Built-in Tools

TeamWeb AI ships with a comprehensive set of built-in tools organized into categories. These tools are available to all assistants by default and can be toggled on or off per assistant from the assistant’s Tools tab.

Content & Canvas

Communication

Knowledge & Research

Notes

User Memories

Notes support an optional per-user scoping mode. When an assistant calls save_note or get_notes with about_user: true, the note is tied to the specific person the assistant is currently speaking with — whether that’s a registered user or an external contact.

This lets assistants build up personal memory about individuals across conversations. For example, an assistant might remember a customer’s preferred communication style, project context, or previous requests.

How it works:

• Saving — save_note with about_user: true creates a note scoped to the current person. If a note with the same title already exists for that person, the content is updated instead of creating a duplicate.
• Retrieving — get_notes with about_user: true returns only notes about the current person. Without the flag (or about_user: false), only general assistant notes are returned.
• Auto-injection — The 5 most recently updated user memories are automatically included in the assistant’s system prompt at the start of each conversation turn, so the assistant doesn’t need to explicitly retrieve them.
• Promotion — When a contact is promoted to a registered user, their memories are automatically transferred to the new user account.

User memories can be viewed and managed from the Notes tab on the assistant detail page (admin), the staff chat memories view, or the contact detail page.

Todos

Inbox

Child Runs

Project Management

Tasks & Scheduling

Conversations

Code Execution

Browser

browser_action

The browser_action tool gives an assistant access to a real managed browser session. Unlike simple HTTP fetching, this lets the assistant interact with dynamic pages that require JavaScript, form input, navigation, and DOM-based waiting.

Supported actions:

• start — Start a browser session for the current conversation. Call this once.
• navigate — Go to an http:// or https:// URL. Cookie/consent banners are auto-dismissed by default; pass auto_dismiss: false to keep the banner if you need to interact with it.
• click — Click an element using a CSS selector.
• click_text — Click the first visible element matching a label. Optional role (e.g. "button", "link", "tab") disambiguates when the same label exists on multiple element types. Prefer this over building speculative CSS selectors.
• type — Clear a field and type text into it using a CSS selector.
• wait_for — Wait for either a selector to appear or the URL to contain a given substring.
• extract — Extract data from the current page or an optional selector scope. See Extract modes below.
• screenshot — Capture a PNG screenshot of the current page or a specific element. Last-resort orientation tool — prefer extract mode='page_model'.
• scroll — Scroll the page (viewport, page, a pixel count) or scroll a specific selector into view. Useful for lazy-loaded list pages.
• find_text — Find visible elements matching a text label and return their selectors. Use when you need a selector for a follow-up extract or wait_for; for clicking, prefer click_text.
• accept_cookies — Best-effort dismissal of a cookie/consent banner. Rarely needed manually since navigate runs this automatically.
• close — Close the session and release browser resources.

Extract modes:

• text — Plain visible text from the page or a selector.
• html — Inner HTML, or outerHTML when a selector is given (use this for self-closing elements like <img>/<input> to read attributes such as src, value, alt).
• links — Array of {href, text} for every <a> in scope. Capped at 50 entries with per-item href/text trimming, so a busy page can’t dump tens of KB into context.
• form_fields — Array of {name, type, value, placeholder} for every form field in scope.
• page_model — Compact JSON skeleton of the current page: visible headings, landmarks (main, nav, etc.), interactive elements (each with role, name, selector), forms with field labels, and detected list groups (≥3 sibling elements with the same shape). Output is bounded to fit comfortably under 8 KB. This is the cheap default for orienting on a new page.
• structured — Schema-driven scraping for lists/grids/cards. Pass a repeat_selector matched once per item and a schema of {name, selector, attr?} field specs. Returns a list of dicts. Reads innerText by default, or the named attribute when attr is set. Caps at max_items (default 50).

Key behaviors:

• Conversation-scoped sessions — Browser sessions belong to the current conversation. They can be reused across multiple tool calls in that conversation by using the same session_key.
• Explicit lifecycle — The assistant must start a session before it can navigate or interact with pages. If the browser runtime expires, the session must be started again.
• Managed profile state — Login cookies and local browser storage persist while the managed browser profile remains active. TeamWeb AI does not replay old browser actions to reconstruct expired sessions.
• Auto-dismiss on navigate — Each navigate call attempts a best-effort cookie/consent banner dismissal once the page is loaded. Failures are swallowed so they cannot fail the navigation. Disable per-call with auto_dismiss: false.
• Inline screenshot results — Screenshots are stored through the media system and returned in the tool result as both a URL and a ready-to-use markdown image snippet. This is intended for normal conversation output, not the canvas or deliverables system.
• Channel-friendly delivery — When a reply containing a local screenshot is sent through email, Telegram, Slack, or other attachment-capable channels, TeamWeb AI promotes that referenced media into a real outbound attachment instead of relying on the raw markdown image URL.

Best practices for the assistant:

• Don’t guess URLs. Use site navigation. If the right URL isn’t already known, run web_search first or navigate to the site root and use extract mode='page_model' to follow real links. Pattern-matching paths from product or destination names ("/cruises/port/copenhagen", “/products/foo”) fails fast with 404s and burns turns.
• Orient with page_model, not screenshot. A page_model extract is much cheaper and gives the assistant a click-ready map. Reach for screenshot only when visual layout is genuinely needed.
• Click by label. click_text removes the find-then-click dance; reserve CSS selectors for cases where labels are ambiguous.
• Scrape lists with structured mode. Define a repeat_selector and a schema rather than dumping mode='text' and parsing prose.
• Use the run journal. For non-trivial multi-page research, write objective, current state, verified findings, and the next action via update_run_journal / append_run_journal_entry early and often. The journal is reloaded on every continuation, including after iteration-limit halts — see Run Journal.

The bundled browser-research skill packages these guidelines plus worked patterns (form-filling, embedded SPA app-state extraction, lazy-loaded results, recovery from CAPTCHAs and dead-ends) into a single activatable instruction set.

Current limits:

• CSS selectors only in v1
• No multi-tab support
• No video recording
• No automatic recovery of expired sessions

Planning

manage_plan

The manage_plan tool lets an assistant create, inspect, and update a structured task plan within the current conversation. Plans support hierarchical (nested) steps, so a top-level step can contain sub-steps for finer-grained tracking.

Actions:

• create — Create a new plan with a title and a list of steps. Steps can be nested to form a hierarchy.
• get — Retrieve the current plan and its step statuses.
• update_step — Mark a step as completed, in-progress, skipped, or failed.
• add_steps — Append new steps to an existing plan.
• remove_steps — Remove steps from the plan by ID.

Plans appear in the workspace sidebar during the conversation, giving you a live view of what the assistant is working on and how far along it is.

Configuration options (set per assistant):

think

The think tool gives the assistant a structured space for self-reflection. When called, the agent writes down its reasoning about the current task — what it knows, what’s uncertain, and what approach it should take. The reasoning is stored as a hidden message (not shown to the user) but is available in API Logs for debugging.

Research on LLM tool-use shows a 21% average improvement in task quality when agents use structured reflection before acting. The tool is especially valuable for multi-step or ambiguous tasks.

Workspace

Workspace files give assistants a scratch-pad for producing and iterating on deliverables — drafts, data files, generated code, and so on. Files persist across turns within the same conversation and are visible in the workspace sidebar.

Key behaviors:

• Scoped to the conversation — Each conversation has its own isolated workspace. Files are not shared between conversations.
• Upsert semantics — write_file overwrites the file if one with the same filename already exists.
• Automatic cleanup — Workspace files are removed after a configurable time-to-live once the conversation ends. The TTL is controlled by the WORKSPACE_TTL_DAYS environment variable (default: 7 days).

Workflow

These tools are automatically available during workflow execution. They do not need to be enabled per assistant – the workflow engine injects them when a workflow run is active.

Skills