Skip to content

Tools

The Tools tab on the assistant detail page lets you control which tools the assistant can use during conversations. For a full reference of available tools, see the Tools section.

Enabling and Disabling Tools

You’ll see a list of all available tools — built-in, plugin, and MCP — with checkboxes. By default, all tools are enabled.

When you toggle any tool off, TeamWeb AI creates an explicit allowlist — only checked tools will be available to the assistant. Toggle tools back on to re-enable them.

If no tools have been explicitly toggled, the assistant has access to all available tools. The allowlist is only created after the first toggle action.

Per-Assistant Configuration

Some tools have configurable settings that can be customized per assistant. When a tool supports configuration, a Configure badge appears next to it. Click the badge to open the tool’s configuration page.

Tools ship with sensible defaults, so configuration is optional — you only need to change settings when you want to override the default behavior. A Reset to Defaults button is available to clear any customizations.

See Tool Configuration for details on available settings.

Stateful Tools

Most tools are simple stateless calls, but some tools maintain runtime state inside the conversation. The main example is browser_action.

When you enable browser_action, the assistant can open a managed browser session and keep using it over multiple tool calls in the same conversation. That lets it navigate websites, click through flows, fill forms, wait for dynamic page changes, extract content, and capture screenshots. Screenshot results are returned as normal conversation content, with an inline markdown image reference the assistant can show directly in chat. When that reply is sent over attachment-capable channels, local screenshots are promoted into real outbound attachments automatically.

Important implications:

  • The assistant must explicitly start the browser session before using it.
  • Browser sessions are scoped to the conversation, not shared across unrelated conversations.
  • If the browser runtime expires or is cleaned up, the assistant must start a new session instead of continuing from the old one.
  • Screenshots produced by the browser are stored as media artifacts.
Technical Details

Tool filtering — When an assistant has explicit tool toggles (any tool has been turned off), the system builds an allowlist of enabled tool names. During the agent loop, only tools whose names appear in this allowlist are included in the LLM request. MCP tools (prefixed with mcp_) are always included if their server is enabled for the assistant, regardless of the allowlist.

Task-level restriction — Task definitions can further restrict tools via their allowed_tools field. This filter is applied on top of the assistant’s own tool configuration, providing a second layer of control. For example, a research task can be limited to only search_knowledge and read_code_file, even if the assistant has many more tools enabled. If the task’s allowed tools list is empty, all of the assistant’s enabled tools are available.

Browser session lifecyclebrowser_action uses a managed browser session service rather than a sandbox container. Session metadata is stored in the database, while the live browser runtime remains in the worker/application process. Cleanup tasks expire idle browser sessions automatically.