Repository navigation
Let outside MCP agents stay signed in longer than 30 days #31474
Replies: 4 comments
|
Yes, please. I'm running a self-hosted Claude agent in Slack that orchestrates T3 over MCP from a container. It needs someone to sign it in again by hand every 30 days, and we only find out when its requests start failing. Either a configurable TTL or refresh tokens would fix it for us. |
|
This also affects built-in peer links, rather than only outside MCP agents, on v0.1.8 ( Source-verifiable sequence:
Expected: an optional longer lifetime or renewal for an actively used pairing. Current behavior is documented, so this is an enhancement request.
Please consider refresh/renewal while the pairing remains authorized, or a configurable MCP/peer-session TTL. Changing pairing-code validity does not change the resulting session lifetime. These paths were checked at the tag and main Written by Hermes Agent (GPT-6.1 Sol), reviewed by a human before posting. |
|
Following up on the peer-link case above (#31474 (comment)). #37980 and #37990 give directly paired clients renewal bound to the device key, and #37980 explicitly refuses peer-link sessions. Could peer links get the same treatment once that lands? Today a link between two of your own machines has to be re-paired every 30 days with a code minted on the target. In practice that means scripting A renewal that keeps the current guarantees could look like:
Seen on 0.1.10-nightly.20261116. |
|
+1, with a read-only angle that might make a first step easier. I run a local companion that only observes native orchestration (which threads exist, who delegated to whom, who is waiting on whom) through The only way to automate it today is Suggestions, in case they help scope this:
Checked on v0.1.11-nightly.20261118.4877: |
Uh oh!
There was an error while loading. Please reload this page.
Problem
Outside agents connected through
/mcpwith OAuth lose access every 30 days.MCP_CLIENT_SESSION_TTLis fixed atDuration.days(30)inapps/server/src/auth/EnvironmentAuth.ts, and the authorization server metadata only advertisesgrant_types_supported: ["authorization_code"], so there is no refresh token. When the session expires, the agent's calls fail until someone opens the approval page again and enters a fresh pairing code.For a personal, always-on agent (mine runs as a systemd service and is driven from Telegram), that means a manual re-approval every month, usually noticed only when a request fails.
The same problem was reported as #26748, which was closed without a fix, so I'm proposing it here as an idea.
Possible directions
refresh_tokensupport for MCP clients, so a session renews while it's in use and still expires when idle. Revocation in Settings → Sharing & access would work as it does today.I'm happy to work on either once the direction is agreed.
All reactions