Skip to main content
A tool is a Python function an agent may call. The core ships a portfolio of 31 tool modules; an agent is given the subset it needs, and each tool that talks to an external service reads its settings from a tool configuration attached to that agent.

​
The portfolio

Ask a running instance rather than trusting this table to stay exhaustive:

​
Tool configurations

A tool module is generic; a tool configuration binds it to something concrete — this database, that Drive folder, this API key. Configurations belong to an agent, so two agents can use database against two different servers. Credentials inside a configuration are encrypted at rest with ENCRYPT_KEY.
Tools that act on Google or Microsoft data do not hold their own credentials: they use the OAuth integration the user connected, so revoking the integration revokes the agent’s access with it. See Integrations.

​
Tools backed by MCP

Two modules are not plain functions but MCP toolsets: database_mcp and web_search_mcp. They are discovered by naming convention — a factory called mcp_tool_set_* is invoked at build time so the toolset is created fresh, with the current key, rather than captured once at import. See MCP.

​
Adding a tool

A module in the portfolio exposes plain functions with type hints and a docstring — that docstring is what the model reads to decide whether to call it, so it is interface, not commentary. Reload the agent (POST /api/agents/{agent_id}/reload) to pick up a change without restarting the server.