When your agent gets a new MCP tool, who actually says yes?

by•

This has been bugging me more as we hook coding agents up to real tools.

The first time you connect an MCP server, you look at the tool list and you're careful. A week later a new tool has shown up, or a schema changed, and the agent just starts calling it. I only notice when something weird comes back.

I've been treating "which tools this agent can call" like permissions, not like a prompt. Reads are easy to allow. Anything that writes, sends, or touches a real account needs a person to say yes, and it shouldn't be able to talk itself into a new tool mid-session.

Curious how other people handle this. Do you freeze the tool list? Re-approve when something changes? Or just keep an eye on the logs and hope you catch it?

7 views

Add a comment

Replies

Best

reads fine and writes need a yes sounds right to me. the scary part is a tool quietly changing what it does after you approved it, so re-approve on change feels worth the annoyance. does anyone diff the tool schema automatically?