentioning the Q4 roadmap"
* **Meeting Notes** — "Create meeting notes from our #standup channel in slack and link them to the project page in Notion"
* **Status Tracking** — "Every morning, check the sprint database and summarize tasks that are overdue"
* **File handoff** — "When someone uploads a file in Slack, save it as an agent document and attach it to the matching Notion project page"

## How to create a Notion integration token

Notion uses internal integrations to provide API access. Each integration generates a token with access to specific pages you share with it.

<Steps>
  <Step title="Create an integration">
    Go to [notion.so/my-integrations](https://www.notion.so/my-integrations) and click **+ New integration**.

    Enter a name (e.g., "Abundly Agent") and select the workspace you want to connect.
  </Step>

  <Step title="Configure capabilities">
    On the **Capabilities** tab, select the permissions your agent needs:

    * **Read content** — required
    * **Update content** — if your agent needs to edit pages
    * **Insert content** — if your agent needs to create new content

    Click **Save changes**.
  </Step>

  <Step title="Copy the token">
    Go to the **Secrets** tab and copy the **Internal Integration Token**.
  </Step>

  <Step title="Share pages with the integration">
    The integration can only access pages you explicitly share with it. For each page or database you want the agent to access:

    1. Open the page in Notion
    2. Click **•••** in the top right → **Connect to**
    3. Select your integration from the list

    Child pages inherit access, so you only need to share the top-level page.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **Notion**. Paste your internal integration token.
  </Step>
</Steps>

<Note>
  The agent can only access pages you explicitly share with the integration. It cannot browse or search your entire workspace.
</Note>

## Working with files in Notion

Your agent can upload files from [agent documents](/features/documents) into Notion and attach them to pages. This includes files that originally came from Slack attachments or chat uploads, as long as they exist as agent documents.

Common patterns:

* Attach an uploaded image or PDF to an existing Notion page
* Include files as part of an automated "Slack to Notion" workflow
* Create a database item and include uploaded files in the page content


# OpenAI
Source: https://docs.abundly.ai/integrations/openai

GPT models, image generation, voice, and vision capabilities

OpenAI provides a wide range of AI capabilities for your agent, from language models to image generation and voice features.

## GPT models

OpenAI's GPT models are available as an alternative to the default model for your agent. Configure model selection under **Settings → Advanced Config**.

<Info>
  See [Model Selection](/features/model-selection) for the full list of available models and guidance on choosing the right one.
</Info>

## Image generation

GPT image generation is the default provider for new agents.

Generate images from text descriptions using OpenAI's GPT image generation.

**How to enable:** Toggle on the **Image generation** capability on your agent, and select GPT.

<Frame>
  <img alt="Image generation capability card" />
</Frame>

**Quality levels:** You can ask the agent to generate images at `low`, `medium`, or `high` quality. Medium is the default and works well for most use cases. Low is faster and cheaper, and fine for drafts and simple scenes. High is best for complex scenes with many details or precise text and typography.

**Source images:** You can provide up to 16 source images as references for editing, compositing, or style transfer.

**Example use cases:**

* "Generate a hero image for our blog post about sustainable energy"
* "Create an illustration showing our product workflow"
* "Create a high-quality poster with the headline 'Summer Sale — 30% off'"
* "Combine these product photos into a lifestyle scene"

## Voice transcription

Transcribe spoken audio to text using OpenAI's Whisper model.

**How to use:** Press the mic button in the chat, or drop an audio file to transcribe.

**Example use cases:**

* Talk instead of typing to your agent
* Drop a meeting recording file and ask the agent to transcribe it

<Info>
  See [Voice Communication](/features/voice) for details on voice input and walk-and-talk mode.
</Info>

## Text to speech

Convert text to natural-sounding speech using OpenAI's TTS models.

**How to use:** Press the speak button under any chat message to hear it read aloud.

**Example use cases:**

* Listen to a long response while making coffee or taking a walk
* Have your agent read out a summary or report

<Info>
  See [Voice Communication](/features/voice) for details on text to speech and walk-and-talk mode.
</Info>

## Realtime voice API

The OpenAI Realtime API enables low-latency, speech-to-speech interactions with AI models. The Abundly platform uses this in combination with [Twilio](/integrations/twilio) to handle voice calls — your agent can make and receive phone calls with natural conversation flow.

**How to enable:**

* Enable **Make Phone Call** and/or **Receive Phone Call** in **Settings → Capabilities**.
* NOTE: These capabilities are currently hidden. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable them for your account.

<Info>
  See [Voice Communication](/features/voice) for details on voice calls and other voice features.
</Info>

## Image recognition

Your agent can analyze and understand images automatically.

**How to use:** Provide an image through the chat interface, agent documents, or email attachments.

**Example use cases:**

* "Look at this screenshot and tell me what's wrong with the layout"
* "Read this invoice image and extract the key details"
* "Read this chart image and add the data to my spreadsheet"


# Outlook
Source: https://docs.abundly.ai/integrations/outlook

Email management for Microsoft 365

Connect your agent to Outlook to read and search your Microsoft 365 emails and create drafts.

## How to enable

<Steps>
  <Step title="Enable the capability">
    Go to **Settings → Capabilities** and enable **Outlook**.
  </Step>

  <Step title="Connect your Microsoft account">
    Click **Create new** to sign in with your Microsoft account and create a personal access token. This token belongs to you and can be managed in your profile settings.

    If you already have an Outlook token from setting up another agent, you can select it from the dropdown instead.
  </Step>
</Steps>

<Note>
  In many organizations, a Microsoft admin must approve the Abundly Outlook app before the connection can be completed. If you see an "Almost There" message after sign-in, ask your admin to grant approval in Microsoft Entra, then try connecting again.
</Note>

## Sending behavior

By default, Outlook gives your agent read and draft tools only.

In capability settings, you can enable **Allow sending email** to let the agent send immediately from your Outlook mailbox.

<Warning>
  Sending is a sensitive action. When **Allow sending email** is on, your agent can send messages directly instead of only drafting them.
</Warning>

## Technical details

<Info>
  This section describes exactly what the integration requests and how access is scoped. Useful when a Microsoft admin needs to review the app before granting consent.
</Info>

### What the agent connects to

Outlook runs against **Microsoft Graph** (`graph.microsoft.com/v1.0`) using delegated OAuth 2.0 through the Microsoft identity platform. The integration is registered as a multitenant app against the `organizations` endpoint, so work and school accounts are supported and personal Microsoft accounts are not.

Everything the agent does goes through the `/me` endpoints of the connected account. The agent cannot reach another user's mailbox, a shared mailbox it wasn't granted, or any mailbox outside the connected account.

### Permissions requested

| Scope                        | Why it's requested                                                               |
| ---------------------------- | -------------------------------------------------------------------------------- |
| `openid`, `profile`, `email` | Sign-in and identifying which account was connected                              |
| `offline_access`             | A refresh token, so the agent keeps working without re-authorizing               |
| `User.Read`                  | Reads the account's email address to label the connection                        |
| `Mail.Read`                  | Lists and reads messages                                                         |
| `Mail.ReadWrite`             | Creates drafts                                                                   |
| `Mail.Send`                  | **Only requested when Allow sending email is enabled** at the moment you connect |

All permissions are **delegated** — the agent acts as the connected user and can never do more in Outlook than that user can.

### Tools exposed to the agent

| Tool                    | What it does                                                                          | Available                               |
| ----------------------- | ------------------------------------------------------------------------------------- | --------------------------------------- |
| `list_outlook_messages` | Lists messages in a folder (Inbox by default), max 50 per call, optional search query | Always                                  |
| `get_outlook_message`   | Reads one message by ID, including the full body                                      | Always                                  |
| `create_outlook_draft`  | Creates a draft — never sends it                                                      | Always                                  |
| `send_outlook_mail`     | Sends immediately from the mailbox and saves to Sent Items                            | Only when **Allow sending email** is on |

### Sending is gated twice

`send_outlook_mail` is off by default and requires two independent things to be true:

1. **Allow sending email** is enabled in the capability settings.
2. The connected token actually holds `Mail.Send` — which it only does if the toggle was already on when you authorized.

Turning the toggle on for an existing connection is therefore not enough on its own. The capability card detects a token without send permission and prompts you to reconnect, which triggers a fresh consent screen listing the send permission.

### Tokens and access

* The connection is stored as a **personal token** that belongs to you, not to the workspace or the agent. Manage it under **Profile settings → Personal Tokens**, where you can see which agents use it and disconnect it.
* Access tokens are refreshed automatically shortly before they expire, using the refresh token from `offline_access`. See [Credentials](/security/credentials) for how Abundly handles credentials generally.
* The agent never sees the token. Authentication is attached when the platform calls Microsoft Graph, after the model has decided what to do.
* Every Outlook tool call is recorded in the agent's [activity log](/features/monitoring).

<Warning>
  An agent using your token acts as **you** in Outlook, and anyone who can use that agent can use its Outlook access. Only connect your mailbox to agents whose access list you're comfortable with, and leave sending off unless the agent genuinely needs it.
</Warning>

### Admin-side controls

Your Microsoft admins keep full control independently of Abundly:

* **Admin consent** — many tenants require an admin to approve "Abundly Outlook Integration" in Microsoft Entra before any user can connect.
* **Conditional Access** — your existing policies apply to the delegated token.
* **Revocation** — removing the enterprise application in Entra, or revoking the user's sessions, cuts off every agent using it immediately.

## Example use cases

* **Email Summary** — "Summarize all emails from my manager this week"
* **Attachment Search** — "Find emails with invoice attachments from the last month and add them to my Notion tracker"
* **Daily Digest** — "Every morning at 8am, check my inbox for urgent emails and post a summary to Slack"
* **SMS Follow-up** — "When a team member sends an SMS to you about a deal, find their recent emails and draft a follow-up"


# Integrations Overview
Source: https://docs.abundly.ai/integrations/overview

Give your agent access to external services and data

You can give your agents access to external services such as Slack, Google Drive, GitHub, etc. These are toggled on and off in the Capabilities section.

<Frame>
  <img alt="Agent capabilities and integrations" />
</Frame>

## Authentication

Most external services require authentication.

* Some integrations use **platform-provided credentials** — for example, Perplexity, and send/receive email. These work out of the box with no setup required, since they use a shared account managed by Abundly.
* Other integrations require **your own credentials** — for example, Slack or Notion. These access your personal accounts or customer-specific data, so you must authorize the connection.

When you toggle on a capability that requires authentication, the capability card will display an OAuth button or an API token field for completing the authentication process. See each integration's page for specifics.

<Info>
  Credentials are stored securely and encrypted. The agent only has access to what you explicitly authorize. See [Credentials](/security/credentials) for details.
</Info>

## On-demand capabilities

You don't need to configure all capabilities upfront. If you ask an agent to do something that requires a capability it doesn't have, it will automatically request it by showing a capability card in the conversation.

<Frame>
  <img alt="Agent requesting a capability with a capability card" />
</Frame>

Toggle the capability on to enable it, complete any required authentication, and the agent will continue with the task.

## User approval

Some capabilities can be configured to require user approval before the agent executes them. For example, you might want to review outgoing emails before they're sent.

When approval is required, the agent will pause and wait for you to approve or reject the action before proceeding. See [User Approval](/security/user-approval) for details on how this works.

## Model selection

The LLM (Large Language Model) is the "brain" of your agent — it powers the agent's reasoning, decision-making, and natural language understanding.

Agents come preconfigured with a good default model, so you normally don't need to think about this.
However, you can choose a different model under **Settings → Advanced Config**. Available models include Claude models (from Anthropic), Gemini models (from Google), and GPT models (from OpenAI).

<Info>
  See [Model Selection](/features/model-selection) for more info about model switching, and guidance on choosing the right model for your agent.
</Info>

## Extending integrations

If the integration you need is not available, you can extend your agent's capabilities by connecting to an [MCP server](/integrations/mcp) or using the [HTTP capability](/features/http).

<CardGroup>
  <Card title="MCP Servers" icon="plug-circle-plus" href="/integrations/mcp">
    Connect any MCP server to extend your agent's capabilities beyond built-in integrations.
  </Card>

  <Card title="HTTP Capability" icon="globe" href="/features/http">
    Connect to any REST API directly — a swiss army knife for custom integrations.
  </Card>
</CardGroup>

## FAQ

<AccordionGroup>
  <Accordion title="Is Abundly integrated with Microsoft ecosystem?">
    Yes. Abundly integrates with [Microsoft Teams](/integrations/msteams) (full remote-chat surface), [Outlook](/integrations/outlook) (email management), and [SharePoint](/integrations/sharepoint) (document management). All three are available to all customers — just enable them in **Settings → Capabilities**.

    We don't use Microsoft's selection of language models—we make independent choices about which AI models to use based on performance for autonomous agent tasks.
  </Accordion>
</AccordionGroup>


# Perplexity
Source: https://docs.abundly.ai/integrations/perplexity

AI-powered web search and deep research

Perplexity powers your agent's ability to search the web and perform in-depth research. This is a platform-provided integration — no setup required.

## How to enable

Go to **Settings → Capabilities** to configure **Web Search** and **Deep Research** — both powered by Perplexity. Web Search is enabled by default.

## Web Search vs Deep Research

**Web Search** returns answers in seconds — ideal for quick facts, news, and lookups. **Deep Research** takes a lot longer but provides comprehensive analysis with extensive citations. See [Web Search & Research](/features/search) for a detailed comparison.

## Example use cases

* **Competitive Analysis** — "Do deep research on the competitive landscape for AI-powered customer support tools, and how it relates to our product"
* **News Monitoring** — "Every morning, find the latest news about our industry and post a summary to Slack"
* **Due Diligence** — "When you receive an email about a new lead, research their company and reply with a brief summary"
* **Market Research** — "What are the current trends in sustainable packaging? Include academic sources."

<Tip>
  The agent includes citation URLs in all responses. For deep research, be specific about what you're looking for — the more context you provide, the better the results.
</Tip>


# Integration pricing
Source: https://docs.abundly.ai/integrations/pricing

How integrations are billed

When your agent uses an external service, who pays for it? The answer depends on whose credentials are being used.

* **User-provided credentials** — If you bring your own account (e.g., Slack, Notion, GitHub), then Abundly doesn't charge for these integrations since you're already paying the provider directly.
* **Platform-provided credentials** — If Abundly provides the account (e.g., web search, image generation), then Abundly pays the provider and converts the cost to credits, which are charged from your account.

So credits are charged whenever your agent uses an external service with platform-provided credentials.

## Example

Suppose your agent has the following capabilities:

<Frame>
  <img alt="Three capability cards showing different authentication types" />
</Frame>

* **Slack** asks you to provide credentials — your account is used, so no credits charged.
* **Stripe** uses OAuth to connect your account — your account is used, so no credits charged.
* **Web Search** works out of the box with no setup. But web search is still an external service (Perplexity) — since no user credentials are provided, Abundly's account is used, so credits are charged on usage.

<Info>
  LLM inference — the "thinking" your agent does — also consumes credits, regardless of which integrations are involved. See [Pricing](/introduction/pricing) for the full picture.
</Info>

<Tip>
  For concrete examples of how many credits different tasks consume, see the [Agent Optimization](/guides/agent-optimization) guide.
</Tip>

## Tracking credit usage

To see how credits are being consumed:

1. Click the Abundly logo in the top left
2. Select **Usage and Limits**
3. View the breakdown by agent or time period

<Tip>
  All Abundly accounts are workspaces, so credits are pooled across your entire workspace. Every agent and member draws from the same balance, making it easy to manage costs in one place. See [Pricing](/introduction/pricing) for more details.
</Tip>

<Tip>
  You can set per-agent limits to prevent any single agent from consuming too many credits. See [Usage & Limits](/features/usage) for details.
</Tip>


# Salesforce
Source: https://docs.abundly.ai/integrations/salesforce

Manage contacts, accounts, opportunities, leads, and more

Connect your agent to Salesforce to read and update any data in your CRM — contacts, accounts, opportunities, leads, cases, tasks, custom objects, and more. The agent can call any Salesforce REST API endpoint, including SOQL queries, so anything you can do in Salesforce, the agent can do too.

## How to enable

Go to **Settings → Capabilities** and enable **Salesforce**, then link a workspace secret containing your Salesforce login URL, Consumer Key, and Consumer Secret. The same secret can be reused across agents and teams.

## Example use cases

* **Lead triage** — "Every morning, list new leads from the last 24 hours and email me a summary grouped by lead source"
* **Opportunity monitoring** — "Check for opportunities closing this month that haven't had activity in 14 days, and create a follow-up task on each"
* **Account research** — "Look up this account and summarize open opportunities, recent cases, and the primary contact"
* **Data hygiene** — "Find duplicate contacts (same email) and report them to me"
* **Custom objects** — "Read this week's records from our custom Project\_\_c object and post a summary to the #sales Slack channel"

## How to set up Salesforce authentication

Salesforce uses **External Client Apps** (the modern replacement for Connected Apps) with the OAuth 2.0 Client Credentials Flow for server-to-server access. The agent authenticates as a Salesforce user you choose ("run-as user"), so everything the agent does is performed and audited as that user.

<Steps>
  <Step title="Find your My Domain login URL">
    In Salesforce **Setup**, search for **My Domain**. The **Current My Domain URL** at the top is what you'll use as the **Login URL** in Abundly.

    <Warning>
      Salesforce gives each org three different URL aliases. Only the canonical **`https://<your-org>.my.salesforce.com`** version works for OAuth. The other two — `…my.salesforce-setup.com` (Setup pages) and `…lightning.force.com` (Lightning UI) — will not authenticate. Take the URL straight from the My Domain page in Setup to be safe.
    </Warning>
  </Step>

  <Step title="Create an External Client App">
    In **Setup**, search for **External Client App Manager** and click **New External Client App**.

    In **Basic Information**:

    * **External Client App Name**: e.g. "Abundly Agent"
    * **API Name**: auto-fills
    * **Contact Email**: your email
    * Leave **Distribution State** as **Local**
  </Step>

  <Step title="Enable OAuth and add scopes">
    Check **Enable OAuth**, then in **App Settings**:

    * **Callback URL**: Enter `https://login.salesforce.com/services/oauth2/success` (a placeholder — Client Credentials Flow doesn't use it, but the form requires a value).
    * **Selected OAuth Scopes**: Move at least **Manage user data via APIs (api)** to the right. For broader access during evaluation, also add **Full access (full)**. You can tighten this later.
  </Step>

  <Step title="Enable Client Credentials Flow">
    Scroll down to **Flow Enablement** and check **Enable Client Credentials Flow**.

    Then in **Security**:

    * Leave **Require secret for Web Server Flow** and **Require secret for Refresh Token Flow** checked (the defaults).
    * **Uncheck** **Require Proof Key for Code Exchange (PKCE) extension**. PKCE doesn't apply to Client Credentials Flow and breaks token issuance if left enabled.

    Click **Create** to save the app.
  </Step>

  <Step title="Set the Run As user">
    On the new app's detail page, open the **Policies** tab and click **Edit**. Scroll to **OAuth Flows and External Client App Enhancements**.

    In the **Run As (Username)** field, enter the Salesforce **username** of the user the agent should act as (the exact username is shown under **Setup → Users**, and isn't always the same as the email).

    <Warning>
      The agent will see and modify data exactly as the run-as user can. We recommend a dedicated integration user with the **Minimum Access – API Only Integrations** profile and only the permission sets the agent needs.
    </Warning>

    Save the policy.
  </Step>

  <Step title="Wait for activation">
    External Client Apps take **about 10 minutes** after creation before they can issue tokens. If you try to authenticate immediately and get an `invalid_client_id` error, wait and retry.
  </Step>

  <Step title="Copy the Consumer Key and Secret">
    On the app's detail page, open the **Settings** tab and click **Consumer Key and Secret**. Salesforce will send a verification code to your email — enter it. Copy the **Consumer Key** and **Consumer Secret**.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities**, enable **Salesforce**, and create a new workspace secret of type **Salesforce**. Paste:

    * **Login URL** — your `https://<your-org>.my.salesforce.com` from step 1. For sandboxes, use the sandbox's My Domain URL or `https://test.salesforce.com`.
    * **Consumer Key**
    * **Consumer Secret**

    All three values live in a single workspace secret, so you only have to enter them once and can share the same secret across multiple agents.
  </Step>
</Steps>

<Note>
  If your org is older and still allows Connected Apps, that path also works (Connected Apps remain functional even though new ones can no longer be created in fresh orgs). The OAuth flow is identical — only the Setup screens differ.

  See [Salesforce's Client Credentials Flow documentation](https://help.salesforce.com/s/articleView?id=xcloud.remoteaccess_oauth_client_credentials_flow.htm) for more details.
</Note>

## API version

The agent calls the Salesforce REST API at `https://{your-instance}/services/data/v66.0/...`. If you need a different API version, just tell the agent — it can target any version your org supports.

## Downloading files

The agent can download files (PDFs, attachments) from Salesforce using `salesforce_api_request` with `saveResponseAsDocument: true`. The response is saved as an agent document for transcription or PDF reading.

For PDFs and attachments, prefer the Connect API endpoint:

`GET /services/data/v66.0/connect/files/{ContentDocumentId}/content`

Some endpoints (such as ContentVersion `VersionData`) return valid file bytes but omit `Content-Type`. If needed, the agent can set `responseDocumentName` (e.g. `contract.pdf`) or `responseDocumentMimeType` (`application/pdf`) when saving.


# SharePoint
Source: https://docs.abundly.ai/integrations/sharepoint

Access documents and work with team sites

Connect your agent to SharePoint to read, create, edit, and fully manage files and folders across your Microsoft 365 document libraries.

## What your agent can do

* **Browse** — List files and folders in sites and libraries
* **Read** — Fetch file content (including many Office formats via extraction)
* **Create and edit text files** — Create or update `.txt`, `.md`, and `.csv` files
* **Organize** — Move or rename files and folders, copy items (cross-site capable), create folders, and delete items to the site recycle bin (recoverable)

<Note>
  Deleting through Abundly sends items to the SharePoint **site recycle bin**, not permanent removal. Your Microsoft 365 admins or site owners can restore them from there if needed.
</Note>

## How to enable

You can connect SharePoint in two ways. Pick the one that fits the agent.

| Mode                   | Acts as                        | Best for                                                                       | Setup effort                  |
| ---------------------- | ------------------------------ | ------------------------------------------------------------------------------ | ----------------------------- |
| **Personal** (default) | The connected user             | Interactive agents where per-user attribution matters                          | Click-through OAuth           |
| **App-only**           | The agent's own Entra identity | Autonomous agents that should not be kicked out by Conditional Access timeouts | One-time admin setup in Entra |

Open the capability card on the agent and use the **Personal / App-only** picker at the top to switch between them.

### Personal mode

<Steps>
  <Step title="Enable the capability">
    Go to **Settings → Capabilities** and enable **SharePoint**, then pick **Personal** at the top of the card.
  </Step>

  <Step title="Connect your Microsoft account">
    Click **Create new** to sign in with your Microsoft account and create a personal access token. This token belongs to you and can be managed in your profile settings.

    If you already have a SharePoint token from setting up another agent, you can select it from the dropdown instead.
  </Step>
</Steps>

<Note>
  In many organizations, a Microsoft admin must approve the Abundly SharePoint app before the connection can be completed. If you see an "Almost There" message after sign-in, ask your admin to grant approval in Microsoft Entra, then try connecting again.
</Note>

### App-only mode

App-only mode runs SharePoint as the agent's own Entra identity using the Microsoft client\_credentials flow with `Sites.Selected` permission. The agent gets access only to the sites an admin explicitly grants it — no broad tenant-wide read, no per-user token expiry.

<Steps>
  <Step title="Register an app in Microsoft Entra">
    In your customer's tenant, go to **Microsoft Entra → App registrations → New registration**. Give it a name (e.g. "Abundly agent"). Accounts in this directory only is fine. No redirect URI needed.

    Copy the **Directory (tenant) ID** and **Application (client) ID** — you'll paste them into Abundly in the last step.
  </Step>

  <Step title="Grant Sites.Selected">
    On the registration, go to **API permissions → Add a permission → Microsoft Graph → Application permissions → Sites.Selected**. Click **Grant admin consent**.
  </Step>

  <Step title="Create a client secret or upload a certificate">
    Under **Certificates & secrets**, create either:

    * **A client secret** — copy the **Value** immediately (it's only shown once). Store it in Abundly as a workspace secret of type **SharePoint app-only — Client secret**.
    * **A certificate** — upload the public X.509 cert; you'll need the **SHA-1 thumbprint** and the matching **PKCS#8 PEM private key** in Abundly. Store the PEM as a workspace secret of type **SharePoint app-only — Certificate** (separate type with a textarea so the multiline PEM is preserved exactly — pasting a PEM into a single-line input strips the newlines and Microsoft rejects it with a decoder error).

    Note the expiry date — you'll want to renew before it hits.
  </Step>

  <Step title="Grant per-site access">
    `Sites.Selected` starts the app with zero site access. A SharePoint or Global Administrator must grant the app access to each specific site you want the agent to reach.

    This is done via PowerShell — either **PnP.PowerShell** (`Grant-PnPAzureADAppSitePermission`) or **Microsoft Graph PowerShell** (`POST /sites/{site-id}/permissions`). Use whichever tool fits your environment.

    You'll need:

    * The app's **Client ID**
    * The target **site URL**
    * The desired **permission level** (`Read`, `Write`, or `FullControl`)

    Repeat for every site the agent should reach.
  </Step>

  <Step title="Configure Abundly">
    In the SharePoint capability card, switch to **App-only**. Paste the tenant ID, client ID, and the secret (or certificate + thumbprint). Save, then click **Test connection** to confirm Microsoft accepts the credential.

    Optionally:

    * **Default site URL** — used by tools when no `siteUrl` is supplied per call.
    * **Granted sites** — one SharePoint site URL per line, listing the sites you granted the app in the previous step. Microsoft doesn't expose an API for the app to discover which sites it has been granted, so paste the list here. The agent sees it in its prompt as the authoritative starting set, and **Test connection** verifies each entry by calling Microsoft Graph for it.
  </Step>
</Steps>

<Note>
  **Behaviour differences in app-only mode**

  Tenant-wide site enumeration isn't supported with `Sites.Selected` — every tool call has to target a specific site (`siteUrl`), drive (`driveId`), or full SharePoint URL (`webUrl`). The agent can still work across multiple sites in a single task; tell it which ones (via **Granted sites**, the agent's instructions, the user's prompt, or **Default site URL**) and it will iterate, calling tools once per site. If the agent tries a site you haven't granted, Microsoft returns 403 and the tool surfaces that.

  If the secret expires or Microsoft rejects the credential, Abundly marks the integration as failing and shows a Reconnect banner on the capability card. The next successful Test connection or token acquisition clears it.
</Note>

<Tip>
  App-only access is per-agent and independent of any user. Anyone with permission to use the agent can use the agent's SharePoint access — apply the principle of least privilege at site-grant time and via the agent's access controls.
</Tip>

## Access scope

After you connect a new SharePoint token, Abundly takes you to a settings page where you can choose which sites to allow and optionally narrow access to specific folders.

You can keep broad access by choosing **Allow All & Continue**, or save a restricted scope and apply least privilege from the start.

<Note>
  You can update these SharePoint access settings later from **Profile settings → Personal Tokens**.
</Note>

<Note>
  Text-based files (txt, md, csv) can be created and edited in place. Office documents can be read, organized (moved, copied, renamed, deleted), and put into folders, but their contents cannot be edited directly.
</Note>

## Example use cases

* **Document Search** — "Find all documents in our Legal site that were modified this week"
* **Content Retrieval** — "Get the contents of the employee handbook and summarize it"
* **Report Generation** — "Every Friday, create a weekly summary report in the Team site based on our Slack activity"
* **Cross-platform Sync** — "Check my Outlook inbox every morning — if you see any invoice emails, save them to the Finance folder in SharePoint"


# Slack
Source: https://docs.abundly.ai/integrations/slack

Chat with your agent in Slack, or drive it from Slack events

Your agent can communicate with you and your team in Slack. People mention it in a channel or DM it, and the agent replies in the same thread — using its full context, capabilities, and history — just like it does in the portal chat. It can also search, read, post messages, add files, and react with emoji.

Each Slack-enabled agent has its own identity in Slack, with its own Slack app, bot user, and token.

## How to enable

Go to **Settings → Capabilities** and enable **Slack**. Then pick a Slack mode, paste in your bot token and Slack app ID, and point your Slack app's Event Subscriptions URL at the webhook Abundly shows you. See the step-by-step guide below.

## Slack modes

The Slack capability has two modes, selectable on the capability card. You can switch modes at any time under **Settings → Capabilities → Slack**.

### Chat mode (default)

The natural choice for most agents. Slack threads and DMs behave as persistent agent chats — the same abstraction as the portal chat. People start a conversation by @mentioning the agent in a channel or sending it a DM, and then keep replying in the same thread without having to @mention every time.

Chat mode already supports everything most agents need on Slack, including:

* Posting proactively to any channel the bot has been invited to
* Sending DMs
* Adding reactions
* Running all Slack tools (search, user lookup, etc.)

You do **not** need Trigger mode just to post to a channel, react, or run Slack tools — Chat mode handles all of that.

### Trigger mode

Each Slack event your app subscribes to is forwarded to the agent as a raw trigger payload for it to interpret with your instructions. There is no automatic reply; the agent only responds or takes action if your instructions call for it.

Trigger mode is intended for ambient monitoring or custom automations on specific event types (for example: "whenever someone reacts with 🚨 in #alerts, file an incident"). It is **not** the right mode for normal interactive Slack use.

<Warning>
  Trigger mode can produce a high volume of events. If multiple Trigger-mode agents share a channel they can end up reacting to each other's messages and spiral into an infinite reply loop, which consumes credits quickly.

  Before using Trigger mode:

  * Narrow the `bot_events` in your Slack app to only the specific event types you care about, rather than using the full recommended manifest.
  * Avoid putting multiple Trigger-mode agents in the same channel unless you've designed them not to react to each other.
  * Consider whether Chat mode already covers your use case — it almost always does.
</Warning>

## Example use cases

* "Post a daily standup summary to #engineering every morning"
* "Whenever a new potential customer is mentioned on the #sales channel, research how the customer fits our ideal profile, and add relevant info to the slack thread"
* "Notify #alerts when you are unable to solve a support ticket"
* "When a support request needs follow-up, add an :eyes: reaction so the team sees it's being handled"
* "When a teammate shares a design mockup in a Slack thread, summarize the feedback and save the file as an agent document for follow-up"

## How to connect to your Slack workspace

<Steps>
  <Step title="Create a Slack app from the manifest">
    Abundly generates an app manifest tailored to your agent. It includes the webhook URL, OAuth scopes, and the event subscriptions Slack needs for threads, DMs, and group DMs.

    1. In Abundly, go to **Settings → Capabilities**, enable **Slack**, and copy the **App Manifest**.
    2. Visit [api.slack.com/apps](https://api.slack.com/apps) and click **Create New App → From a manifest**.
    3. Select your workspace, paste the manifest, then click **Next** and **Create**.

    The manifest looks like this (your agent's name and webhook URL are filled in automatically):

    ```json Manifest template theme={null}
    {
      "display_information": {
        "name": "YOUR_AGENT_NAME",
        "description": "YOUR_AGENT_DESCRIPTION",
        "background_color": "#24a18d"
      },
      "features": {
        "app_home": {
          "messages_tab_enabled": true,
          "messages_tab_read_only_enabled": false
        },
        "bot_user": {
          "display_name": "YOUR_AGENT_NAME",
          "always_online": true
        }
      },
      "oauth_config": {
        "scopes": {
          "bot": [
            "channels:read",
            "chat:write",
            "groups:read",
            "im:read",
            "users:read",
            "channels:history",
            "groups:history",
            "im:history",
            "channels:join",
            "im:write",
            "mpim:history",
            "reactions:write",
            "reactions:read",
            "files:read",
            "files:write"
          ]
        }
      },
      "settings": {
        "event_subscriptions": {
          "request_url": "YOUR_WEBHOOK_URL",
          "bot_events": [
            "message.channels",
            "message.groups",
            "message.im",
            "message.mpim"
          ]
        },
        "org_deploy_enabled": false,
        "socket_mode_enabled": false,
        "token_rotation_enabled": false
      }
    }
    ```
  </Step>

  <Step title="Install the app and copy your credentials">
    After creating the app:

    1. Copy the **App ID** from the Basic Information page.
    2. On the same page, under **App Credentials**, click **Show** next to **Signing Secret** and copy it.
    3. Go to **Install App** in the left sidebar and click **Install to Workspace**, then authorize the app.
    4. Copy the **Bot User OAuth Token** from the **OAuth & Permissions** page.
  </Step>

  <Step title="Configure in Abundly">
    Back in **Settings → Capabilities → Slack**, paste the bot token, app ID, and signing secret, then save. Click **Test connection** to confirm Abundly can reach your workspace.

    The signing secret lets Abundly confirm that incoming events genuinely come from Slack. Without it, anyone who knows your agent's webhook URL could send it fabricated messages, so we strongly recommend setting it.
  </Step>

  <Step title="Verify the Event Subscriptions URL in Slack">
    Slack does **not** verify the Request URL when you paste the manifest — it just saves it. You have to trigger the verification yourself, once, before any Slack events will reach your agent.

    1. Go back to your app at [api.slack.com/apps](https://api.slack.com/apps) and open **Event Subscriptions** in the left sidebar.
    2. The Request URL is already filled in from the manifest, but you'll see a banner saying the URL isn't verified. Click **Click here to verify** (or **Retry**).
    3. Slack sends a test request to Abundly and the banner should change to **Verified** ✓.
    4. Click **Save Changes** at the bottom of the page.

    <Tip>
      Abundly's webhook responds to Slack's verification request immediately, so this works whether or not you've pasted the bot token yet — you can do this step in any order.
    </Tip>
  </Step>

  <Step title="Invite to channels">
    In each Slack channel where you want the agent to work, type `/invite @YourAgentName`.
  </Step>
</Steps>

## How conversations work in Chat mode

With the default manifest and Chat mode, the agent listens broadly but replies selectively:

* **@mentions in a channel** start a new chat — the agent replies in the thread with full context.
* **DMs and group DMs** start (or continue) a chat — the agent replies naturally.
* **Follow-up thread replies** in a thread the agent already joined continue the same chat, so you don't need to @mention on every message.
* **Replies to agent-initiated threads** — When the agent posts a top-level message in a channel (e.g. a scheduled update or proactive notification), replies to that thread also continue the conversation without needing @mention.
* **All other messages** are ignored silently, so normal channel chatter doesn't spam the agent.

These conversations show up in the sidebar alongside portal chats, and can be filtered using the chat filter pill (e.g. *Hide remote chats*). The agent has the same instructions, memory, and capabilities as it does in the portal — so handoffs between Slack and the portal are seamless.

<Tip>
  Slack-originated chats are marked as "remote" and are read-only in the portal — reply from Slack to keep the thread in the agent's working context.
</Tip>

## Customizing scopes and events

The manifest includes defaults that work for most use cases. You can customize them in your Slack app settings:

| Setting        | Description                                                                                         |
| -------------- | --------------------------------------------------------------------------------------------------- |
| **Scopes**     | What the agent can do in Slack. See the [full list of scopes](https://api.slack.com/scopes).        |
| **Bot events** | Which events Slack sends to the agent. See the [full list of events](https://api.slack.com/events). |

If you don't want the agent to see all channel messages, remove `message.channels` and `message.groups` and add `app_mention` instead — the agent will then only receive @mentions and DMs.

## Accessing other Slack API methods

Beyond posting messages, reading channels, and reacting to messages, your agent can call any [Slack Web API method](https://api.slack.com/methods) — for example pinning messages, adding bookmarks, or listing user groups. Just ask in plain language:

* "Pin the latest status update in #launch"
* "Add a bookmark to #onboarding linking to our new-hire checklist"
* "List all user groups in the workspace and tell me who's in @design-team"

If a method requires an OAuth scope your app doesn't have yet, the agent will tell you exactly which scope to add. Open your Slack app settings, add the scope under **OAuth & Permissions → Bot Token Scopes**, reinstall the app to the workspace, and update the new bot token in Abundly.

## Working with file attachments

Your agent can both receive files from Slack and send files back into Slack messages and DMs.

**Incoming files.** When people share files in Slack, your agent can download those attachments into [agent documents](/features/documents) and use them in multi-step workflows — for example, reading and summarizing an attached PDF, saving it to the document library, or forwarding it to another capability like Notion.

**Outgoing files.** The agent can attach files when it posts to a channel or sends a DM. Supported sources include agent documents, images and other files produced by other capabilities, external URLs, and raw SVG content (auto-converted to PNG). Multiple files can be attached to a single message.

Example prompts:

* "Generate a pie chart of last month's sales by region and post it to #sales with a short summary"
* "When the weekly report PDF is ready, DM it to @Jane"
* "Take the diagram you just drew and share it in the thread"

The default manifest already includes `files:read` and `files:write`, which are needed for file workflows.

## Troubleshooting

<AccordionGroup>
  <Accordion title="Agent isn't responding to messages">
    1. **Check the mode** — In **Settings → Capabilities → Slack**, confirm you're in Chat mode if you expect the agent to reply automatically. Trigger mode does not auto-reply unless your instructions say so.
    2. **Check event subscriptions** — In your Slack app settings, go to **Event Subscriptions** and ensure:
       * The Request URL matches the webhook URL shown in Abundly.
       * The "Enable Events" toggle is ON.
       * The bot events from the manifest are present (`message.im` for DMs, plus `message.channels` / `message.groups` for public / private channel messages — or `app_mention` with `app_mentions:read` if you've narrowed to mention-only).
    3. **Invite the agent** — Type `/invite @YourAgentName` in the channel, or DM the bot directly.
    4. **Test the connection** — Use the **Test connection** button in Abundly to confirm the bot token is valid.
  </Accordion>

  <Accordion title="Agent replies only to @mentions, never to regular channel messages">
    By default, the agent listens to all messages in channels it's in, but only *replies* to @mentions, DMs, and follow-ups inside threads it's already part of. That's intentional — it keeps the agent from jumping into every conversation. To change this behavior, adjust your agent's instructions or customize the bot events in your Slack app.
  </Accordion>

  <Accordion title="Agent is responding to too many messages">
    By default the agent receives all messages in channels, groups and DMs it's a member of, so it can keep a conversation going in a thread without requiring an @mention every time. If you'd rather have it respond only when directly @mentioned:

    1. In your Slack app's **Event Subscriptions**, replace `message.channels` and `message.groups` in the bot events with `app_mention`.
    2. Add `app_mentions:read` to `scopes.bot` on **OAuth & Permissions**.
    3. Reinstall the app from the **Install App** page, then update the new bot token in Abundly.

    You can also remove the agent from noisy channels rather than changing scopes.
  </Accordion>

  <Accordion title="Permission errors">
    Ensure your Slack app has these minimum scopes:

    * `channels:read`, `groups:read`, `im:read` — To see public channels, private channels, and DMs
    * `channels:history`, `groups:history`, `im:history`, `mpim:history` — To read message content in each conversation type
    * `chat:write` — To send messages
    * `im:write` — To open DM conversations
    * `users:read` — To resolve user IDs and display names

    If you've added new scopes, reinstall the app from the **Install App** page in the left sidebar, then update the new bot token in Abundly.
  </Accordion>
</AccordionGroup>


# Snowflake
Source: https://docs.abundly.ai/integrations/snowflake

Turn your Snowflake data into answers — query and explore your warehouse in plain English

Your agent can explore your Snowflake databases and run read-only queries so you get insights, reports, and data checks without writing SQL. Ask in natural language and get answers. Authentication uses a Snowflake [Programmatic Access Token (PAT)](https://docs.snowflake.com/en/user-guide/programmatic-access-tokens) — no password — and queries time out after 30 seconds.

<Frame>
  <img alt="Screenshot: Snowflake capability card in Settings → Capabilities with account, login name and PAT configured" />
</Frame>

## How to enable

Go to **Settings → Capabilities** and enable **Snowflake**, then link a workspace secret containing your **Account Identifier**, **Login Name**, and **Programmatic Access Token (PAT)**. The same secret can be reused across agents and teams.

## Example use cases

* **Revenue and rankings** — "Query our Snowflake sales table and show top 10 customers by revenue this year"
* **Scheduled reports** — "Every Friday, run a query for weekly signups and post a summary to Slack"
* **Data reconciliation** — "Compare yesterday's order count in Snowflake with the count in our CRM and flag any mismatch"
* **Schema discovery** — "List the tables in our analytics schema and describe the columns in the events table"
* **Ad-hoc analysis** — "What were our top-performing campaigns last month by spend and conversions? Use the marketing schema in Snowflake"

<Frame>
  <img alt="Dr Plotts answering a question using Snowflake data" />
</Frame>

<Frame>
  <img alt="Dr Plotts chart from Snowflake data" />
</Frame>

<Note>
  The agent supports read-only operations only (e.g. querying data, listing tables, describing schemas). No DDL or DML.
</Note>

## How to connect Snowflake

<Steps>
  <Step title="Get your account identifier and login name">
    In Snowflake, open **Account Details**. Note your **Account Identifier** (format: <code>\<org>-\<account></code>) and **Login Name** (username).
  </Step>

  <Step title="Create a Programmatic Access Token (PAT)">
    Generate a PAT in Snowflake for programmatic access — see [Snowflake's PAT documentation](https://docs.snowflake.com/en/user-guide/programmatic-access-tokens) for how to create one. The PAT is used instead of a password for the integration.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities**, enable **Snowflake**, and create a new workspace secret of type **Snowflake**. Paste in the **Account Identifier**, **Login Name**, and **Programmatic Access Token (PAT)**.

    All three values live in a single workspace secret, so you only have to enter them once and can share the same secret across multiple agents.
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="BigQuery" icon="database" href="/integrations/bigquery">
    Query and explore your Google BigQuery data from the agent.
  </Card>

  <Card title="Steep" icon="chart-line" href="/integrations/steep">
    Query metrics from your Steep semantic layer without writing SQL.
  </Card>
</CardGroup>


# Steep
Source: https://docs.abundly.ai/integrations/steep

Query your Steep metrics in plain English — same definitions your team already uses

Your agent can discover and query your Steep metrics so you get answers in chat, in reports, or in automations without opening a dashboard or writing SQL. If you use Steep so your team shares one definition of MRR, conversion, or whatever matters — the agent uses those same [defined metrics](https://help.steep.app/introduction/why-steep). You ask in plain English; you get aggregated data (and the generated SQL when relevant) from the semantic layer you've already set up.

<Frame>
  <img alt="Screenshot: Steep capability card in Settings → Capabilities with API key configured" />
</Frame>

## How to enable

Go to **Settings → Capabilities** and enable **Steep**. Create an API key in Steep under **Settings → API** and paste it into the capability card. The key is stored as a team secret.

## Example use cases

* **Metric lookups** — "What was our MRR last month according to Steep?"
* **Scheduled reports** — "Every Monday, get the key product metrics from Steep and email a summary to the leadership team"
* **Cross-tool workflows** — "Pull this week's conversion metrics from Steep and add them to the weekly update in Notion"
* **Trends and analysis** — "Query Steep for signups by region over the last quarter and summarize the trend"
* **Board prep** — "Get our standard KPIs from Steep for the last 30 days and put the numbers into a one-pager for the board meeting"

<Frame>
  <img alt="Dr Plotts answering a question using Steep metrics" />
</Frame>

<Frame>
  <img alt="Dr Plotts returning a metric definition from Steep" />
</Frame>

<Note>
  The agent talks to Steep's API (api.steep.app). Ensure your API key has access to the metrics and data the agent needs.
</Note>

## How to connect Steep

<Steps>
  <Step title="Create an API key">
    In Steep, go to **Settings → API** and create an API key.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **Steep**. Paste the API key into the capability card. It is stored as a team secret (secret type: Steep).
  </Step>
</Steps>

## Learn more

<CardGroup>
  <Card title="BigQuery" icon="database" href="/integrations/bigquery">
    Query and explore your Google BigQuery data from the agent.
  </Card>

  <Card title="Snowflake" icon="snowflake" href="/integrations/snowflake">
    Query and explore your Snowflake data warehouse from the agent.
  </Card>
</CardGroup>


# Trello
Source: https://docs.abundly.ai/integrations/trello

Manage boards, lists, and cards

Connect your agent to Trello to manage cards, respond to board events, and automate your workflow.

## How to enable

Go to **Settings → Capabilities** and enable **Trello**. Then provide your API key, token, and board ID(s) — see below for details.

## Example use cases

* **Triage New Cards** — "When a new card is created in the Inbox list, analyze it and move it to the appropriate list based on priority"
* **Comment Responses** — "When someone comments on a card asking a question, research the answer and reply"
* **Weekly Cleanup** — "Every Friday, archive all cards in the Done list"
* **Cross-platform Sync** — "When you receive an email about a project update, create a card in the Updates list"

## Webhook events

When you configure a board ID, the agent automatically subscribes to that board and can react to events on it — card changes, comments, members, labels, checklists, attachments, and more.

By default, agents respond to a small set of common events (card created, card moved between lists, comment added, label removed). You can customize which events trigger the agent in the Trello capability settings, and optionally scope each event type to specific lists so the agent only fires when something happens in the lists you care about.

Mention the events you want to act on in your agent's instructions so it knows what to do when they fire.

## How to get your Trello credentials

<Steps>
  <Step title="Create a Power-Up">
    Go to [trello.com/power-ups/admin](https://trello.com/power-ups/admin) and click **New** to create a Power-Up. Fill in the required fields (name, email, etc.) and click **Create**.
  </Step>

  <Step title="Generate your API key">
    Open the Power-Up you created and go to the **API Key** tab. Click **Generate a new API key** and copy it.
  </Step>

  <Step title="Generate a token">
    In the same API Key section, click the **Token** link next to your key. Choose your permissions and expiration, then click **Allow**. Copy the token that appears.
  </Step>

  <Step title="Find your board ID">
    Open your Trello board in a browser. The board ID is in the URL: `trello.com/b/{boardId}/...`

    You can watch multiple boards by providing comma-separated IDs.
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities** and enable **Trello**. Enter your API key, token, and board ID(s).
  </Step>
</Steps>


# Twilio
Source: https://docs.abundly.ai/integrations/twilio

SMS and phone call infrastructure

<Warning>
  Twilio-based capabilities are currently hidden. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable them for your account.
</Warning>

Twilio powers SMS and phone call capabilities for your agent. For phone calls, the Abundly platform uses Twilio together with the [OpenAI Realtime API](/integrations/openai#realtime-voice-api) to handle natural voice conversations.

## How to enable

Go to **Settings → Capabilities** and enable the capabilities you need:

* **Send SMS** — Send text messages
* **Receive SMS** — Respond to incoming texts
* **Make Phone Call** — Initiate outbound calls
* **Receive Phone Call** — Answer incoming calls

See [Multi-channel communication](/features/communication) for an overview of agent communication channels, [Email & SMS](/features/messaging) for messaging details, and [Voice communication](/features/voice) for voice interactions.

## Example use cases

* **Appointment Reminders** — "Send a text reminder to all attendees 24 hours before their appointment"
* **SMS Triage** — "When you receive an SMS, analyze the request and respond or escalate as needed"
* **Outbound Calls** — "Call this lead and schedule a demo"
* **Inbound Support** — "When called about a technical issue, interview the user to learn about the problem, and then say that we will get back to them shortly. Then send an email summary to me and add a ticket to the support backlog in Notion."

<Note>
  You can configure a whitelist of allowed phone numbers and require approval before sending messages or making calls.
</Note>


# Twitter / X
Source: https://docs.abundly.ai/integrations/twitter

Search, analyze, and post on social media

Connect your agent to Twitter/X to search tweets, analyze users, and post content.

## How to enable

Go to **Settings → Capabilities** and enable the capabilities you need:

* **Twitter/X Search** — Search tweets, get user info, and analyze followers. Requires an API key from [TwitterAPI.io](https://twitterapi.io).
* **Twitter/X Post** — Post tweets and replies from your account. Requires Twitter OAuth credentials — see below.

## Example use cases

* **Brand Monitoring** — "Every morning, search Twitter for mentions of our company name and post a sentiment summary to Slack"
* **Competitor Tracking** — "Search Twitter for tweets from @competitor1 and @competitor2 this week and add a summary to my Notion research database"
* **Content Publishing** — "When interesting research articles are mentioned in slack channel #interesting-stuff, suggest a twitter post about it, wait for approval and then post it."
* **Community Engagement** — "Search Twitter for questions about our product, draft helpful replies, and send them to me for approval before posting"
* **Lead Research** — "When you receive an email about a new lead, search Twitter for their company and summarize their recent social activity"

## How to get Twitter/X credentials

### For Twitter/X Search

Twitter/X Search uses the [TwitterAPI.io](https://twitterapi.io) service. Sign up for an account and copy your API key.

### For Twitter/X Post

Posting requires OAuth credentials from an X Developer account:

<Steps>
  <Step title="Create an X Developer app">
    Go to [developer.x.com](https://developer.x.com) and sign in. Create a new project and app if you don't have one.
  </Step>

  <Step title="Enable user authentication">
    In your app settings, go to **User authentication settings** and enable it. Set **App permissions** to **Read and write**.
  </Step>

  <Step title="Generate credentials">
    Go to the **Keys and tokens** tab and generate:

    * **API Key** and **API Key Secret**
    * **Access Token** and **Access Token Secret**
  </Step>

  <Step title="Configure in Abundly">
    Go to **Settings → Capabilities**, enable **Twitter/X Post**, and enter all four credentials.
  </Step>
</Steps>

<Note>
  The X API Free tier has limited write access. You may need a Basic or Pro tier to post tweets via the API.
</Note>


# WhatsApp
Source: https://docs.abundly.ai/integrations/whatsapp

Connect your agent to WhatsApp so it can send and receive messages

Your agent can have direct conversations with people over WhatsApp — receiving messages and responding in natural conversation, just like in the portal. Each person who messages the WhatsApp number gets their own ongoing 1:1 chat with the agent.

## Example use cases

* A customer support agent that handles inquiries via WhatsApp 24/7
* A booking assistant that confirms appointments through WhatsApp messages
* An internal helper that team members can reach from their phone without opening the portal

## Prerequisites

* A Meta (Facebook) developer account
* A phone number to register with WhatsApp Business (can be new or existing)
* The Abundly agent you want to connect

## How to connect

<Steps>
  <Step title="Create a Meta App">
    Go to [developers.facebook.com](https://developers.facebook.com), create a new app, and add the **WhatsApp** product. This gives you a test phone number and API access.
  </Step>

  <Step title="Create a System User and generate a permanent token">
    In [Meta Business Settings](https://business.facebook.com/settings), go to **Users → System Users**, create one, and generate a token with these permissions:

    * `whatsapp_business_messaging`
    * `whatsapp_business_management`

    The temporary token from the API Setup panel expires hourly — the System User token is permanent.
  </Step>

  <Step title="Enable the WhatsApp capability">
    In the Abundly portal, open your agent's capabilities and enable **WhatsApp**. The settings panel shows your webhook URL, verify token, and a place to store your credentials.
  </Step>

  <Step title="Add your credentials">
    Create a WhatsApp secret with your **Phone Number ID** (from Meta's API Setup panel) and the **Access Token** you generated.
  </Step>

  <Step title="Configure the webhook in Meta">
    In your Meta App Dashboard under **WhatsApp → Configuration**, set the **Callback URL** and **Verify Token** shown in the capability panel. Then subscribe to the **messages** webhook field.
  </Step>

  <Step title="Publish the app">
    Under **App Settings → Basic**, fill in a privacy policy URL, app icon, and category, then click **Publish**. This switches the app to live mode so real messages are delivered to your webhook.

    <Note>
      Publishing does not make your app public — it just enables production webhook delivery.
    </Note>
  </Step>

  <Step title="Subscribe the app to your WABA">
    This step is sometimes needed if webhooks don't arrive after publishing. Call the Graph API to explicitly subscribe your app:

    ```bash theme={null}
    curl -X POST "https://graph.facebook.com/v21.0/YOUR_WABA_ID/subscribed_apps" \
      -H "Authorization: Bearer YOUR_ACCESS_TOKEN"
    ```
  </Step>

  <Step title="Send a test message">
    Send a WhatsApp message to your registered number. You should see it arrive in the agent's activity.
  </Step>
</Steps>

## Access control

Only whitelisted phone numbers can message the agent. Add allowed numbers in the capability settings panel under **Allowed Senders** — the agent will ignore messages from anyone else.

<Note>
  WhatsApp enforces a 24-hour session window. After a user messages, the agent can reply freely for 24 hours. After that, only pre-approved template messages can be sent to initiate a new conversation.
</Note>

## Learn more

<CardGroup>
  <Card title="Email & SMS" icon="envelope" href="/features/communication">
    Other messaging channels your agent can use
  </Card>

  <Card title="Slack" icon="slack" href="/integrations/slack">
    Connect your agent to Slack for team communication
  </Card>
</CardGroup>


# Getting Started
Source: https://docs.abundly.ai/introduction/getting-started

Create your first AI agent in minutes

<Tip>
  **Prefer video?** Watch our [9-minute walkthrough](/introduction/welcome#see-the-platform-in-action) showing how to build an agent from scratch.
</Tip>

## Step 1: Create an account

[Sign up at app.abundly.ai](https://app.abundly.ai/signup). New accounts include a free trial with credits — no credit card required.

## Step 2: Create an agent

<CardGroup>
  <Card title="Start from Scratch" icon="plus" href="https://app.abundly.ai">
    Click "New Agent" in your dashboard and give it a name.
  </Card>

  <Card title="Start from a Use Case" icon="copy" href="https://www.abundly.ai/use-cases">
    Open a use case and click "Build this agent" to create a starter agent with a pre-filled kickoff prompt. If needed, Abundly guides you through sign up/sign in and workspace selection first.
  </Card>
</CardGroup>

## Step 3: Configure and test

<Steps>
  <Step title="Define the job">
    Tell the agent what it should do, either by chatting with it or writing instructions directly. The agent can write it's own instructions based on the conversation, which is very convenient.
  </Step>

  <Step title="Add capabilities">
    Enable the capabilities the agent needs, for example email, Slack, image generation, code generation, etc. If you chat with the agent, it will ask for what it needs. See [Capabilities](/features/capabilities) for the full list.
  </Step>

  <Step title="Test and tweak">
    Put your agent to work, review its responses, give feedback, and refine the instructions until you're happy.
  </Step>

  <Step title="Done!">
    Your agent is ready to use!
  </Step>
</Steps>

<Info>
  Once your basic agent is working, you can add advanced features like triggers, schedules, multi-agent collaboration, and more. See [Features](/features/communication) for the full list.
</Info>

## Next steps

<CardGroup>
  <Card title="Capabilities" icon="plug" href="/features/capabilities">
    Explore tools and integrations for your agents.
  </Card>

  <Card title="Best Practices" icon="book" href="/guides/agent-design">
    Learn how to design effective agents.
  </Card>

  <Card title="Use Cases" icon="lightbulb" href="https://www.abundly.ai/use-cases">
    Get inspiration from real-world examples.
  </Card>

  <Card title="Features" icon="star" href="/features/communication">
    Discover advanced platform features.
  </Card>
</CardGroup>


# Platform Overview
Source: https://docs.abundly.ai/introduction/platform-overview

A quick tour of the Abundly interface

When you log in to Abundly, you'll land on the **Workspace page** — your home base for creating and managing AI agents and workspace members.

## Workspace page

<Frame>
  <img alt="The Workspace page showing agent cards and navigation" />
</Frame>

This is where you see all agents in your workspace. The page organizes agents into sections to help you find what you need faster:

* **Starred** — Agents you've bookmarked for quick access. Star an agent from its card.
* **Recently Used** — Agents you've opened recently (you can hide this section from your preferences).
* **All agents** — A flat, searchable list with an A–Z / Recent-first sort toggle.

You can filter by tag, search by name, and the filter state persists in the URL. Click **+ New agent** to create one.

The left sidebar provides access to workspace-level features:

* **Agents** — The main view showing all agents for you and your workspace
* **Members** — Invite colleagues and manage who has access to what
* **Management** — A dedicated section for admin-oriented tools, including **Agent management** (security, access controls) and **Usage & Limits** (credit monitoring and spending limits).
* **Help & support** — Links to documentation, what's new (the latest product update and changelog), contact support, service status, and more

<Info>
  You can get back to the workspace page at any time by clicking the Abundly logo in the top left.
</Info>

## Agent page

When you click on an agent, you'll be taken to the agent page. There you can configure, monitor, and interact with the agent.

<Frame>
  <img alt="The agent detail view showing chat, instructions, and capabilities" />
</Frame>

### Agent selector

The agent selector dropdown at the top left lets you quickly switch between agents.

### Sidebar links

These sidebar links let you configure and monitor the agent:

* <Icon icon="file-lines" /> **[Instructions](/features/instructions)** — The agent's job description and behavioral guidelines
* <Icon icon="file" /> **[Docs](/features/documents)** — Files and documents the agent can reference
* <Icon icon="bolt" /> **[Capabilities](/features/capabilities)** — Configure [integrations](/integrations/overview) and [MCP servers](/integrations/mcp) for your agent.
* <Icon icon="clipboard-list" /> **[Activity](/features/monitoring)** — Activity log and pending approval requests
* <Icon icon="gear" /> **Settings** — Agent configuration and advanced options:
  * Basic information — Name, avatar, and description.
  * Communication — Email address and chat start message
  * Usage & Limits — Daily credit usage and limits
  * Actions — Turn agent on/off, clone, delete
  * [Agent-to-agent communication](/features/multi-agent) — Configure which other agents this agent can communicate with
  * [User access](/security/access-control) — Control which workspace or team members can access this agent
  * [API access](/features/agent-api-endpoint) — Expose your agent as an HTTP API, an MCP server, or a webhook receiver
  * Public sharing — Make this agent available for others to clone
  * Alarms — Monitor which tasks the agent has scheduled for itself
  * Advanced config — [Model selection](/features/model-selection), per-context thinking settings, and more

### Chat links

Below the sidebar are the chat links:

* **+ New Chat** — Start a fresh conversation with the agent
* **Recent chats** — Access previous conversations to continue where you left off

### Chat

The main area where you interact with your agent. The chat input at the bottom lets you:

* **Type messages** — Just like any chat interface
* **[Use voice](/features/voice)** — Click the mic button to speak instead of type
* **Upload files** — Share documents, images, or data for the agent to process (you can also drag and drop files into the chat)
* **Walk & Talk mode** — Have a mostly hands-free conversation while moving around

See [Chat](/features/chat) for more details.

### Chat settings

The top bar provides an interactive **model picker** — click it to switch the AI model for this chat only — and toggles and links to chat features (see [Model Selection](/features/model-selection)):

* **Chat documents** - Opens the chat document panel so you can browse files connected to the current conversation.
* **Help** - Opens contextual help for the chat experience.
* **Thinking** - Turning this on makes the agent think out loud before responding, which can improve quality but can be slower.
* **Show tool calls and thinking** - Shows details about the tools the agent is using, and the agent's reasoning process (if thinking is enabled). Useful for getting a better understanding of how the agent is working.

<Info>
  If you want thinking to be on by default for new chats, you can configure that in Advanced Config under Settings.
</Info>

### Account menu

Click your profile image at the bottom left to access personal settings, switch workspaces (if you are a member of multiple workspaces), or sign out.

## Next steps

<CardGroup>
  <Card title="Getting Started" icon="rocket" href="/introduction/getting-started">
    Create your first agent step by step
  </Card>

  <Card title="Instructions" icon="file-lines" href="/features/instructions">
    Learn how to configure agent behavior
  </Card>

  <Card title="Capabilities" icon="plug" href="/features/capabilities">
    Explore available tools and integrations
  </Card>

  <Card title="Chat" icon="comments" href="/features/chat">
    Deep dive into the chat interface
  </Card>
</CardGroup>


# Pricing
Source: https://docs.abundly.ai/introduction/pricing

Understanding Abundly pricing and credit usage

Abundly uses a credit-based pricing model that scales with your usage.

<Card title="View Plans & Pricing" icon="tags" href="https://app.abundly.ai/prices">
  See current plans, pricing, and credit details at app.abundly.ai/prices
</Card>

## Your workspace

Every Abundly account is a workspace — a shared environment where humans and agents collaborate together. Pricing is credit-based, and your member limit depends on your plan:

* **Unlimited agents** — create as many agents as you need
* **Team members** — invite your team (seat count depends on your plan)
* **You pay for credits** — consumed when agents perform actions

## How credits work

When your agents perform actions — sending emails, responding to chat messages, analyzing documents, making API calls, etc. — they consume credits. The cost depends on the amount of content the agent reads/writes, and the tools it uses.

Simple tasks use fewer credits, while complex multi-step operations use more. See the [Agent Optimization](/guides/agent-optimization) guide for detailed examples.

## How billing works

Credits are managed at the account level and shared by all your agents:

* **Initial credits** — you receive credits when you sign up
* **Monthly subscription credits** — each subscription includes a monthly credit allotment that refills at the start of every billing period. Unused subscription credits don't carry over — you always start a new month with your full plan allotment.
* **Top-ups** — purchase additional credits whenever you need them. Top-up credits stay on your account until used and aren't affected by the monthly refill.

## Managing costs

You have full control over your credit consumption:

* **Track usage** in the Usage Reports dashboard
* **Set limits** per agent to prevent runaway consumption
* **Optimize agents** by simplifying instructions and reducing context size

<CardGroup>
  <Card title="Usage & Limits" icon="gauge-high" href="/features/usage">
    Learn how to track and control credit consumption.
  </Card>

  <Card title="Agent Optimization" icon="gauge-high" href="/guides/agent-optimization">
    Reduce costs and improve agent performance.
  </Card>
</CardGroup>

## FAQ

<AccordionGroup>
  <Accordion title="What happens when I run out of credits?">
    * **Agents automatically pause** when your credit balance reaches zero
    * **No data loss** — all configurations, documents, and chat history remain safe
    * **Easy restart** — agents resume immediately once you add more credits
  </Accordion>

  <Accordion title="How do I buy more credits?">
    Go to **Billing** in your account dashboard and look for the **Credits** tab to top up. You need an active paid subscription to purchase additional credits.

    Workspaces that aren't on a Stripe subscription can top up against a **FortNox invoice** from the same Credits tab — confirm the purchase, an invoice is issued, and the credits are added to your workspace right away.

    See [Plans & Pricing](https://app.abundly.ai/prices) for subscription options.
  </Accordion>

  <Accordion title="How do I get an invoice or receipt for my payments?">
    Open the **Payments** tab on the Billing page to see every successful charge, including subscription renewals and credit top-ups. Each row has:

    * **Invoice** — a Stripe-hosted invoice with your workspace billing details and applicable VAT/tax, suitable for accounting
    * **Receipt** — a simple proof-of-payment receipt

    Invoices are generated automatically for both subscription payments and credit top-ups. Very old top-up charges may only have a Receipt button.
  </Accordion>

  <Accordion title="Can I upgrade from trial at any time?">
    Yes — go to your account settings and choose a plan. Your subscription benefits start immediately and all your data remains intact. Contact [support@abundly.ai](mailto:support@abundly.ai) if you need help.
  </Accordion>

  <Accordion title="Can I apply a promotion code?">
    Yes — workspace admins can apply Stripe promotion codes directly from the **Plans / Billing** page. You don't need to switch plans first; just enter the code on your current subscription.
  </Accordion>
</AccordionGroup>


# Welcome
Source: https://docs.abundly.ai/introduction/welcome

Learn about the Abundly agent platform

An autonomous AI agent is a digital colleague that carries out knowledge work on behalf of you and your colleagues. We are quickly moving towards a world with **teams of humans and AI agents working side by side**.

Agents need a place to live, and that's what the Abundly agent platform is - an enterprise-grade **platform to build, configure, manage and collaborate with autonomous AI agents**. The platform is designed to be easy to use for non-technical users, while providing advanced features, security, and governance for power users.

Abundly is more than just a tool-enabled AI chat. It is an **operating system for AI agents** - a place where agents and humans work together like colleagues.

<Frame>
  <img alt="Teams of humans and agents working side by side" />
</Frame>

## See the agent platform in action

Watch this 9-minute video showing how to build an advanced AI agent that processes invoices, communicates with teams via email, flags issues via Slack, makes voice calls for urgent matters, and creates interactive dashboards.

<iframe title="Abundly Platform Demo" />

## What makes Abundly different

* **True Autonomy** — Schedule recurring tasks, react to triggers (emails, webhooks, schedules), and operate 24/7 without constant prompting.
* **Multi-Modal Communication** — Native support for SMS, email, Slack, voice calls, and more. Chat is just one of many channels.
* **Conversational Configuration** — Configure agents by talking to them, like onboarding a new colleague. They write their own instructions based on the conversation. No need to write code or draw complex diagrams.
* **Intelligent Problem Solving** — Agents handle uncertainty and adapt to unexpected events. You control the flexibility vs. predictability tradeoff.
* **Collaborative by Design** — Agents are part of your workspace and can interact with multiple people, they are not just personal assistants.
* **Multi-Agent Collaboration** — Agents can handle complex tasks by collaborating with other agents and delegating work to subagents.
* **Dynamic Apps & Scripts** — Agents can create interactive apps to ease human collaboration, and scripts to automate their work on the fly.
* **Enterprise Security and Governance** — Comprehensive monitoring, access controls, audit logs, and usage management.

## What you can do

These are examples of prompts and instructions that work directly in Abundly, but are difficult or impossible on many other platforms:

* "Dig through last week's messages on the #support channel in slack, and create an overview doc in Notion."
* "Every morning, check my todos in HubSpot and if something seems urgent, ping me on Slack."
* "Whenever a new customer lead is added to our Trello board, do research on the companies mentioned there, and attach it to the card."
* "Go through these 300 lines of investment prospects in this spreadsheet, and analyze each one based on the following criteria..."
* "If an urgent security-related problem is mentioned in Slack, notify me via SMS."
* "Create a database to track all support issues and keep it up-to-date as they are resolved."
* "Create an interactive dashboard to show all ongoing support tickets, and historical trends."

## Next steps

<CardGroup>
  <Card title="What is an AI Agent?" icon="circle-question" href="/introduction/what-is-an-agent">
    Understand the core concepts behind autonomous AI agents.
  </Card>

  <Card title="Getting Started" icon="rocket" href="/introduction/getting-started">
    Create your first agent in minutes.
  </Card>

  <Card title="Features" icon="star" href="/features/communication">
    Explore the platform's capabilities.
  </Card>

  <Card title="Use Cases" icon="lightbulb" href="https://www.abundly.ai/use-cases">
    See what others are building with Abundly.
  </Card>
</CardGroup>


# What is an AI Agent?
Source: https://docs.abundly.ai/introduction/what-is-an-agent

Understanding autonomous AI agents and how they work

An Abundly agent is a multilingual, autonomous digital colleague that carries out knowledge work on behalf of you and your workspace.

## A new type of coworker

Autonomous AI agents are a new concept that takes some getting used to.

Knowledge work in companies is traditionally carried out by either code or by humans. **Agents are a new type of worker** — not quite code, not quite human, but somewhere in between.

<Frame>
  <img alt="Code vs Agent vs Human" />
</Frame>

An agent is:

* **Faster than humans** but slower than code
* **More predictable than humans** since it isn't affected by sleep, personal conflicts, or mood swings—but less predictable than code.
* **More intelligent than code** since it can deal with uncertainty and fuzzy inputs—but less intelligent than a human for most tasks.
* **Cheaper than humans** for most tasks, but more expensive than code

A well-crafted agent with the right tools and context is like a super-colleague, able to carry out complex tasks that would be impossible to do with code, and prohibitively slow and expensive to do with humans.

## Augment, not replace

The Abundly agent platform is designed to augment humans rather than replace them. Most professionals spend significant time on low-level routine tasks — work that takes time away from higher-value activities but can't be automated with code because it involves fuzzy inputs, uncertainty, and judgment.

An AI agent handles those tasks, freeing humans to focus on strategic work and make better use of their time.

## Benefits of AI agents

Compared to doing tasks 100% manually:

* **Lower cost** — Despite token usage costs, total cost is often lower than manual work.
* **Higher speed** — Agents process information faster, don't need sleep or breaks, and can do any number of tasks in parallel. You never need to wait to find time in the agent's calendar—they are always available.
* **Better control** — Agents act based on written instructions, and will tend to do the same task the same way, with less variation than if the task was done manually.
* **Higher quality** — Agents are persistent and thorough. They read their given context and documents carefully and take every word into account, every time, while we humans sometimes get tired or distracted and take shortcuts.

<Info>
  Agents are not infallible. Like humans, they can make mistakes and need oversight. But with good instructions and context, they tend to make fewer mistakes than humans.
</Info>

## The four components of an agent

An Abundly agent consists four main components: LLM (Large Language Model) + Mission + Tools + Autonomy

<Frame>
  <img alt="What is an AI Agent" />
</Frame>

<AccordionGroup>
  <Accordion title="LLM (Large Language Model)" icon="brain">
    The external "brain" — a <Tooltip>LLM</Tooltip> — that allows the agent to process information and make decisions.

    Abundly integrates with all major LLMs (Claude, GPT, Gemini) and provides a unified interface. By default we select the best model for agentic behavior, but advanced users can choose which model to use based on cost, speed, and capability tradeoffs.
  </Accordion>

  <Accordion title="Mission" icon="bullseye">
    The agent's job description, or instructions, written in natural language. The instructions are used by the agent in all contexts — whether responding to a chat message, reacting to a trigger, or calling a tool.

    Instructions are versioned and can be updated over time by the user or by the agent itself. You choose the level of detail to provide, depending on how predictable vs creative the agent should be. Additional context can be provided via documents or links.
  </Accordion>

  <Accordion title="Tools" icon="wrench">
    The tools the agent has access to. Tools are grouped into **capabilities** — for example, the Slack capability includes tools like posting messages, listing channels, and reading threads.

    You decide which capabilities the agent should have, and the agent will ask for additional capabilities if it needs them. Available capabilities include web search, web scraping, deep research, code execution, HTTP for interacting with any API, phone calling, SMS, and 40+ integrations like Slack, Gmail, Trello, Sharepoint, Notion, GitHub, and more.

    You can require human approval for sensitive actions (e.g., sending an email to an external domain) and set constraints like email address whitelists.
  </Accordion>

  <Accordion title="Autonomy" icon="robot">
    The ability to act independently, beyond just responding to chat messages:

    * Schedule recurring tasks ("every Monday at 9am...")
    * React to triggers (incoming email, webhooks, calendar events, etc.)
    * Proactively contact humans or other agents to share information or ask for input

    **Transparency:** You can see what the agent is doing in real-time and after the fact.

    **Control:** You choose how much autonomy the agent should have.
  </Accordion>
</AccordionGroup>


# Access Control
Source: https://docs.abundly.ai/security/access-control

Role-based permissions for users and agents

Control who can access which agents and what actions they can perform. Access control operates on three layers: workspace membership, team membership (if enabled), and per-agent access settings.

<Frame>
  <img alt="Workspace page showing members with their workspace roles (Admin, Creator, Collaborator, or Guest) and team memberships" />
</Frame>

## Workspace roles

Every user in your workspace has a role that determines their base level of access:

| Role             | Description                                                                                                         |
| ---------------- | ------------------------------------------------------------------------------------------------------------------- |
| **Admin**        | Full control over the workspace—can manage members, billing, capabilities, and all agents                           |
| **Creator**      | Standard workspace member—can access agents based on agent-level settings, and can create, clone, and import agents |
| **Collaborator** | Same as Creator, but cannot create, clone, or import agents, or manage workspace secrets and API capabilities       |
| **Guest**        | Limited access—can only access agents they've been explicitly granted access to                                     |

Workspace admins automatically have Full Access to all agents in the workspace, regardless of agent-specific settings. See [Workspace Members](/features/workspace-members) for how to assign roles.

## Agent access levels

Each agent has access settings that control what workspace members can do with it. There are four levels:

| Level           | What it allows                                                                   |
| --------------- | -------------------------------------------------------------------------------- |
| **Full Access** | Full control: delete agent, change access settings, edit configuration, interact |
| **Edit**        | Modify instructions, capabilities, documents, and interact with the agent        |
| **Use**         | Interact with the agent (chat), but cannot change anything                       |
| **Nothing**     | Cannot see or interact with the agent                                            |

The hierarchy is: **Full Access > Edit > Use > Nothing**

## Default access vs. user-specific access

Agent access is controlled through two settings, found in the agent's **Settings > User access** tab:

<Frame>
  <img alt="Agent user access settings showing the default access level and user-specific access rules" />
</Frame>

**Default access** (top section) sets the baseline for everyone who can see the agent—all workspace members if the agent is not assigned to a team, or all members of that team if it is. If you set this to "Edit," everyone in that scope can modify the agent.

**User-specific access** (bottom section) lets you give individual users a different access level. For example, you might set the default to "Use" but give a specific colleague "Full Access".

To remove a user-specific override entirely and revert the user to the workspace default, click the red trash button next to their row. A confirmation modal prevents accidental removal.

<Note>
  For most access levels, the system grants the **highest** access between the default and user-specific access—so user-specific rules are typically used to grant **more** access. The exception is **Nothing (No Access)**: when set as a user-specific override, it acts as an explicit deny, blocking access regardless of the default level.
</Note>

### Private agents

When you set the default access to "Nothing," the agent becomes private:

* Other workspace or team members cannot see the agent—it won't appear in their agent list
* Workspace administrators (and team administrators, if the agent is assigned to a team) can still see the agent in the management view — including its description, instructions, agent-to-agent connections, and exposure hooks (MCP server, HTTP API, webhooks, widget) — but cannot interact with it directly
* Only users with explicit user-specific access can interact with it

This is useful for agents that handle sensitive data or are still in development.

### Inviting guests by email

The **Invite by email** section on the agent's User access tab lets you grant access to this agent by email address — including to people who are not in your workspace yet:

* If the email belongs to an existing workspace member, they get the chosen access level directly.
* Otherwise an invitation is sent (the email is optional — the invite also appears when the person logs in). When they accept, they join the workspace as a **Guest** and receive the chosen access level on the agent.

Guests can log in but only see the agents they've been granted access to — no workspace, team, or member surfaces. They don't count toward your plan's member limit, and they're only visible to workspace admins on the Members page (and to team admins in their teams). Pending guest invitations are listed on the agent's User access tab, where they can be revoked.

<Note>
  From the UI, guests can be granted **Use** or **Edit** access. Their access to the agent works exactly like any other user-specific access rule.
</Note>

## Admin-restricted agents

In each agent's **Settings > User access** tab, users with Full Access can narrow who in the portal can see and interact with the agent. There are three tiers:

| Tier                         | Who can see the agent                                                                                     |
| ---------------------------- | --------------------------------------------------------------------------------------------------------- |
| **No restriction** (default) | Anyone with access via the rules above                                                                    |
| **Team admins only**         | Workspace admins, plus team admins of this agent's team (only available when the agent belongs to a team) |
| **Workspace admins only**    | Workspace admins only                                                                                     |

When a restriction is active, the agent — its chats, instructions, capabilities, and configuration — is hidden from anyone outside the chosen admin tier, regardless of any default or user-specific access rules above.

Use the tighter tiers for agents that work with internal data — billing, analytics, member lists, audit logs — that the rest of the workspace or team shouldn't see. Some capabilities (for example workspace and team analytics) are also gated behind the matching admin tier for the same reason.

<Note>
  Admin restrictions narrow the **portal** only. External entry points an admin has enabled — embeddable widget, HTTP API, MCP, webhooks, public clone link — keep working under their own authentication. Agent-to-agent communication is also unaffected, so other agents can still reach the agent if they're configured to.
</Note>

## Teams

For larger workspaces, [teams](/features/teams) add another layer of access control between the workspace and individual agents:

* Agents can belong to a team
* Users must be members of that team to access the agent (unless they're workspace admins)
* Team admins can manage agents within their team

If an agent belongs to a team, access is determined by the user's role in that specific team—not their workspace-wide role.

<Note>
  Teams are an enterprise feature. Contact [support@abundly.ai](mailto:support@abundly.ai) to enable them for your workspace.
</Note>

## Agent-to-agent access

Agent-to-agent access is configured per agent in Settings → Agent Communication, and has two sides:

* **Discoverability (inbound)** — who is even *allowed* to reach this agent: **No one**, **Team** (agents in the same team, default), or **Everyone** (any agent in the workspace, including across teams).
* **Outbound** — which other agents this agent is configured to call. Picked explicitly per target from the communication graph.

<Frame>
  <img alt="Agent-to-agent communication settings showing which agents can communicate with this agent" />
</Frame>

Agents only exchange messages when both sides align: the target must be discoverable by the caller, and the caller must explicitly allow the target in its outbound list. Tightening a target's discoverability (e.g. from **Everyone** back to **Team**) immediately blocks existing outbound connections from callers that can no longer reach it.

See [Multi-Agent Collaboration](/features/multi-agent) for more on agent-to-agent communication patterns.

<Warning>
  Agent-to-agent communication can create indirect access paths. If Agent A can talk to Agent B, and Agent B has access to sensitive data, users who talk to Agent A may be able to access that data indirectly. Be deliberate about setting an agent's discoverability to **Everyone** — it broadens the set of agents (and, transitively, users) that can reach it.
</Warning>

## Best practices

**Start with restrictive defaults.** Set the default access to "Use" or "Nothing" for sensitive agents, then grant higher access to specific users who need it.

**Use teams for scale.** If you have many agents or multiple departments, teams make access management much simpler than per-agent user rules.

**Review access after membership changes.** When someone leaves or changes roles, audit their access to ensure it's still appropriate.

**Keep at least one user with Full Access.** The platform prevents you from removing all Full Access from an agent—every agent needs at least one user with Full Access.

## FAQ

<AccordionGroup>
  <Accordion title="What's the difference between 'Edit' and 'Full Access'?">
    **Edit** access lets you modify how the agent works—instructions, capabilities, documents, and other settings.

    **Full Access** includes everything in Edit, plus the ability to delete the agent and change who else can access it.
  </Accordion>

  <Accordion title="Can I restrict a user to less access than the default?">
    Yes—set their user-specific access to **Nothing (No Access)**. This acts as an explicit deny, blocking access regardless of the default level. To remove an override and revert a user to the default instead, click the red trash button next to their row.
  </Accordion>

  <Accordion title="Can Guests use agents?">
    Guests cannot access agents through the default access level—they need explicit user-specific access to each agent. This makes the Guest role useful for external collaborators who should only see specific agents.
  </Accordion>

  <Accordion title="Do I need to be in a team to access an agent in that team?">
    Yes. If an agent belongs to a team, you must be a member of that team to access it. The exception is workspace administrators, who automatically have access to all teams.
  </Accordion>

  <Accordion title="Can agents access each other's data?">
    Only if you explicitly configure agent-to-agent communication between them. By default, agents are isolated—they can't see each other's documents, databases, or conversations.
  </Accordion>

  <Accordion title="What happens when someone leaves the workspace?">
    Remove them from the workspace in the Members page. They'll immediately lose access to all agents and data. Their past conversations and actions remain in the logs.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Agent Management" icon="sliders" href="/features/agent-management">
    Workspace-level agent and capability controls
  </Card>

  <Card title="Teams" icon="users-rectangle" href="/features/teams">
    Organize agents and users by department or function
  </Card>

  <Card title="Multi-Agent" icon="arrows-left-right" href="/features/multi-agent">
    How agents collaborate and communicate
  </Card>
</CardGroup>


# Attack Detection
Source: https://docs.abundly.ai/security/attack-detection

Automated screening of untrusted trigger content before agent execution

The platform includes automated attack detection that screens incoming trigger content before the agent acts on it. This provides protection against manipulation attempts—especially prompt injection through channels like email and SMS where anyone can send content to the agent.

## How it works

For security-sensitive triggers, the platform adds an attack detection step before execution:

1. **Trigger arrives** — The agent receives an event from an untrusted channel (e.g., an incoming email or SMS)
2. **Attack detection** — The platform analyses the trigger content for manipulation attempts
3. **Decision** — If an attack is detected, the trigger is blocked and admins are alerted. Otherwise, the agent proceeds to act on the event.

<Note>
  Attack detection runs in a separate context from the main agent. This is critical: the detection evaluates the raw trigger content before the agent has a chance to process it, catching manipulation attempts before they can influence the agent's behavior.
</Note>

## What it checks for

The attack detection evaluates trigger content for:

| Check                 | What it looks for                                                   |
| --------------------- | ------------------------------------------------------------------- |
| **Prompt injection**  | Attempts to override the agent's instructions through crafted input |
| **Data exfiltration** | Instructions designed to trick the agent into leaking data          |
| **Jailbreak**         | Attempts to bypass the agent's safety boundaries                    |
| **Reconnaissance**    | Probing to discover what the agent has access to                    |

## When an attack is blocked

When the platform detects an attack:

1. The agent does not act on the trigger
2. The block is logged to the [activity log](/features/monitoring) with the detection reasoning
3. Alert emails are sent to configured recipients
4. A diary entry is written documenting the blocked attack

## When does it run?

Attack detection runs automatically for:

* **Incoming emails** — Anyone can send an email to your agent
* **Incoming SMS** — Anyone can text your agent's number
* **Script escalations** — When a script escalates to the full agent pipeline (unless the script explicitly opts out)

Other trigger types (scheduled tasks, webhooks from authenticated sources, etc.) do not require attack detection because the content source is already controlled.

## Configuring alerts

Workspace administrators can configure attack detection alert behavior in **Workspace → Settings**.

| Setting                     | What it does                                                                                            |
| --------------------------- | ------------------------------------------------------------------------------------------------------- |
| **Alert recipients**        | Choose whether alerts go to all workspace admins, admins plus additional emails, or a custom email list |
| **Include blocked content** | Optionally include the blocked trigger payload in alert emails for debugging                            |

## FAQ

<AccordionGroup>
  <Accordion title="Does attack detection check every trigger?">
    No. It only screens untrusted trigger content—primarily incoming emails and SMS. Triggers from authenticated or controlled sources (like scheduled tasks or authenticated webhooks) skip this step.
  </Accordion>

  <Accordion title="Can attack detection be bypassed by clever prompting?">
    Attack detection runs before the agent processes the content, and in a separate context. The detection model evaluates the raw input independently, making it resistant to the manipulation techniques it's designed to catch.
  </Accordion>

  <Accordion title="What if attack detection blocks a legitimate message?">
    Check the activity log to understand why it was blocked. If the message was legitimate, review the content for patterns that might look suspicious (e.g., instructions that resemble prompt injection). You can also adjust the agent's instructions to make its expected interactions clearer.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Guardrails" icon="fence" href="/security/guardrails">
    Code-enforced constraints like whitelists and approval requirements
  </Card>

  <Card title="User Approval" icon="user-check" href="/security/user-approval">
    Human-in-the-loop for sensitive actions
  </Card>

  <Card title="Activity Monitoring" icon="chart-line" href="/features/monitoring">
    See attack detection results in the activity log
  </Card>
</CardGroup>


# Credentials
Source: https://docs.abundly.ai/security/credentials

Secure storage of secrets and API keys

Agents need credentials to access external services. Depending on the integration, credentials are handled in different ways—some require no setup, others need API keys, and some use OAuth to connect your personal accounts.

<Frame>
  <img alt="Three capabilities showing different credential patterns: Slack with user-entered token, Web Search with no credentials needed, and Google Drive with a connected personal token" />
</Frame>

## Three credential patterns

### Platform-provided

Some capabilities work out of the box with no credentials required. The platform uses its own accounts and you pay through credits.

**Example:** [Web Search](/integrations/perplexity) uses Perplexity under the hood. You just enable the capability and start using it—credits are deducted for usage.

### User-entered credentials

For services where you bring your own account, you enter credentials directly in the capability settings. Typically this means pasting an API token or key.

**Example:** [Slack](/integrations/slack) requires a bot token and app ID from your Slack workspace. You create a Slack app, copy the credentials, and paste them into the capability settings.

### Personal tokens (OAuth)

For services that access your personal data, you authorize via OAuth. This creates a personal token tied to your account.

**Example:** [Google Drive](/integrations/google-drive) requires access to your Google account. You click "Connect" and authorize through Google's OAuth flow. The resulting token is yours—other users who interact with the agent can trigger actions, but only within the scope you authorized.

## Managing personal tokens

Personal tokens are managed in your profile settings under **Personal Tokens**. You can see all your connected services, add new tokens, rename tokens, review which agents use each token, and revoke access.

OAuth tokens refresh automatically in normal use. For expired tokens, you can manually refresh when available. For Google Drive, if Google has revoked access or refresh no longer works, you can reconnect the existing token in place so agents keep their file access. See [Google Drive](/integrations/google-drive#if-your-token-expires-or-is-revoked).

<Frame>
  <img alt="Personal Tokens page showing connected Google Drive and Calendar tokens" />
</Frame>

When you add a new personal token, you choose the service and authorize access:

<Frame>
  <img alt="Add Personal Access Token modal showing available OAuth services" />
</Frame>

Currently supported OAuth services:

* Google Drive
* Google Calendar
* Gmail
* SharePoint
* Outlook

<Tip>
  Once created, a personal token can be used across multiple agents. You don't need to re-authorize for each agent—just select your existing token when configuring the capability.
</Tip>

### Multiple accounts on one agent

For Gmail, an agent can connect more than one account—for example a personal inbox and a work inbox managed by the same assistant. Connect each account as its own personal token, then refer to the accounts by email address in the agent's instructions (e.g. "work mail is [anna@company.com](mailto:anna@company.com)").

## How credentials are secured

| Measure               | Description                                                         |
| --------------------- | ------------------------------------------------------------------- |
| **Encrypted storage** | Stored encrypted in database—not readable even with database access |
| **Key separation**    | Decryption keys stored separately from encrypted data               |
| **Access control**    | Only authorized agents can use specific credentials                 |
| **No LLM exposure**   | Credentials are never included in prompts or visible to the LLM     |
| **Auto-refresh**      | OAuth tokens are automatically refreshed before they expire         |

<Note>
  The LLM never sees your credentials. It only knows which capabilities are available (e.g., "Gmail - Read and send emails"). When an agent calls an external API, the platform injects authentication after the LLM has generated the request. This prevents credentials from leaking through model outputs or logs.
</Note>

## Best practices

**Use least privilege.** When creating API keys or OAuth grants, give only the permissions the agent actually needs. A read-only key is safer than one with write access.

**Rotate after membership changes.** When someone leaves your workspace, rotate any credentials they might have had access to through agents.

**Revoke unused tokens.** In your Personal Tokens page, remove tokens for services you no longer use.

**Never put credentials in instructions.** Credentials in agent instructions would be visible to anyone with access to the agent and potentially exposed in logs. Always use the credential system.

## If a credential is compromised

If you suspect a credential has been compromised:

1. **Revoke immediately** — For personal tokens, go to your Personal Tokens page and disconnect the service. For user-entered credentials, regenerate the API key at the source (e.g., Slack, GitHub).
2. **Update agents** — Any agents using that credential will lose access until you reconnect with new credentials.
3. **Review activity** — Check the [activity log](/features/monitoring) for any suspicious actions taken before the credential was revoked.

<Tip>
  Need the same credential available to several agents? Store it as a **Secret** in **Shared Assets → Secrets**. Secrets can be scoped to the whole workspace, one or more teams, or a hand-picked set of agents, and they're used for authenticating [HTTP requests](/features/http) and custom API capabilities. The same multi-team scoping applies to **API Capabilities** in **Shared Assets → API Capabilities**.
</Tip>

## Learn more

<CardGroup>
  <Card title="Slack" icon="slack" href="/integrations/slack">
    Set up a Slack app and connect it to your agent
  </Card>

  <Card title="Google Drive" icon="google-drive" href="/integrations/google-drive">
    Connect your Google account and share files
  </Card>

  <Card title="Integrations" icon="plug" href="/integrations/overview">
    All available integrations and how to configure them
  </Card>
</CardGroup>


# Guardrails
Source: https://docs.abundly.ai/security/guardrails

Technical constraints enforced by code, not LLM reasoning

Guardrails are technical constraints enforced by platform code—not by LLM reasoning. This is a critical distinction: even if an agent is manipulated through prompt injection or given conflicting instructions, guardrails cannot be bypassed. The platform simply won't execute the action.

## Why guardrails matter

Telling an agent "don't email external domains" in its instructions is helpful, but not foolproof. LLMs can be tricked, confused, or make mistakes. Guardrails provide a hard limit that works regardless of what the agent thinks it should do.

The agent never gets a chance to override this—the check happens at the platform level before the action executes.

## Capability-specific guardrails

Different capabilities have different guardrail options. Configure these in each capability's settings.

### Email

<Frame>
  <img alt="Email capability settings showing allow-list and approval options" />
</Frame>

| Setting                                          | Options                                                   | What it does                                  |
| ------------------------------------------------ | --------------------------------------------------------- | --------------------------------------------- |
| **Restrict who the agent can email**             | Off, or on with an allow-list of addresses and/or domains | Hard limit on who the agent can contact       |
| **If the agent tries to email inside the list**  | Send the email, or ask for your approval                  | Behavior for recipients on the allow-list     |
| **If the agent tries to email outside the list** | Block the email, or ask for your approval                 | Behavior for recipients not on the allow-list |

A typical setup is restrict-on with "send inside" and "ask outside" — the agent can email known contacts freely, and you review anything new. When restriction is off, you instead get a single "ask before every email" toggle. See [User Approval](/security/user-approval) for how the approval flow works.

### Other capabilities

* **SMS and Phone calls** — Same allow-list plus inside/outside-list approval model. See [Email & SMS](/features/messaging) and [Voice](/features/voice) for details.
* **HTTP requests** — Optional approval for each request. See [HTTP Requests](/features/http) for details.

## Configuring guardrails

1. Open the agent and navigate to **Capabilities** in the sidebar
2. Click on the capability you want to configure (Email, SMS, Phone, etc.)
3. Choose whether to restrict recipients, configure the allow-list, and set the inside-list and outside-list rules
4. Test by trying an action that should be blocked or require approval

<Tip>
  You can see guardrail settings at a glance in the capability card — each card summarises the current allow-list and
  approval rules in one line.
</Tip>

## Workspace-level controls

Workspace administrators can control which capabilities are available to agents across the workspace. This is configured in **Workspace management** under the **Capabilities** tab. They can also limit which **LLM models** agents may pick for new work under **Workspace management → Model selection** (models disabled platform-wide by Abundly cannot be turned on there).

<Frame>
  <img alt="Workspace management capabilities tab showing workspace-wide settings and capability toggles" />
</Frame>

| Control                | What it does                                               |
| ---------------------- | ---------------------------------------------------------- |
| **Capability toggles** | Enable or disable specific capabilities workspace-wide     |
| **Default mode**       | Whether new capabilities are allowed or blocked by default |
| **Custom MCP servers** | Allow or block users from adding their own MCP servers     |

Capabilities disabled at the workspace level cannot be enabled by individual agents.

Workspace administrators can also configure attack detection notification behavior in **Workspace → Settings**.

| Security setting            | What it does                                                                                                                |
| --------------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| **Alert recipients**        | Choose whether blocked attack alerts go to all workspace admins, admins plus additional emails, or a custom email list only |
| **Include blocked content** | Optionally include the blocked trigger payload in alert emails for debugging                                                |

<Note>
  If your workspace uses [teams](/features/teams), team administrators can set additional restrictions for their team,
  but cannot enable capabilities that are disabled at the workspace level.
</Note>

## Best practices

**Start restrictive.** Begin with tight guardrails—require approval for external communication, limit capabilities to what's needed. You can always loosen them as you gain confidence.

**Combine an allow-list with "ask for approval" outside it.** This lets agents work freely with known contacts inside your organization while routing anything new through you. It's a good balance of autonomy and safety.

**Review blocked and approved actions.** Check the [activity log](/features/monitoring) periodically to see what's being blocked or approved. This helps you identify false positives (legitimate actions being blocked) and tune your configuration.

**Match guardrails to risk.** A customer-facing agent needs tighter guardrails than an internal research assistant. See [Risk & Autonomy](/security/risk-autonomy) for guidance on assessing risk.

## FAQ

<AccordionGroup>
  <Accordion title="Can an agent bypass guardrails through clever prompting?">
    No. Guardrails are enforced by platform code, not LLM reasoning. The agent can't talk its way past an allow-list check—the platform simply won't execute the blocked action.
  </Accordion>

  <Accordion title="What happens when an action is blocked?">
    If the outside-list rule is "Ask for approval", the action goes to the approval queue (see [User
    Approval](/security/user-approval)). If the outside-list rule is "Block", the action fails outright and the agent is
    notified that the recipient isn't allowed.
  </Accordion>

  <Accordion title="Can I set guardrails that apply to all agents?">
    Yes. Workspace administrators can disable capabilities workspace-wide in Workspace management → Capabilities. Per-agent guardrails (like allow-lists) are still configured individually.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="User Approval" icon="user-check" href="/security/user-approval">
    How the approval queue works
  </Card>

  <Card title="Attack Detection" icon="shield-halved" href="/security/attack-detection">
    Automated screening of untrusted trigger content
  </Card>

  <Card title="Agent Management" icon="sliders" href="/features/agent-management">
    Workspace-level capability controls
  </Card>
</CardGroup>


# Infrastructure & Compliance
Source: https://docs.abundly.ai/security/infrastructure

Data residency, encryption, audit trails, and regulatory compliance

Abundly is built on enterprise-grade cloud infrastructure with security and compliance at its foundation. All data is stored in EU data centers, encrypted at rest and in transit, and protected by comprehensive access controls. This page provides the technical details enterprise security teams need.

## Data residency

All customer data is stored in EU data centers, specifically in Stockholm, Sweden:

| Component         | Location                      |
| ----------------- | ----------------------------- |
| **Database**      | AWS eu-north-1 (Stockholm)    |
| **Agent Service** | GCP europe-north2 (Stockholm) |
| **File Storage**  | GCP europe-north2 (Stockholm) |
| **Task Queue**    | GCP europe-north2 (Stockholm) |

<Note>
  EU data residency ensures compliance with GDPR and other European data protection regulations. Data does not leave the EU.
</Note>

## Encryption

| Type           | Standard                                                         |
| -------------- | ---------------------------------------------------------------- |
| **At rest**    | AES-256 encryption for all stored data                           |
| **In transit** | TLS 1.2+ / HTTPS for all communication                           |
| **Secrets**    | RSA-OAEP with SHA-256 (see [Credentials](/security/credentials)) |

All communication between components is encrypted. HTTPS is enforced for all endpoints.

## Web and API transport hardening

In addition to TLS encryption, Abundly uses security headers across the web portal and API surfaces to reduce browser-based attack risk.

| Surface               | Header strategy                                                                                                                                     | Security goal                                                                          |
| --------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- |
| **Web portal**        | Strict browser headers including Content Security Policy (CSP), frame protections, permissions policy, referrer policy, and MIME-type protections   | Reduce XSS and clickjacking risk, restrict browser features, and limit data leakage    |
| **User app renderer** | Served from a separate `usercontent` subdomain with its own CSP profile that allows the embedding and external assets needed by user-generated apps | Isolate untrusted user-generated content from the main app via cross-origin sandboxing |
| **Agent service API** | API-focused hardening headers (HSTS and related Helmet protections) for non-HTML responses                                                          | Enforce HTTPS and protect API consumers from common browser-side header attacks        |

<Note>
  CSP is applied on web responses where script and content execution rules matter. API endpoints focus on transport and response hardening headers.
</Note>

### User-generated content isolation

[Interactive Apps](/features/apps) and HTML documents created by agents can contain arbitrary code. To prevent this code from accessing the main application or its session, the platform serves all rendered user content from a dedicated `usercontent.abundly.ai` subdomain that is separate from the main `app.abundly.ai` domain.

Because the browser treats these as different origins, code running inside an embedded app is fully sandboxed by the browser's same-origin policy: it cannot read cookies, local storage, or the DOM of the main app, even when displayed inside it. The main app domain refuses to serve the renderer, and the usercontent domain refuses to serve anything else, so the boundary cannot be bypassed by linking or redirecting.

## Compliance status

| Standard            | Status        | Notes                                     |
| ------------------- | ------------- | ----------------------------------------- |
| **GDPR**            | Compliant     | EU data residency in Stockholm            |
| **Data Encryption** | Met           | AES-256 at rest, TLS 1.2+ in transit      |
| **Access Controls** | Comprehensive | Role-based permissions                    |
| **ISO 27001**       | Certified     | Information security management certified |
| **SOC 2 Type II**   | Planned       | On our roadmap                            |

## Audit trails

Every agent action is logged with complete context:

| Field         | Description                                             |
| ------------- | ------------------------------------------------------- |
| **Timestamp** | When the action occurred                                |
| **Actor**     | Which agent performed the action                        |
| **Trigger**   | What initiated the action (email, schedule, chat, etc.) |
| **Plan**      | What the agent intended to do                           |
| **Execution** | What tools were called and what happened                |
| **Result**    | The outcome of the action                               |

<Note>
  Audit logs are immutable and cannot be modified or deleted after creation.
</Note>

You can access audit information through:

* **Activity log** — Real-time and historical view of agent actions with full details
* **Agent diary** — High-level summary of what each agent has been doing
* **Action history** — Append-only log of workspace configuration changes (agents, teams, secrets, API capabilities, workspace settings), scoped to workspace or team admins. See [Usage & Limits](/features/usage#action-history).

See [Activity Monitoring](/features/monitoring) for details on using these tools.

## Data retention

| Data Type               | Retention                                                           |
| ----------------------- | ------------------------------------------------------------------- |
| **Account Information** | While account active; deleted within 30 days of account deletion    |
| **User Content**        | Retained to provide services; deleted within 30 days after deletion |
| **Log Data**            | Up to 90 days for security and troubleshooting                      |
| **Usage Data**          | Anonymized data may be retained for analytics                       |
| **Backups**             | Data may remain in backups up to 9 months                           |

## System architecture

The platform is built on cloud-native infrastructure:

| Component         | Technology                | Location            |
| ----------------- | ------------------------- | ------------------- |
| **Web Portal**    | Vercel, Next.js/React     | Vercel Edge Network |
| **Agent Service** | Google Cloud Run, Node.js | GCP europe-north2   |
| **Database**      | MongoDB Atlas             | AWS eu-north-1      |
| **Task Queue**    | Google Cloud Tasks        | GCP europe-north2   |
| **File Storage**  | Google Cloud Storage      | GCP europe-north2   |

## Availability and disaster recovery

| Feature                     | Implementation                         |
| --------------------------- | -------------------------------------- |
| **Cloud-native resilience** | Automatic failover via Vercel and GCP  |
| **Automated backups**       | Daily with point-in-time recovery      |
| **Backup retention**        | 9 months with cross-region replication |
| **Health monitoring**       | Continuous monitoring with alerting    |
| **RTO (Recovery Time)**     | 24 hours for critical services         |

We maintain a documented Disaster Recovery Plan covering database recovery, secrets restoration, and service continuity.

## Security monitoring

| Activity                      | Approach                                       |
| ----------------------------- | ---------------------------------------------- |
| **Infrastructure monitoring** | Automated via GCP, MongoDB Atlas, Vercel       |
| **Alert configuration**       | Email notifications to technical team          |
| **Active monitoring**         | Daily dashboard review                         |
| **Incident response**         | Critical alerts within 1 hour (business hours) |
| **Penetration testing**       | Annual third-party testing                     |

## Legal documentation

<CardGroup>
  <Card title="Privacy Policy" icon="shield" href="https://www.abundly.ai/privacy">
    How we collect and use data
  </Card>

  <Card title="Terms of Service" icon="file-contract" href="https://www.abundly.ai/terms">
    Service agreement and terms
  </Card>

  <Card title="Data Processing Agreement" icon="file-signature" href="https://www.abundly.ai/dpa">
    DPA for enterprise customers
  </Card>

  <Card title="Sub-processors" icon="building" href="https://www.abundly.ai/subprocessors">
    List of third-party data processors
  </Card>
</CardGroup>

## Enterprise compliance

For enterprise customers, we can provide:

* Custom data retention policies
* Dedicated compliance documentation
* Audit support and reports
* Custom DPA terms

<Card title="Contact Us" icon="envelope" href="mailto:support@abundly.ai">
  Need custom compliance arrangements? Contact our team.
</Card>


# Security Overview
Source: https://docs.abundly.ai/security/overview

Multi-layer security architecture for enterprise AI agents

## Security at a glance

|                    |                                       |
| ------------------ | ------------------------------------- |
| **Data residency** | EU only (Stockholm, Sweden)           |
| **Encryption**     | AES-256 at rest, TLS 1.2+ in transit  |
| **Compliance**     | ISO 27001 certified, GDPR compliant   |
| **Credentials**    | Never exposed to LLMs                 |
| **Audit trails**   | Complete logging of all agent actions |

For full details, see [Infrastructure & Compliance](/security/infrastructure).

***

Abundly Agents are autonomous by nature. If you allow them, they can make decisions and take actions on your behalf. This is very useful, but creates security considerations that go beyond traditional software. An agent with access to email and customer data could, if poorly configured, leak sensitive information. An agent stuck in a loop could consume credits rapidly. An agent with broad capabilities needs more oversight than one with narrow, read-only access.

To manage this risk, the platform provides multiple layers of protection.

<Tip>
  Security is an important part of the [agent design](/guides/agent-design) process. This is similar to if you hire an intern. Does the intern really need access to the company bank account? If they do need access to sensitive data AND email, how do you ensure they won't accidentally leak secrets? See [Risk & Autonomy](/security/risk-autonomy) for guidance on making these decisions.
</Tip>

## Multi-layer security architecture

Security is provided through multiple layers, each working together building upon each other to provide a comprehensive umbrella of protection.

<Frame>
  <img alt="Security layers showing User, Agent, Platform, and LLM as concentric circles" />
</Frame>

### LLM layer (core)

Abundly uses Claude from [Anthropic](/integrations/anthropic) by default—an industry leader in AI safety. Claude is trained using Constitutional AI, where the model learns to follow explicit principles for helpful, honest, and harmless behavior. Before each release, Anthropic runs extensive safety evaluations including red teaming and testing with external partners like the US and UK AI Safety Institutes. Claude refuses unsafe requests, resists manipulation, and explains its reasoning when declining.

You can [switch to other models](/features/model-selection) from OpenAI or Google—all major providers have safety measures—but Anthropic leads in this space and is our recommended default.

<Note>
  Want to learn more about Anthropic's safety approach? See their [Responsible Scaling Policy](https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy), [research publications](https://www.anthropic.com/research), and [trust center](https://trust.anthropic.com/).
</Note>

### Platform layer

The Abundly platform uses the LLM as the "brain" of the agent. On top of the basic security provided by the LLM, the platform code provides a comprehensive set of additional security measures.

* An internal system prompt optimized to make the LLM function as a responsible autonomous AI agent. This provides important context and guidelines that has a big impact on behavior of the agent, helping them understand their role, abilities, limitations, and risks.
* [Attack detection](/security/attack-detection) screens untrusted trigger content (like incoming emails) for manipulation attempts before the agent acts on them.
* [Guardrails](/security/guardrails) enforce constraints that can't be bypassed by prompts, for example whitelists for email addresses or phone numbers.
* [Access controls](/security/access-control) limit who can do what, for example only allowing certain users to access certain agents.

### Agent layer

On top of the platform, the agent itself is also a security layer. It is limited to its [instructions](/features/instructions) and [capabilities](/features/capabilities). An agent without email access can't send emails, no matter what it's asked to do.

### User layer (outer)

The final layer is the user - you! You decide which capabilities each agent has, configure guardrails, require [approval](/security/user-approval) for sensitive actions, and monitor what agents do. The platform gives you tools; you decide how to use them.

A key part of the agent design process is decide the scope of the agent's job, which tools it should have access to, and how much autonomy it should have. See [Risk & Autonomy](/security/risk-autonomy) for guidance on making these decisions.

## Key security features

<CardGroup>
  <Card title="Attack Detection" icon="shield-halved" href="/security/attack-detection">
    Automated screening of untrusted trigger content before agent execution.
  </Card>

  <Card title="Guardrails" icon="fence" href="/security/guardrails">
    Technical constraints enforced by code—whitelists, limits, and restrictions that can't be bypassed.
  </Card>

  <Card title="User Approval" icon="user-check" href="/security/user-approval">
    Require human approval before agents execute sensitive actions like sending emails.
  </Card>

  <Card title="Access Control" icon="lock" href="/security/access-control">
    Role-based permissions for users, agents, and agent-to-agent communication.
  </Card>

  <Card title="Credential Security" icon="key" href="/security/credentials">
    Encrypted storage of secrets and API keys—never exposed to the LLM.
  </Card>

  <Card title="Audit Trails" icon="clipboard-list" href="/security/infrastructure">
    Complete logging of all agent activities for monitoring and compliance.
  </Card>

  <Card title="Usage Limits" icon="gauge-high" href="/features/usage">
    Set daily credit limits per agent to prevent runaway costs.
  </Card>
</CardGroup>

## Infrastructure & compliance

The platform is GDPR-compliant with all data stored in EU data centers (Stockholm). All data is encrypted at rest (AES-256) and in transit (TLS 1.2+). See [Infrastructure & Compliance](/security/infrastructure) for details on data residency, encryption, audit trails, and certification status.

## FAQ

<AccordionGroup>
  <Accordion title="Is Abundly safe for production use?">
    Yes, if you apply basic security design principles. The platform provides multiple layers of protection—from LLM training to platform guardrails to user oversight—but you need to configure them appropriately. A customer-facing agent needs more guardrails than an internal research assistant. See [Risk & Autonomy](/security/risk-autonomy) for guidance.
  </Accordion>

  <Accordion title="Can agents be manipulated through prompt injection?">
    Yes, but only if badly designed. An agent with broad capabilities and no guardrails could potentially be tricked into harmful actions. That's why the platform includes multiple defenses: [attack detection](/security/attack-detection) screens untrusted trigger content in a separate context before the agent acts on it, and [guardrails](/security/guardrails) are enforced by platform code rather than LLM reasoning, so they can't be bypassed by clever prompts. The key is matching your security configuration to your risk profile—see [Risk & Autonomy](/security/risk-autonomy).
  </Accordion>
</AccordionGroup>


# Risk & Autonomy
Source: https://docs.abundly.ai/security/risk-autonomy

Understanding and managing risk based on your agent's role

Not all agents carry the same risk. An internal research assistant with read-only access is fundamentally different from a customer-facing agent that can send emails and modify databases. Understanding this spectrum—and where your agent falls on it—is the first step to building something that's both useful and safe.

<Frame>
  <img alt="Risk vs Utility Tradeoff showing that broader scope and more tool access increase both utility and risk" />
</Frame>

## Security starts at design time

During the [agent design process](/guides/agent-design), think through:

* What capabilities does this agent need?
* What data does it need access to?
* Who will it communicate with?
* What could go wrong, and how bad would it be?

These questions shape everything: the agent's instructions, its capabilities, the guardrails you configure, and how much human oversight is needed.

<CardGroup>
  <Card title="Agent Design Canvas" icon="object-group" href="https://www.abundly.ai/agent-design-canvas">
    A structured framework for designing agents that actually work—including security considerations.
  </Card>

  <Card title="Agent Design Training" icon="graduation-cap" href="https://www.abundly.ai/courses">
    Courses on designing, building, and securing effective AI agents.
  </Card>
</CardGroup>

## The intern metaphor

Agent design can be compared to hiring a new intern.

**Does the intern need access to the company bank account?** Probably not. Most jobs don't require it, and removing that access entirely eliminates a whole category of risk.

**What if they need access to customer data AND email?** Now you have a potential leak vector. The intern could accidentally (or be manipulated into) sharing sensitive information externally. You need to think about:

1. **Do they really need both?** Can you separate the tasks so one role has data access and another handles email?
2. **If yes, what safeguards?** Maybe require approval for external emails. Maybe restrict email to specific domains. Maybe monitor closely at first.

The same logic applies to agents. The platform gives you tools to manage these risks—[guardrails](/security/guardrails), [approval requirements](/security/user-approval), [monitoring](/features/monitoring)—but you need to decide which ones are appropriate for each agent.

## Risk vs Reward

| Factor            | Lower Risk            | Higher Risk                  |
| ----------------- | --------------------- | ---------------------------- |
| **Scope**         | Specific, narrow task | Broad responsibilities       |
| **Tools**         | Read-only access      | Write, send, call, delete    |
| **Communication** | Internal team only    | Customer-facing, public      |
| **Data access**   | Public information    | Confidential, PII, financial |
| **Reversibility** | Easy to undo          | Permanent or hard to reverse |

An agent that scores "lower risk" across all factors needs minimal guardrails. An agent that scores "higher risk" on several factors may be more capable and valuable, but needs serious attention to security configuration.

## Examples across the spectrum

### Low-risk: Ticket priority agent

Imagine an agent whose only job is to set the priority field on incoming support tickets. That's it—no email, no customer data, no external communication. Just read the ticket and set a priority.

What's the worst that can happen? A ticket gets the wrong priority. Not the end of the world. This agent is safe essentially out of the box, just by nature of its limited scope and tools. Minimal guardrails needed.

### High-risk: Customer inquiry handler

Now imagine an agent that handles incoming customer inquiries. It has access to public-facing email, a database with customer information, and can reach out to people and make decisions with real impact.

This could be a VERY valuable agent—but it takes more work to make it safe:

* **[Better models](/features/model-selection)** — Use the best available (usually also the most expensive)
* **[Better instructions](/features/instructions)** — Spend more time crafting and iterating on clear guidelines
* **More testing** — Test edge cases and adversarial scenarios thoroughly
* **[Guardrails](/security/guardrails)** — Code-enforced constraints like whitelists that can't be bypassed
* **[Human-in-the-loop](/security/user-approval)** — Require approval for important decisions
* **[More monitoring](/features/monitoring)** — Watch closely, especially early on
* **Gradual autonomy** — Start with limited scope (like a trainee), expand as it proves itself

Higher reward, but more work to manage the risk.

### The reality: Most agents are in between

Most agents fall somewhere between these extremes. The most crucial decisions are usually best left to humans, while the agent handles the grunt work—the boring, repetitive tasks—and provides data and insights the human can use.

## Who decides?

A crucial part of agent design is deciding who makes decisions—the agent or the human. This is a spectrum:

<Frame>
  <img alt="Decision spectrum from Agent in control to Human in control, showing four modes: Agent decides, Agent decides and informs human, Agent suggests and waits for approval, Agent asks human to decide" />
</Frame>

| Mode                                   | When to use                                             | Example                              |
| -------------------------------------- | ------------------------------------------------------- | ------------------------------------ |
| **Agent decides**                      | Low-stakes, reversible actions where speed matters      | Setting ticket priority              |
| **Agent decides, informs human**       | Medium-stakes actions where visibility is important     | Sending internal status updates      |
| **Agent suggests, waits for approval** | Higher-stakes decisions where human judgment adds value | Issuing a customer refund            |
| **Agent asks human to decide**         | Critical decisions, edge cases, ambiguous situations    | Escalating a complaint to leadership |

You can adjust this as you gain confidence in the agent. Start with more human oversight, then gradually give the agent more autonomy as it proves itself.

## Guiding principles

* **Principle of Least Privilege** — Give agents only the capabilities and data they need for their job—no more. An agent that doesn't have email access can't accidentally send a bad email.

* **Principle of Earned Trust** — Start narrow and expand as the agent proves itself. Begin with approval requirements, then relax them once you're confident in the agent's behavior.

<Note>
  Advanced agents with broad scope CAN be safe—it just requires more careful design and ongoing attention. Don't be afraid of powerful agents; be thoughtful about how you build them.
</Note>

## Learn more

<CardGroup>
  <Card title="Guardrails" icon="fence" href="/security/guardrails">
    Configure whitelists, limits, and code-enforced constraints.
  </Card>

  <Card title="User Approval" icon="user-check" href="/security/user-approval">
    Require human approval for sensitive actions.
  </Card>

  <Card title="Monitoring" icon="chart-line" href="/features/monitoring">
    Track agent activity through diary and activity logs.
  </Card>
</CardGroup>


# Subprocessors
Source: https://docs.abundly.ai/security/subprocessors

Third parties that may process customer data when using Abundly agents and integrations.

Customer data can be shared with subprocessors depending on which capabilities, AI models, and integrations are enabled in a workspace.

* **Workspace review** — Workspace admins can open **Workspace management → Data processors** in the Abundly portal to see the list of subprocessors, as well as authorize or object to the use of a subprocessor.

* **Infrastructure** — Hosting and core platform providers are described in [Infrastructure](/security/infrastructure).

* **Legal** — Your Data Processing Agreement (DPA) with Abundly and any workspace-specific terms govern how subprocessors may be introduced or restricted.


# User Approval
Source: https://docs.abundly.ai/security/user-approval

Require human approval before agents execute sensitive actions

User approval puts a human checkpoint before sensitive agent actions. When approval is required, the agent pauses and waits for you to review and approve (or reject) the action before it executes.

<Frame>
  <img alt="Email approval request widget showing the pending email with Approve and Reject buttons" />
</Frame>

## When to use approval

Approval is most valuable when:

* **External communication** — Emails, SMS, or calls to people outside your organization
* **High-stakes actions** — Anything that would be difficult or embarrassing to undo
* **Early in agent lifecycle** — While you're building confidence in the agent's behavior
* **Sensitive data access** — Actions involving confidential information

As you gain confidence in an agent, you can relax approval requirements — moving from "ask before every action" to "ask only when reaching new recipients" to "no approval needed."

## What actions can require approval

Approval is configured per capability in the agent's capability settings.

Communication capabilities (Send Email, Send SMS, Make Phone Call) let you restrict the agent to an allow-list of recipients, then choose what happens for recipients inside and outside the list independently:

* **Inside the allowed list** — send straight through, or ask for your approval
* **Outside the allowed list** — block, or ask for your approval

A common setup is "send to known recipients without asking, ask before reaching anyone else." If you don't restrict the agent at all, you instead get a single "ask before every action" toggle — useful when you want broad reach but a human in the loop on every send.

Other capabilities (HTTP requests, custom tools) typically expose a simple "require approval" toggle. See [Guardrails](/security/guardrails) for the full list of capabilities and their configuration options.

## Where approval requests appear

### In chat

When an agent needs approval during a conversation, an approval widget appears inline in the chat. You'll see:

* The action the agent wants to take (e.g., "Send email")
* Details of the action (recipient, subject, content)
* **Approve** and **Reject** buttons
* Status badge showing "Pending"

Click Approve to execute the action, or Reject to cancel it.

### Activity page

All approval requests—from chat and from background triggers—are listed on the agent's Activity page under the **Approvals** tab.

This is where you'll see approval requests that happened outside of a chat conversation, such as:

* An email trigger that needs approval to respond
* A scheduled task that wants to send a report

<Note>
  When an approval is created from a non-chat trigger (scheduled task, incoming email, webhook, etc.), the agent's owners get an in-app notification in the bell menu — and an email too if you've enabled that channel in your notification preferences. Approvals created during a chat conversation don't trigger a notification, since you already see the approval card in the chat. If the agent has no owner, the notification escalates to team admins or workspace admins as a backup.
</Note>

## The approval flow

Here's what happens when an agent attempts an action that requires approval:

1. **Agent calls tool** — The agent tries to send an email (or SMS, or make a call, etc.)
2. **Platform checks settings** — Is approval required for this capability? Is the recipient on the allow-list?
3. **Approval record created** — If approval is needed, a pending approval record is created
4. **Agent pauses** — The agent reports that the action is awaiting approval
5. **You review** — In chat or on the Activity page, you see the pending action
6. **You decide** — Optionally tick **Notify agent**, then click Approve or Reject
7. **Action executes (or doesn't)** — Approved actions execute immediately; rejected actions are cancelled
8. **Agent notified (optional)** — If you ticked Notify agent, the agent learns the outcome and can take follow-up actions; otherwise it stays in the dark

## Notifying the agent of your decision

Each approval card has a **Notify agent** checkbox next to the Approve and Reject buttons. Tick it before clicking, and the agent learns the outcome and can take follow-up steps — for example updating a CRM after a sent email, or trying a different approach when you reject a call. Leave it unchecked, and the action either executes or doesn't and that's the end of it.

<Frame>
  <img alt="Approval card on the Approvals tab with the Notify agent checkbox highlighted and its tooltip visible" />
</Frame>

The checkbox is off by default, so you decide per decision whether the agent needs to react. Inside a chat the checkbox is hidden — the agent is always notified there so the conversation can continue.

## Configuring approval

Configure approval in each capability's settings. Open the agent, go to **Capabilities**, and find the capability you want to configure. Choose whether to restrict recipients to an allow-list, then set the inside-list and outside-list rules.

See [Guardrails](/security/guardrails) for step-by-step instructions and details on configuring allow-lists.

## Best practices

**Start with approval, then relax.** When deploying a new agent, require approval for external communication. Once you've seen the agent handle a variety of situations correctly, you can switch to "ask only outside the allowed list" or disable approval entirely.

**Combine an allow-list with "ask for approval" outside it.** This gives you the best of both worlds: the agent communicates freely with known contacts, and you review anything new.

**Watch the bell menu.** Approvals created by background triggers fire an in-app notification (and an email, if you've opted in) to the agent's owners. The Activity page is still where you'll see the full list of pending approvals.

**Review before approving.** Take a moment to actually read what the agent wants to do. The approval flow is only useful if you're reviewing, not just clicking Approve.

## FAQ

<AccordionGroup>
  <Accordion title="Can I approve multiple actions at once?">
    Not currently. Each approval request must be handled individually. This is intentional—batch approval would defeat the purpose of human review.
  </Accordion>

  <Accordion title="Do approvals expire?">
    No. Approval requests remain pending until you explicitly approve or reject them. If you don't want a pending approval, reject it.
  </Accordion>

  <Accordion title="Will I be notified of pending approvals?">
    Yes, for approvals created by background triggers (scheduled tasks, incoming emails, webhooks, etc.). Agent owners receive an in-app notification in the bell menu, and an email too if email notifications are enabled in your account settings. If the agent has no owner, the notification escalates to team admins or workspace admins. Approvals created during a chat conversation don't trigger a notification — you already see the approval card in the chat.
  </Accordion>

  <Accordion title="What happens if I reject an action?">
    The action is cancelled and won't execute. If you ticked **Notify agent** before rejecting, the agent learns about the rejection and can adjust; otherwise the action just doesn't happen and the agent isn't told. If the agent was actively waiting for the action to complete (e.g. in a chat), it always receives a failure response.
  </Accordion>
</AccordionGroup>

## Learn more

<CardGroup>
  <Card title="Guardrails" icon="fence" href="/security/guardrails">
    Configure allow-lists and approval requirements
  </Card>

  <Card title="Monitoring" icon="chart-line" href="/features/monitoring">
    Track agent activity and review what happened
  </Card>

  <Card title="Risk & Autonomy" icon="scale-balanced" href="/security/risk-autonomy">
    Decide when approval is appropriate for your use case
  </Card>
</CardGroup>


# Platform summary
Source: https://docs.abundly.ai/summary/index

A comprehensive single-page summary of all Abundly documentation — ideal for LLM context

This page is a condensed summary of the entire Abundly documentation. It covers what the platform does, how agents work, available features and integrations, security architecture, and practical guides.

For full details on any topic, follow the links to the relevant documentation pages.

## Introduction

### Welcome

Abundly is an enterprise-grade platform to build, configure, manage, and collaborate with autonomous AI agents — an operating system where teams of humans and AI agents work side by side. Unlike simple tool-enabled AI chat, Abundly agents offer true autonomy (schedules, triggers, 24/7 operation), multi-modal communication (SMS, email, Slack, voice calls), conversational configuration, multi-agent collaboration, dynamic apps and scripts, and enterprise security and governance. Example use cases include cross-platform automation ("dig through last week's Slack messages and create a Notion overview doc"), scheduled monitoring ("every morning, check my HubSpot todos and ping me on Slack if something is urgent"), and bulk data processing across spreadsheets and integrations.

[Full page →](/introduction/welcome)

### What is an AI agent?

An Abundly agent is a multilingual, autonomous digital colleague — faster and cheaper than humans, more intelligent and flexible than code, sitting in between. Agents are designed to augment humans by handling routine knowledge work that involves fuzzy inputs, uncertainty, and judgment, freeing people for strategic tasks. Every agent has four components: an **LLM** (Claude, GPT, Gemini, or others) as its brain, a **Mission** (natural-language instructions that are versioned and can be self-updated), **Tools** (grouped into capabilities like Slack, Gmail, code execution, web search, and 40+ integrations), and **Autonomy** (schedules, triggers, and proactive communication). Benefits over manual work include lower cost, higher speed, better control through written instructions, and higher consistency.

[Full page →](/introduction/what-is-an-agent)

### Getting started

Sign up at app.abundly.ai for a free trial with credits — no credit card required. Create an agent from scratch or clone one from the Agent Catalogue. Configure your agent by chatting with it (it writes its own instructions from the conversation), enable the capabilities it needs, then test and refine.

[Full page →](/introduction/getting-started)

### Platform overview

The **Team page** is your home base, showing all agents as cards with access to Members, Agent Management, and Usage & Limits in the sidebar. The **Agent page** is where you configure and interact with an individual agent — it includes Instructions, Docs, Capabilities, Activity, and Settings (covering communication, usage limits, agent-to-agent communication, user access, API endpoint, model selection, and more). The chat interface supports text, voice input, file uploads, Walk & Talk mode, fact checking, and a thinking toggle for more deliberate reasoning.

[Full page →](/introduction/platform-overview)

### Pricing

Abundly uses a credit-based pricing model: you pay for credits consumed when agents perform actions. You can create unlimited agents, while workspace member seats depend on your plan — see [app.abundly.ai/prices](https://app.abundly.ai/prices) for current plans and credit allotments. Each subscription includes a monthly credit allotment that refills at the start of every billing period and doesn't carry over, while top-up credits stay on your account until used. You can set per-agent spending limits and track usage through the Usage Reports dashboard.

[Full page →](/introduction/pricing)

## Features

### Agent Instructions

Instructions are your agent's job description in natural language — always top of mind, included in every chat, trigger, and scheduled run. You can edit them directly, or chat with the agent and ask it to draft and update its own instructions; the **Update Instructions** capability is on by default with chat-only and edit-access guardrails. Every change is versioned automatically, so you can compare versions, see who made each change (you or the agent), approve or undo individual changes, and revert to any earlier version.

[Full page →](/features/instructions)

### Configurable Capabilities

A capability is a feature or integration the agent can use — platform features like document editing or image generation, external integrations like Slack or Google Drive, custom API capabilities, or MCP servers. Each agent starts with a basic set, and you add more as needed; tabs on the capabilities page split them into **Built-in**, **API**, and **MCP**. Start with the minimum: too few and the agent can't do its job, too many and it gets confused about which to use. If the agent needs something that isn't enabled, it surfaces a capability card on demand so you can toggle it on without leaving the conversation.

[Full page →](/features/capabilities)

### Multi-channel Communication

Your agent isn't stuck in a chat window — it can talk to you through chat, email, SMS, voice calls, Slack, Teams, and other collaboration tools, and you decide which channels are enabled. Chat is always on; email is built-in and toggleable; SMS and voice calls are in beta and require contacting support. Beyond text, agents work with images, audio (including transcription and voiceovers via [ElevenLabs](/integrations/elevenlabs) and [OpenAI](/integrations/openai)), documents, and diagrams — useful for daily briefings, audio email summaries, or analyzing photos a customer sends in. If multiple channels are enabled, guide the agent in its instructions on which to prefer ("only call me for urgent issues").

[Full page →](/features/communication)

### Advanced Chat

Each agent has a feature-rich chat similar to ChatGPT or Claude, but treated as just one of many possible communication channels. Each conversation is its own local context: previous messages plus the agent's instructions, documents, and capabilities — anything you want available outside that chat must be saved to instructions or an agent document (diary entries are the exception). The chat supports file uploads, voice input, text-to-speech, edits to sent messages, model switching mid-conversation, multi-user collaboration with live-streamed responses, starred and searchable conversation history, and a context indicator that suggests starting a **New chat with summary** when conversations get long. Agents can read, create, and update [agent documents](/features/documents) and [databases](/features/agent-databases) from chat, render rich responses (formatted text, diagrams, images, voiceovers), generate [interactive apps](/features/apps) on the fly, and run code when the **Code Execution** capability is enabled.

[Full page →](/features/chat)

### Email & SMS

Every agent gets its own email address (e.g. `myagent.myworkspace@agent.abundly.ai`) — it doesn't access your inbox unless you explicitly enable that. Toggle **Send Email** and **Receive Email** in capabilities; SMS works similarly but is hidden behind support enablement and needs a dedicated phone number for inbound. Agents handle attachments, CC/BCC, replies, forwards, and rich HTML formatting much like a human would. Guardrails let you restrict the agent to an allow-list of addresses or whole domains, then independently choose what happens for recipients inside the list (send, or ask for approval) and outside it (block, or ask for approval) — approval requests appear in chat or in **Activity** for triggers, where each card has a **Notify agent** checkbox to optionally tell the agent the outcome.

[Full page →](/features/messaging)

### Voice Communication

Voice features let you talk to your agent and have it respond in audio across the chat and on phone calls. The microphone button transcribes speech to text in most languages (with editable transcripts before sending), the speak button reads any message aloud, and **Walk & Talk** mode is a mostly hands-free chat variant for walking, driving, or lunch — preserving the agent's full intelligence by routing through the normal text pipeline rather than a real-time voice model. Under **Settings → Basic Information → Voice** you pick the agent's voice (OpenAI built-in, plus [ElevenLabs](/integrations/elevenlabs) premade, professional, and cloned voices when configured). Phone calls (beta) let anyone reach your agent — useful for the **receptionist pattern** where the call captures the request and the agent does the heavy lifting after with a "phone call ended" trigger; settings give per-call control over capabilities, custom instructions, voice and turn detection, transcription tuning, escalation transfer numbers, whitelists, and approval.

[Full page →](/features/voice)

### Agent Documents

Each agent has its own file repository that works out of the box, with both you and the agent able to create, read, and edit documents under full version history. Supported types include Markdown and code (native), SVG (editable text rendered as graphics), PDFs, Word, PowerPoint, and spreadsheets (transcribed to text with originals preserved), images (vision-described), and audio/video (transcribed). Each document has a visibility level — **Full** (always in context), **Summary** (default — agent knows it exists, reads on demand), or **Hidden** — and a separate **Publishing** control for public links. Documents support folder organization, precise edits, exact/full-text/semantic search, **Chat about this** to start a focused conversation, downloads (text, `.docx`, originals, ZIP folders), and publishing (with read/create/update/delete permissions for databases and apps). Agent documents are stored in EU data centers in Stockholm; external documents (Google Drive, SharePoint, etc.) stay on the original provider's servers.

[Full page →](/features/documents)

### Agent Databases

Agent databases (also called data documents) are special [agent documents](/features/documents) that store records as JSON, similar to a lightweight MongoDB collection — instant to create with no setup ("create a database to track processed invoices"). Use them when you have many items of the same type (invoices, tickets, products); use a text document for prose. The agent inserts, updates, deletes, and queries records via natural language, supports **direct queries** (exact field matching) and **semantic search** powered by RAG (find feedback by meaning, not just keywords), and pairs naturally with [interactive apps](/features/apps) and [scripts](/features/scripts) for dashboards, forms, and automation. Schemas are defined automatically in TypeScript and editable; you can import/export JSON and publish databases with read/create/update/delete permissions. For larger or shared data, connect external databases (PostgreSQL, MongoDB, etc.) via [MCP](/integrations/mcp) — agent databases handle hundreds of records and include RAG, while external databases scale to millions, support joins, and serve multiple systems.

[Full page →](/features/agent-databases)

### Interactive Apps

Apps are [agent documents](/features/documents) containing executable code that renders as a UI — dashboards, forms, calculators, data browsers — that you can build by asking ("create a dashboard for our OKRs"). Two types are supported: **React apps** with Chakra UI, Recharts, and Lucide icons built in, and **HTML apps** for custom layouts, 3D, or maps; both run sandboxed and can read/write [agent databases](/features/agent-databases), load other documents, pull libraries from public CDNs, and call a curated set of agent tools (HTTP requests via API capabilities, BigQuery, Snowflake, Notion, etc.). Larger apps can be split into multi-file folders with an `index.html`/`index.jsx`/`index.tsx` entry file. Apps live as **chat documents** (scoped to the conversation, easy to iterate) until you promote them to **agent documents** for reuse and publishing. Published apps share data via published databases; apps that call agent tools work in the agent chat but not on public links.

[Full page →](/features/apps)

### Code Execution

When **Code Execution** is enabled, the agent can write and run JavaScript on the fly in the chat or as saved **scripts** (which are [agent documents](/features/documents) with version history). Code is faster, cheaper, and more reliable than token-by-token reasoning for large data, precise calculations, format conversions, and batch operations — agents still use intelligence to write the script, then let the script do the heavy lifting. Scripts have access to all the agent's capabilities (web search, email, [agent databases](/features/agent-databases), external APIs), so you can build workflows like weekly customer enrichment or competitive monitoring. Scripts can also be linked directly to triggers — scheduled tasks, webhooks, HTTP API endpoints, and MCP tools — bypassing the LLM entirely for fast, deterministic execution; if part of the work needs reasoning, the script can hand off to the full agent by returning `{ escalate: true, message: "..." }`.

[Full page →](/features/scripts)

### Virtual Machine

The Virtual Machine capability gives your agent an ephemeral Linux VM with bash, Python, and Node.js pre-installed — ideal when in-chat code execution isn't enough. The agent spins up a fresh VM (`create_vm`), runs commands (`execute_bash`), transfers files in and out (from chat attachments or [agent documents](/features/documents)), and destroys the VM when done; the same VM is reused across steps so disk state and installed packages persist. Use it for Office documents (`.xlsx`, `.pptx`, `.docx`, `.pdf` via openpyxl, python-pptx, python-docx, pypdf, pandoc), large CSV/JSON/Parquet processing with pandas, package installs (pip, npm, apt, headless browsers like Playwright), and longer-running multi-step jobs.

[Full page →](/features/virtual-machine)

### Scheduled Tasks

Scheduled tasks let your agent work proactively by waking itself up at specific times. **Recurring tasks** run on any pattern (daily, weekly, monthly) and live in the agent's [instructions](/features/instructions), so saving instructions creates or updates them automatically. **One-off tasks** are scheduled through chat and disappear after running. From **Agent Instructions → Scheduled Tasks** you can enable/disable tasks, view details, and **Run now** to trigger immediately for testing. Tasks can be **script-linked** to bypass the LLM entirely for deterministic, high-volume work, and each task can override the agent's default model and thinking setting to balance cost and capability per job. When a task runs, the agent has access to its instructions, capabilities, the task description, and any [agent documents](/features/documents) it has access to.

[Full page →](/features/scheduling)

### Context and Memory

What your agent knows at any moment determines how it responds, so context management is fundamental. **Global context** — [instructions](/features/instructions) and [agent documents](/features/documents) (subject to visibility settings) — is always available. **Local context** is per-interaction: a chat conversation knows its own messages, a trigger sees the trigger details (e.g. 5 previous Slack messages in the channel). Memory across contexts comes from instructions and documents themselves: agents read and write their own, so you can store preferences, decisions, and accumulated knowledge there ("create a database to store everyone's preferences and check them before communicating"). For accumulating facts and deduplicating recurring information, use [agent databases](/features/agent-databases) with semantic search. The legacy Memory capability is retired for new agents but still works for agents that already have it.

[Full page →](/features/memory)

### Web Search, Scraping, and Deep Research

Three capabilities give agents access to current information beyond their training data, all built in with no API keys required. **Web Search** ([Perplexity](/integrations/perplexity)-powered) is fast and good for general lookups and current news; **Web Scraping** ([Firecrawl](/integrations/firecrawl)-powered) reads specific URLs and supports an inline extraction prompt so only the relevant content comes back, similar to delegating to a sub-agent; **Deep Research** runs comprehensive multi-source analysis over 10+ minutes and uses significantly more credits — best for high-value strategic work. Web Search and Web Scraping are on by default; Deep Research is opt-in. Combine them — search to find a page, scrape to read the live content for 100% accurate answers — and remember web search returns citation URLs you can verify, reducing hallucination risk.

[Full page →](/features/search)

### Multi-Agent Collaboration and Task Delegation

Abundly supports two patterns for agents working together. **Delegation** spins up a temporary sub-agent for a specific task (with chosen capabilities, optional document references, and even a different LLM) — primarily for **context management**: a parent agent reviewing a contract against three standards can delegate one sub-agent per standard, keeping its own context clean and supporting parallel processing. The **Delegate Task** capability is on by default and sub-agents can't delegate further. **Agent-to-agent communication** connects persistent, fully-configured agents — a Twitter expert, a reusable ICP checker — configured per agent under **Settings → Agent Communication** with **Discoverability** (No one / Team / Everyone) for inbound and explicit outbound connections; calling agents see only the target's name and description and can optionally share documents with read or read-write access. Be deliberate about agent-to-agent connections, since they increase cost and can give indirect access to sensitive data via Agent A. Before adding more agents, consider task documents — separate guidance documents the agent reads on demand. For external systems, agents can call out via [MCP](/integrations/mcp) or [HTTP](/features/http), and be called via [API access](/features/agent-api-endpoint).

[Full page →](/features/multi-agent)

### Model Selection

Each agent picks the LLM that powers it. Leaving the model on **(no preference)** lets Abundly choose the best general-purpose model for you and update the default as new ones land. You can pick model **aliases** (e.g. "Claude Sonnet Latest", recommended) that auto-upgrade after we verify them, or **specific versions** for full control. Most workspaces include Anthropic Claude (Opus, Sonnet, Haiku), OpenAI GPT (5.2 Pro, 5.2, 5 Mini, 5 Nano), Google Gemini (Pro 3, Flash 2.5, Flash Lite 2.5), and xAI Grok, with optional families like DeepSeek or self-hosted. Models can be set per agent and per **context** (chat, scheduled tasks, email, Slack, Teams, agent messages), with further per-task overrides on individual scheduled tasks. Many models support a **thinking** mode that trades extra processing time and credits for better reasoning; behavior depends on model (supports / requires / does not support thinking).

[Full page →](/features/model-selection)

### Agent Personalization

Personalization makes agents feel like real teammates rather than generic tools. Profile pictures can be auto-generated from the agent's name and description, generated from a custom prompt, uploaded, or surprised by the AI; the agent can update its own picture if you ask. Personality is shaped through [instructions](/features/instructions) — even one sentence ("be brief and to-the-point", "act like an eager intern", "communicate like the butler Jeeves") changes how responses feel, and you can vary tone by context (formal in email, brief in Slack). The **Chat Start Message** can be a fresh contextual greeting per chat or a fixed line, descriptions and tags help organize agents and feed [multi-agent communication](/features/multi-agent), and agents can update their own name, picture, and description on request. The philosophy: agents shouldn't pretend to be human, but treating them like teammates with a bit of personality usually works better.

[Full page →](/features/personalization)

### Agent Evals

Agent Evals turn "working correctly" into something you can verify automatically — define test cases once, run them after any change to catch regressions, compare models side by side, and iterate with confidence. Each eval has a **Test Prompt** (simulating a user message or scenario), an optional **Tools** override (Included, **Excluded**, or **Faked** with canned responses — useful to prevent side effects, control inputs deterministically, or speed up tests), and an optional **Grader** (LLM-based Pass/Fail or Rating 1-10, scoped to full transcript or final response only) plus optional **Validation rules** that fail fast on tool-call assertions. Run individual evals or **Run All**, configure multiple **Eval models** to compare, and pick a separate **Grader model**; results are organized by run with list and grid views, error counts separated from fails, and credits split between test and grader usage. Sub-agent calls during evals are captured as nested transcripts, agent-to-agent recursion is capped at 3 levels, and evals can be exported/imported as JSON. Turning on the **Evals capability** lets the agent create, edit, run, and review its own evals — the foundation of eval-driven development.

[Full page →](/features/evals)

### Activity Monitoring

Two built-in tools, surfaced under **Activity**, give you visibility into what your agent is doing autonomously. The **agent diary** is a high-level human-readable log of activities and reasoning — written automatically after triggers and after significant chat moments when **Write diary entries** is enabled. The **activity log** is a live technical log of every trigger the agent processes, with the trigger event, **pre-filter**, **attack detection** (prompt injection / jailbreak checks for untrusted email and SMS), and **execution** with all tool calls and credits per phase (when Token optimisation mode is on). Each entry has a **Chat about this** button for drilling in, and turning on the **Activity Log** capability lets the agent inspect its own (or connected agents') trigger history and credit usage. Workspace admins configure who receives blocked-attack alerts.

[Full page →](/features/monitoring)

### Usage & Limits

The **Usage & Limits** page gives workspace and team administrators visibility and control over credit consumption. Usage Reports show current balance, average daily use, balance history, and top-consuming agents, with CSV export across configurable date ranges, intervals, and aggregations. **Agent limits** set a workspace-wide default daily limit plus per-agent overrides — when an agent hits its limit it stops immediately and resets at UTC midnight, with yellow badges at 80% and red when exceeded; you can manually **Reset Today** to resume early. **Team monthly limits** (when [teams](/features/teams) are enabled) cap each team's monthly spend and pause every agent in the team when reached, in addition to per-agent daily limits. A sidebar usage widget shows credits used today and a progress bar (turn on **Token optimisation mode** to enable it). Both LLM tokens (input, output, reasoning) and tool usage count toward credits, regardless of trigger source.

[Full page →](/features/usage)

### Workspace Members

The Members page is where workspace admins invite users and assign roles. Every user has one of four workspace roles: **Admin** (full control over members, billing, capabilities, and all agents), **Creator** (standard access based on agent-level settings; can create agents), **Collaborator** (can use and edit shared agents, but cannot create agents or manage secrets and API capabilities), or **Guest** (limited, invitation-only access to specific agents). The role hierarchy is Admin > Creator > Collaborator > Guest. Invites accept multiple addresses at once with optional names and (when [teams](/features/teams) are enabled) direct assignment to teams, and pending invitations count toward your plan's member cap. Every workspace must keep at least one Admin, members can leave a workspace they belong to (unless they're the last admin), and you can fully delete your account from **Profile settings → Danger Zone** with a 30-day grace period. The page also shows pending invitations (with Renew/Revoke) and any active or pending **Support access** so you can review and revoke at any time.

[Full page →](/features/workspace-members)

### Teams

Teams are an enterprise feature (enabled on request) that partition a workspace's agents and users into logical units — by department, project, or however fits. Each user can belong to multiple teams with different team-level roles (Admin, Member, Guest) layered on top of their workspace role, and each team can have its own capability settings that override workspace defaults (more restrictive only — workspace-disabled capabilities can't be enabled per team). Workspace admins create and delete teams and set monthly credit limits from the Team management page; team admins manage members, capability settings, and limits for their team. Agents belong to exactly one team, or are **global agents** with no team — your workspace controls whether everyone, only admins, or nobody can create global agents. Global agents are visible only to workspace admins.

[Full page →](/features/teams)

### Agent management

The Agent Management page is a centralized table for workspace and team administrators to search, filter, sort, and update agents in bulk. Columns cover the agent (with privacy icon, model, and team badges), today's usage, last activity, last updated, credit details visibility, and a free-text business value summary; expanding a row shows tags, enabled capabilities (including admin-policy disabled state), API capabilities, MCP servers, and agent admins. Filter by status, privacy, teams, tags, model, admin, capabilities, API capabilities, MCP servers, usage range, and date ranges; bulk actions include enable, disable, move to team, set daily limit, and set model. The **Agent limits** tab lets you manage workspace defaults, per-agent overrides, and reset daily counters in one place. Capability availability policies are configured separately under **Workspace management → Capabilities** and **Team management → Capabilities**.

[Full page →](/features/agent-management)

### Agent Sharing and Cloning

Cloning duplicates an agent inside your workspace or to another workspace you belong to — useful for variations, templates, or per-team copies. You can also **Export** an agent to a JSON file (optionally including documents, evals, and API access settings) and **Import** it as a new agent — handy for backups or moving between environments. Cloning copies name, description, tags, profile picture, instructions, capabilities, model, scheduled triggers, MCP servers (without auth), and the daily limit; it never copies chat history, tool credentials, access settings, instruction history, public sharing, or daily usage counters. The person performing the clone becomes Admin of the new agent, which gets fresh default access settings. **Public sharing** publishes the agent so anyone with a link can clone it to their own workspace — visitors see the name, description, creator, and (on click) the instructions, with optional document inclusion. Clones are completely independent: updates to the original don't propagate.

[Full page →](/features/sharing)

### HTTP Requests: Connect to Any API

The HTTP capability is your agent's swiss-army-knife for talking to any online service with a REST API — well-known APIs often work straight from training data, while less familiar ones can be learned through experimentation, documentation links, web search, or your own description. It supports all standard HTTP methods (GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS) and combines well with [Web Search](/features/search) for discovering APIs, with [code execution](/features/scripts) for processing large responses, and with [delegation](/features/multi-agent) for one-pass heavy payloads. Two guardrails on the capability card: **Allow requests to any URL** (when off, the agent is locked to endpoints defined by enabled API Capabilities) and **Require Approval** (No / Yes for all / Yes except GET). For authenticated APIs, store credentials as **Secrets** in **Shared Assets → Secrets** and define an **API Capability** that injects them server-side — supporting bearer tokens, API key headers, basic auth, query parameters, custom headers, OAuth 2.0 client credentials, GCP service accounts, and signed JWTs (RS256/ES256), all scoped to specific teams or agents. The capability is hidden by default and enabled per workspace by support; ask the agent to "codify" working API calls into its instructions for fast, reliable reuse.

[Full page →](/features/http)

### API Access

The **API Access** tab in agent settings exposes your agent to external systems in four ways, all backed by the same API keys: **HTTP API** for scripts and backend services, **MCP server** for AI apps like Claude Desktop, Cursor, and n8n, **Chat widget** for embedding a floating AI chat bubble on your website, and **Webhooks** for receiving callbacks from services like Stripe, GitHub, or Trello. HTTP API and MCP authenticate with API keys (sent as `X-Agent-Access-Key` or `Authorization: Bearer`); webhooks rely on the URL itself as the secret. For each endpoint, MCP tool, or webhook integration you specify a **handler** — either a natural-language prompt to the agent or a [script](/features/scripts) that runs deterministically without invoking the LLM — plus optional model overrides. The chat widget (gated behind a workspace flag) supports allowed-origins enforcement, capability scoping, a separate daily credit limit, system prompt override, demo response override, model override, and brand styling — and stays hidden if the agent is disabled, the origin isn't allowed, or the limit is exhausted. All exposure modes log to the [activity log](/features/monitoring) and respect the agent's normal capabilities and guardrails.

[Full page →](/features/agent-api-endpoint)

## Integrations

### Integrations Overview

You give your agent access to external services like Slack, Google Drive, or GitHub by toggling capabilities on. Some capabilities ship with platform-provided credentials (e.g. Perplexity, send/receive email) and work out of the box; others require your own credentials via OAuth or an API token, with all secrets stored encrypted. If a capability isn't enabled when you ask the agent to do something, it requests it on demand by showing a capability card in the conversation. Some capabilities can be configured to require user approval before the agent acts. When no built-in integration fits, you can extend your agent with an [MCP server](/integrations/mcp) or the [HTTP capability](/features/http).

[Full page →](/integrations/overview)

### MCP Servers

MCP (Model Context Protocol) is an open standard for connecting AI applications to external tools. You give your agent access to an MCP server by pasting its URL into the MCP Servers tab — the platform auto-detects whether OAuth or an API token is needed, discovers the available tools, and lets you toggle them individually. This lets you extend your agent beyond built-in integrations, tap into the community ecosystem, or expose your own internal systems (CRM, inventory, custom databases) as agent capabilities without waiting for platform updates. Only add servers from sources you trust — MCP credentials are encrypted with hybrid AES+RSA, and OAuth tokens are never exposed to the model directly. You can also go the other way and expose your Abundly agent as an MCP server so external apps like Claude Desktop, Cursor, or n8n can call it as a tool.

[Full page →](/integrations/mcp)

### Integration pricing

Whether an integration consumes credits depends on whose credentials are being used. With user-provided credentials (Slack, Notion, GitHub, Stripe, etc.), you're already paying the provider directly, so Abundly doesn't charge for usage. With platform-provided credentials (e.g. web search via Perplexity, image generation), Abundly pays the provider and converts the cost into credits charged to your account. LLM inference also consumes credits regardless of which integrations are involved. Credits are pooled across your entire workspace, so every agent and member draws from the same balance — you can track consumption under **Usage and Limits** and set per-agent limits to cap spend.

[Full page →](/integrations/pricing)

### Atlassian

Connect your agent to Atlassian to work across Jira, Confluence, and other Atlassian Cloud products — searching issues, creating pages, and managing projects. An Atlassian organization admin allows Abundly access via the Rovo MCP Server settings once, after which each user authenticates with their own account through OAuth. Use it for issue triage, sprint reporting, documentation sync between Jira epics and Confluence, and cross-platform updates with Slack.

[Full page →](/integrations/atlassian)

### Anthropic

Anthropic provides the Claude family of language models, known for strong reasoning and safety features. Claude is available as one of the LLM "brain" options for your agent, configured under **Settings → Advanced Config**. See [Model Selection](/features/model-selection) for the full list and guidance.

[Full page →](/integrations/anthropic)

### Autodesk

Connect to Autodesk Platform Services (APS) to upload CAD files, convert DWG drawings, translate 3D models (e.g. RVT → IFC, DWG → SVF2), extract model properties and element hierarchies, and generate standalone HTML 3D viewers. With APS app credentials alone, your agent can upload, convert, translate, and view; after connecting an Autodesk account via OAuth, it can also list ACC hubs, browse project folders, and download files from Autodesk Construction Cloud. Use it to convert drawings, extract Revit door schedules, drill into mechanical systems, or pull the latest IFC from a Coordination folder.

[Full page →](/integrations/autodesk)

### BigQuery

Your agent can explore your Google BigQuery datasets and run read-only SELECT queries on your behalf, so you get revenue breakdowns, scheduled reports, and data-quality checks without writing SQL. It discovers schemas, executes queries (with a 30-second timeout and 1 GB billing limit per query), and returns results along with bytes processed so you can keep an eye on cost. Authentication uses a GCP service account JSON key with BigQuery Metadata Viewer, Data Viewer, and Job User roles, stored as team secrets.

[Full page →](/integrations/bigquery)

### Cloud Storage

Your agent can list, read, write, and delete objects in any S3-compatible bucket — AWS S3, Cloudflare R2, Google Cloud Storage (HMAC), MinIO, Backblaze B2, or any provider that speaks the S3 API. Use it to archive email attachments, run scheduled CSV exports, fetch reference data like config files, clean up old objects, or move files into Slack workflows. Downloads are capped around 5 MiB and uploads around 10 MiB per object; credentials are stored as a workspace secret.

[Full page →](/integrations/cloud-storage)

### ElevenLabs

ElevenLabs provides high-quality text-to-speech for your agent. It's one of the voice providers available in the **Text to Speech** capability — your agent can pick a specific ElevenLabs voice per request when generating audio files like newsletter voiceovers or daily audio briefings. ElevenLabs voices (premade, professional, and cloned) also appear as options for the agent's own voice used by the Speak button and [Walk & Talk mode](/features/voice#walk--talk-mode), alongside OpenAI voices.

[Full page →](/integrations/elevenlabs)

### Firecrawl

Firecrawl powers your agent's ability to read and extract content from web pages. The **Web Scraping** capability is enabled by default, letting your agent summarize articles, extract structured data like pricing from product pages, or compare content across multiple URLs. Combine it with Perplexity for deeper research workflows — Perplexity finds the URLs, Firecrawl extracts the detail.

[Full page →](/integrations/firecrawl)

### Giphy

Giphy lets your agent search and share animated GIFs. It's bundled into the **Image Generation** capability alongside AI image generation. Use it to add a celebration GIF to team announcements, illustrate sentiment in weekly reports, or drop a thumbs up into a Slack post.

[Full page →](/integrations/giphy)

### GitHub

Connect your agent to GitHub to browse files, create and review pull requests, manage issues, and react to repository events. Authentication uses a personal access token (fine-grained tokens recommended), and you can specify watched repositories so the agent reacts to webhooks for new issues, pull requests, or pushes. Common patterns include code review assistance, generating release notes from commits, automated issue triage with labels, and dependency update PRs.

[Full page →](/integrations/github)

### Google Calendar

Connect your agent to Google Calendar to view events, schedule meetings, and block focus time. Authentication uses a personal access token tied to your Google account, and after connecting you choose which calendars the token can access — broad access or least-privilege scope. The capability is currently hidden; email [support@abundly.ai](mailto:support@abundly.ai) to enable it.

[Full page →](/integrations/google-calendar)

### Google Drive

Connect Google Drive in one of two scopes: **Google Drive** (only specific files you share with the agent, plus files it creates) or **Google Drive (Full Access)** (all approved drives and folders, hidden by default). Both use a personal access token; with Full Access you pick which drives and folders to allow during OAuth. Your agent can read Google Docs, Sheets, Slides, PDFs, and common Office files; edit Docs and Sheets; and create new Docs and Sheets in your Drive. Use **Reconnect** to re-run consent if a token's access has been revoked.

[Full page →](/integrations/google-drive)

### Google Gemini

Google provides the Gemini family of language models and image generation. Gemini models are available as an alternative LLM under **Settings → Advanced Config**. Image generation uses Gemini Pro 3 (also known as NanoBanana) — toggle on the **Image generation** capability and select Google to generate infographics, cover images for daily news summaries, and other visuals.

[Full page →](/integrations/google-gemini)

### HubSpot

Connect your agent to HubSpot to read and update any data in your CRM — contacts, companies, deals, tickets, tasks, emails, and custom objects. The agent can call any HubSpot REST API endpoint your token has access to, so anything you can do in HubSpot, the agent can do. Authentication uses a private app access token with scopes you choose (you must be a HubSpot super admin to create private apps). Use it for contact lookups, deal stuck-in-stage monitoring, company research, follow-up task creation, and ticket status updates.

[Full page →](/integrations/hubspot)

### Microsoft Teams

Connect your agent to Microsoft Teams so it appears as its own named bot — users @mention it in channels, group chats, or DMs to trigger the agent directly. Each Teams-enabled agent needs its own Teams app, bot identity, and client secret, set up across the Developer Portal, Azure Portal, and Teams Admin Center. Common patterns include triggering a deployment agent with `@Releaser`, kicking off research from `@Researcher`, or scheduled standup posts to a channel. The integration is currently hidden; email [support@abundly.ai](mailto:support@abundly.ai) to enable it.

[Full page →](/integrations/msteams)

### Notion

Connect your agent to Notion to read pages, query databases, and create or update content. Authentication uses an internal integration token with the read/update/insert permissions you choose, and the agent can only access pages you explicitly share with the integration (child pages inherit access). Use it for CRM-style database updates from forwarded emails, content search, meeting notes linked from Slack, overdue task summaries, and file handoffs from Slack to Notion via [agent documents](/features/documents).

[Full page →](/integrations/notion)

### OpenAI

OpenAI provides a wide range of capabilities for your agent: GPT models as an alternative LLM, image generation via GPT, voice transcription via Whisper (mic button or audio file drop), text to speech (Speak button), and image recognition for screenshots, invoices, and charts. The OpenAI Realtime API also powers low-latency voice calls — Abundly combines it with [Twilio](/integrations/twilio) for the **Make Phone Call** and **Receive Phone Call** capabilities (currently hidden, contact support to enable).

[Full page →](/integrations/openai)

### Outlook

Connect your agent to Outlook to read, search, and draft emails in Microsoft 365. Authentication uses a personal access token tied to your Microsoft account; in many organizations a Microsoft admin must approve the Abundly Outlook app in Microsoft Entra first. By default the agent only has read and draft tools — toggle **Allow sending email** to let it send directly. Use it for inbox summaries, attachment search, daily digests posted to Slack, and SMS-driven email follow-ups. The integration is currently hidden; email [support@abundly.ai](mailto:support@abundly.ai) to enable it.

[Full page →](/integrations/outlook)

### Perplexity

Perplexity powers your agent's web search and deep research. It's a platform-provided integration with no setup — **Web Search** (enabled by default) returns answers in seconds for quick facts and lookups, while **Deep Research** takes longer but produces comprehensive analysis with extensive citations. Use it for competitive analysis, daily news monitoring, due diligence on new leads, and market research with academic sources.

[Full page →](/integrations/perplexity)

### SharePoint

Connect your agent to SharePoint to list, read, create, and edit documents in your Microsoft 365 document libraries. Authentication uses a personal access token tied to your Microsoft account, and after connecting you choose which sites and folders to allow. Text-based files (txt, md, csv) can be read and edited; for Office documents the agent can read content but not edit it directly. Use it for document search, content retrieval from handbooks, weekly summary reports, and cross-platform sync between Outlook and SharePoint folders. The integration is currently hidden; email [support@abundly.ai](mailto:support@abundly.ai) to enable it.

[Full page →](/integrations/sharepoint)

### Slack

Your agent communicates with you and your team in Slack with its own identity — its own Slack app, bot user, and token. People @mention it or DM it, and it replies in the same thread with full context, capabilities, and history, just like in the portal chat. Two modes are available: **Chat mode** (default), where Slack threads behave as persistent agent chats, and **Trigger mode**, where each subscribed Slack event is forwarded as a raw payload for ambient monitoring or custom automations (use cautiously to avoid reply loops). Beyond posting and reacting, the agent can call any [Slack Web API method](https://api.slack.com/methods) — pinning messages, adding bookmarks, listing user groups — and it can both download attachments into [agent documents](/features/documents) and send files (including generated charts and SVGs) back into Slack. Setup uses an Abundly-generated app manifest with the right scopes, event subscriptions, and webhook URL pre-filled.

[Full page →](/integrations/slack)

### Snowflake

Your agent can explore your Snowflake databases and run read-only queries so you get insights, reports, and reconciliation checks without writing SQL. Authentication uses a Snowflake Programmatic Access Token (PAT) — no password — with your Account Identifier and Login Name, stored as a team secret. Queries time out after 30 seconds; no DDL or DML is supported. Use it for revenue rankings, scheduled weekly signup summaries posted to Slack, cross-source data reconciliation, schema discovery, and ad-hoc campaign analysis.

[Full page →](/integrations/snowflake)

### Steep

Your agent can discover and query your Steep metrics in plain English, using the same shared definitions (MRR, conversion, etc.) your team already uses in the semantic layer. You get aggregated data — and the generated SQL when relevant — without opening a dashboard or writing SQL. Authentication uses a Steep API key stored as a team secret. Use it for metric lookups, scheduled KPI reports emailed to leadership, cross-tool workflows that pull metrics into Notion updates, and board-prep one-pagers.

[Full page →](/integrations/steep)

### Trello

Connect your agent to Trello to manage cards, respond to board events, and automate your workflow. Authentication uses an API key, token, and one or more board IDs generated through a Trello Power-Up. When you configure board IDs, the agent automatically subscribes to webhooks for card created, card moved, and card commented events — mention these triggers in your agent's instructions to drive workflows like inbox-list triage, comment auto-replies, weekly Done-list cleanup, or creating cards from incoming emails.

[Full page →](/integrations/trello)

### Twilio

Twilio powers your agent's SMS and phone call infrastructure. Available capabilities include **Send SMS**, **Receive SMS**, **Make Phone Call**, and **Receive Phone Call** — phone calls combine Twilio with the [OpenAI Realtime API](/integrations/openai#realtime-voice-api) for natural voice conversation. Use it for appointment reminders, SMS triage, outbound demo scheduling, and inbound support intake. You can configure a whitelist of allowed phone numbers and require approval before sending. Twilio-based capabilities are currently hidden; contact [support@abundly.ai](mailto:support@abundly.ai) to enable them.

[Full page →](/integrations/twilio)

### Twitter / X

Connect your agent to Twitter/X with two separate capabilities: **Twitter/X Search** (search tweets, get user info, analyze followers — uses an API key from TwitterAPI.io) and **Twitter/X Post** (post tweets and replies — uses OAuth credentials from an X Developer app with read-and-write permissions). Use it for brand-mention monitoring with sentiment summaries to Slack, competitor tracking, content publishing with approval workflows, community engagement on product questions, and lead research from new email contacts. The X API Free tier has limited write access — Basic or Pro may be required to post.

[Full page →](/integrations/twitter)

## Guides

### Find your use case

Not every task benefits from an agent — some are better handled by code, others by humans. The sweet spot for agents lies in tasks that are routine and time-consuming, don't require deep expertise or creativity, involve fuzzy inputs that pure code can't handle, and carry correctable (not mission-critical) stakes. To find candidates, map how your team spends time along two axes: frequency/time investment and value of your time. Tasks that are frequent but feel like poor use of a human's time are your best starting points. Good examples include screening incoming emails, generating weekly summaries from multiple sources, and reviewing documents against a checklist. Start with simple use cases to build intuition for how agents think and what context they need before tackling complex workflows.

[Full page →](/guides/find-use-case)

### Agent design

Agent design is the process of figuring out what an agent should do, how it does its work, how success is measured, and how it interacts with humans and systems. The **Agent Design Canvas** is a framework covering ten aspects: Purpose, Triggers, Action Plan, Interfaces, Capabilities, Impact, Input, Knowledge & State, Output, and Success. The Action Plan is the heart of the design — numbered steps showing the workflow with clear human-AI handoffs. Apply the principle of least privilege when choosing capabilities, and define success in measurable business terms. Every agent needs a clearly defined human owner responsible for its behavior, analogous to an editor-in-chief who sets guardrails and follows up when things go wrong. Prototype early with fake data or simulated integrations — a useful first prototype can often be built in 30 minutes. Keep instructions clean by limiting context to what the agent actually needs, and plan for iteration: getting an agent from "works OK" to "awesome" typically takes a few rounds of refinement.

[Full page →](/guides/agent-design)

### Agent optimization

Every agent consumes credits when it works, with costs varying by task complexity and LLM choice. A simple news check might use around 50 credits, while screening 300 acquisition targets could use 5,000. To reduce costs, evaluate run frequency, trim context to essentials, store reference information in agent documents (with Summary visibility) rather than instructions, simplify instruction logic, and consider switching to a cheaper model if your use case allows it. In chat, the context indicator in the action bar shifts from gray through yellow, orange, and red as context grows — start a new chat when it gets high to keep responses fast and cheap. Use the Usage & Limits dashboard to track per-agent consumption and set daily limits to prevent runaway costs.

[Full page →](/guides/agent-optimization)

### Troubleshooting

Common issues fall into a few categories: contacting support from inside the app via **Help & support → Contact support** (a structured email to our support inbox with your verified email as reply-to, optionally including recent pages, browser info, and a screenshot), "Overloaded" errors from LLM provider traffic spikes (wait and retry, or temporarily switch models), integration failures caused by missing OAuth scopes, expired tokens, missing Slack channel invites, misconfigured webhooks, or external rate limits, and invalid verification codes that result from requesting multiple codes — codes expire after 30 minutes and a successful sign-in invalidates remaining ones. For anything not covered, contact [support@abundly.ai](mailto:support@abundly.ai).

[Full page →](/guides/troubleshooting)

## Use Cases

### Use Cases

Abundly publishes detailed use cases at [abundly.ai/use-cases](https://www.abundly.ai/use-cases), each showing the challenge, solution, and results. Below is a categorized summary.

**Support & knowledge base:**

* **HR Support Agent** — Answers employee questions about policies and processes, escalates sensitive issues to HR colleagues, and keeps the knowledge base current.
* **Helpdesk Agent** — Handles incoming support questions with consistent answers, escalates complex issues with context, and learns from every interaction.

**Sales & commercial:**

* **Lead Generation Agent** — Searches the web for companies matching your criteria, verifies relevance, and delivers contact details.
* **Lead Analysis Agent** — Researches incoming leads, scores them against your ideal customer profile, and alerts your team to high-potential opportunities.
* **Sales Proposal Agent** — Transcribes meeting notes, drafts tailored proposals, and formats everything to your company template.
* **Commercial Contracts Agent** — Organizes your contracts, answers questions about terms and pricing, and enables analysis across your entire contract portfolio.

**Finance & legal:**

* **Procurement Agent** — Reviews procurement documentation against standards, validates compliance, and flags contradictions.
* **Target Screening Agent** — Researches potential investment targets, scores them against your criteria, and surfaces the most promising opportunities.

**IT & development:**

* **Security Vulnerability Agent** — Monitors vulnerability disclosures, filters for relevance to your codebase, and alerts the right people.
* **Product Updates Agent** — Monitors your repository, translates commits into plain-language release notes, and posts them to Slack.

[Full page →](/use-cases/index)

## Security

### Security overview

Abundly uses a multi-layer security architecture to protect autonomous AI agents: the **LLM layer** (Anthropic Claude with Constitutional AI by default), the **platform layer** (system prompts, attack detection, guardrails, access controls), the **agent layer** (scoped by instructions and capabilities), and the **user layer** (your configuration and oversight decisions). Data residency is EU-only (Stockholm, Sweden), with AES-256 encryption at rest and TLS 1.2+ in transit. The platform is ISO 27001 certified and GDPR-compliant, and credentials are never exposed to LLMs. Complete audit trails log all agent actions.

[Full page →](/security/overview)

### Risk and autonomy

Not all agents carry the same risk — an internal read-only assistant is fundamentally different from a customer-facing agent with email and database access. You assess risk across five factors: scope, tools, communication reach, data sensitivity, and reversibility. Two guiding principles apply: **Least Privilege** (give agents only what they need) and **Earned Trust** (start narrow, expand as the agent proves itself). A decision spectrum ranges from "agent decides autonomously" for low-stakes tasks to "agent asks human to decide" for critical situations.

[Full page →](/security/risk-autonomy)

### Attack detection

The platform runs an automated attack-detection step on untrusted incoming trigger content — primarily incoming email and SMS, plus escalated scripts that don't opt out — before the agent acts. The detector evaluates the raw trigger payload in a separate context (so it can't be influenced by the same malicious input) for prompt injection, data exfiltration, jailbreak attempts, and reconnaissance. Triggers from authenticated or controlled sources (scheduled tasks, authenticated webhooks) skip this step. When an attack is detected the trigger is blocked, the block is logged to the activity log with the detection reasoning, alert emails are sent to configured recipients, and a diary entry documents what happened.

[Full page →](/security/attack-detection)

### Guardrails

Guardrails are technical constraints enforced by platform code, not LLM reasoning — they cannot be bypassed through prompt injection or conflicting instructions. For email, SMS, and phone capabilities, you can restrict the agent to an allow-list of addresses, domains, or phone numbers, then choose independently what happens for recipients inside the list (send, or ask for approval) and outside it (block, or ask for approval). Workspace administrators can enable or disable capabilities workspace-wide, and team administrators can add further restrictions within their teams.

[Full page →](/security/guardrails)

### User approval

User approval places a human checkpoint before sensitive agent actions like external communication. For communication capabilities you set an allow-list and pick independently what happens inside it (send, or ask for approval) and outside it (block, or ask for approval); other capabilities expose a simple approval toggle. Approval requests appear inline in chat or on the agent's Activity page, where each card has a **Notify agent** checkbox — tick it before approving or rejecting to let the agent take follow-up actions, leave it unchecked to end the action there.

[Full page →](/security/user-approval)

### Access control

Access control operates on three layers: workspace roles (Admin, Member, Guest), agent-level access (Admin, Edit, Use, Nothing), and optional team membership for larger workspaces. You set a default access level per agent — applied to workspace members, or to team members if the agent is assigned to a team — and override it with user-specific rules. The system always grants the highest level between the two. Agents are isolated by default and cannot communicate with each other unless you explicitly enable agent-to-agent communication between specific pairs.

[Full page →](/security/access-control)

### Credentials

Credentials follow three patterns: platform-provided (no setup needed, billed via credits), user-entered API keys (e.g., a Slack bot token), and personal tokens via OAuth for Google Drive, Google Calendar, Gmail, SharePoint, and Outlook. Credentials are stored encrypted with decryption keys separated from the data, and the LLM never sees them — the platform injects authentication after the LLM has generated a request, preventing leaks through model outputs or logs. OAuth tokens auto-refresh, and personal tokens are managed in your profile settings under **Personal Tokens** and can be reused across multiple agents without re-authorizing. Best practices include using least privilege when issuing keys or OAuth grants, rotating credentials after workspace membership changes, revoking unused tokens, and never placing credentials in agent instructions. If a credential is compromised, revoke it immediately, update affected agents, and review the activity log; workspace-wide credentials are available on request via support.

[Full page →](/security/credentials)

### Infrastructure and compliance

All customer data resides in EU data centers in Stockholm, Sweden (AWS eu-north-1 and GCP europe-north2). Encryption covers AES-256 at rest, TLS 1.2+ in transit, and RSA-OAEP with SHA-256 for secrets. The platform is ISO 27001 certified and GDPR-compliant, with SOC 2 Type II on the roadmap. Audit logs are immutable and capture timestamp, actor, trigger, execution, and result for every agent action. Automated daily backups with 9-month retention and cross-region replication support disaster recovery with a 24-hour RTO for critical services.

[Full page →](/security/infrastructure)


# Use Cases
Source: https://docs.abundly.ai/use-cases/index

Examples of AI agents in action across different business functions

We publish detailed use cases on [abundly.ai/use-cases](https://www.abundly.ai/use-cases), showing the challenge, solution, and results for each. Here's a brief summary of a few of them.

## Support & knowledge base

* **[HR Support Agent](https://www.abundly.ai/use-cases/hr-support-agent)** — Answers employee questions about policies and processes, escalates sensitive issues to HR colleagues, and keeps the knowledge base current.
* **[Helpdesk Agent](https://www.abundly.ai/use-cases/helpdesk-agent)** — Handles incoming support questions with consistent answers, escalates complex issues with context, and learns from every interaction.

## Sales & commercial

* **[Lead Generation Agent](https://www.abundly.ai/use-cases/lead-generation-agent)** — Searches the web for companies matching your criteria, verifies relevance, and delivers contact details.
* **[Lead Analysis Agent](https://www.abundly.ai/use-cases/lead-analysis-ai-agent)** — Researches incoming leads, scores them against your ideal customer profile, and alerts your team to high-potential opportunities.
* **[Sales Proposal Agent](https://www.abundly.ai/use-cases/proposal-writer-agent)** — Transcribes meeting notes, drafts tailored proposals, and formats everything to your company template.
* **[Commercial Contracts Agent](https://www.abundly.ai/use-cases/commercial-contracts-agent)** — Organizes your contracts, answers questions about terms and pricing, and enables analysis across your entire contract portfolio.

## Finance & legal

* **[Procurement Agent](https://www.abundly.ai/use-cases/procurement-agent)** — Reviews procurement documentation against standards, validates compliance, and flags contradictions.
* **[Target Screening Agent](https://www.abundly.ai/use-cases/target-screening-agent)** — Researches potential investment targets, scores them against your criteria, and surfaces the most promising opportunities.

## IT & development

* **[Security Vulnerability Agent](https://www.abundly.ai/use-cases/security-vulnerability-agent)** — Monitors vulnerability disclosures, filters for relevance to your codebase, and alerts the right people.
* **[Product Updates Agent](https://www.abundly.ai/use-cases/product-updates-agent)** — Monitors your repository, translates commits into plain-language release notes, and posts them to Slack. This agent also generates [our changelog](/changelog).

## Learn more

<CardGroup>
  <Card title="Browse all use cases" icon="arrow-up-right-from-square" href="https://www.abundly.ai/use-cases">
    Detailed case studies with challenges, solutions, and results
  </Card>

  <Card title="Find your use case" icon="magnifying-glass" href="/guides/find-use-case">
    Framework for identifying tasks suited for agents
  </Card>
</CardGroup>