Tools (MCP)

How agents actually call the integrations you enable.

Discovery

When a session starts, every enabled MCP tool is listed in the agent's system prompt with its name, args, and description. The agent picks tools the way it picks any other capability.

Calling

Tool calls render as cards in chat:

• Pending — agent has called the tool, waiting for the response.
• Result — collapsed by default; click to expand the JSON.

For an always_allow tool, the call goes through immediately. For an always_ask tool, an approve/deny card appears and the session pauses until you click.

Approving

When the prompt comes up, you'll see the tool name and the exact arguments. Approve to run it; deny to send the agent a "denied" result so it can take a different path.

If you trust the agent on this tool, flip the policy to always_allow from the agent's profile to skip the prompt next time.

Errors

Tools fail. The agent gets the error and decides what to do — retry, switch approach, or tell you. You see the failure inline.

Tip

If a tool keeps mis-firing, narrow the agent's scope in extra_prompt: "Only use the GitHub tool for read operations." Cleaner than disabling.

To wire up integrations, see Tools (MCP).