Tools

Tools are the capabilities agents can use. Three kinds:

• Built-in tools: capabilities the platform provides directly (memory, task history, task setup, feedback). Never passed in tools; the configurable ones are toggled with enable_* fields. See Built-in tools.
• 3rd-party integrations: external services like Gmail, Slack, and Google Ads, connected via OAuth.
• Desktop tools: when sandbox execution is enabled, every run also gets computer, bash, and a file editor over a full Linux desktop, automatically. See Computer use.

The tools array below is only for 3rd-party integrations and custom tools. Passing a built-in name (e.g. "memory") returns a 422.

List available integrations

Assign tools to an agent

Set the base tool set when creating an agent. These tools are available on every run:

Override per task or per run

Tasks and runs can override the agent's tool set:

The most specific level wins: run over task over agent. If a run omits tools, the task's tools apply, then the agent's.

This cascade covers catalog tools. Custom tools attach to an agent only — a task- or run-level tools list containing a custom slug is rejected. For one agent per end-user acting on your own API, see Your API as MCP tools.

Account-level connections are made through the Developer Dashboard; OAuth flows are handled automatically, and client.apps.list() (above) shows what's connected.

End-user connections (multi-tenant OAuth)

Connect apps programmatically for your end-users. Each customer authorizes their own Gmail, Slack, etc.

1. Initiate connection

2. Complete connection

After the user completes OAuth and is redirected back:

3. Check status and disconnect

Isolation

• Each end-user gets their own OAuth connection, scoped by user_id
• Runs for that user automatically use their connected accounts
• There is no fallback from user_id-scoped runs to account-level connections

Next: Built-in tools · Computer use · Human-in-the-loop · Users