Connected agents
Settings › Connected agents (/settings/grants) lists every agent you let in
through the MCP's OAuth sign-in: Claude, an editor, or another harness.
What a row says
- the client's registered name, as it named itself when it registered;
- its scope (
roorrw) and any OpenID scopes (openid,profile,email); - when you granted it: the first consent, kept across every refresh;
- when it was last used, to the minute, or "never used";
- when its token ends or renews, for a grant that expires.
Last used is written at most once a minute per token, so a busy agent costs one write a minute, not one per call. A refresh keeps both dates.
Revoking
Revoke is one tap: a same-origin POST /settings/grants/<id>/revoke. The
token and its refresh tokens go at once, so the agent's next call to /mcp or
/api/v1 gets 401, and it cannot refresh its way back. The page says what it
ended. A token.deleted event lands in the audit trail.
Only your own grants are listed and only they can be revoked here. Tokens you made by hand are under Settings › Access codes.
In the terminal
Connected agents in the palette lists the same rows; x revokes the
selected one after a confirm and says the same receipt.