Skip to content

Framework does not handle refreshed tokens correctly #1250

Open

Description

@pematth

Initial Checks

Description

I use python sdk with Keycloak as external authz provider. I allowed my mcp client (in my test scenario it is latest VSC) to get AT and RT (using offline_access scope). I have implemented a TokenVerifier that is doing validation of the MCP keycloak JWT. And then in some MCP tool, i use get_access_token() from the framework to use the JWT. I log both values and i can clearly see that it works fine initially, but after the token was refreshed on client-side, i still get the old accesst token from get_access_token() function. Not sure if you intended to support token refresh in a clean way. But IMHO that is a bug that should be fixed soon.

Example Code

Python & MCP Python SDK

1.12.4

Activity

  1. lucheng-hugh commented on Oct 28, 2025

    @lucheng-hugh

    I confirmed the bug still exists in the latest 1.19.0

  2. bgaidioz commented on Dec 3, 2025

    @bgaidioz

    This GitHub issue was referenced in FastMCP because it describes the same problem (PrefectHQ/fastmcp#1863). It’s also related to the implementation that uses a context variable. The fix is here: PrefectHQ/fastmcp#2505

  3. bgaidioz commented on Feb 5, 2026

    @bgaidioz

    Hello.

    Here's a patch that I believe fixes the issue.

    I added a regression test that fails as one would expect without the patch (the old token is seen on the server side). CI failed on coverage for the new branch in _handle_request. I did some efforts for that, added a small, low‑level unit test to cover it, but it’s fairly close to internals and I’m not sure this is the preferred approach.

    This isn't a direct port of the patch to fastmcp. Setting the context var felt reasonable to me.

    That fixes the demo repro and our own MCP server when running against keycloak.

    I hope this helps.

  4. grahonan commented on Feb 12, 2026

    @grahonan

    I also ran into this and would love to see it fixed soon.

  5. added
    P1Significant bug affecting many users, highly requested feature
    and removed
    P0Broken core functionality, security issues, critical missing feature
    on Feb 19, 2026
  6. maxisbey commented on Feb 19, 2026

    @maxisbey
    Contributor

    This may be addressed as part of the broader auth rework in #1743. Keeping as a standalone issue for now in case it can be fixed independently before the rework lands.

  7. Ptraaah commented on Oct 31, 2026

    @Ptraaah

    I dug into this and can confirm the bug on the v1 line. Small repro against a Streamable HTTP server with a plain token verifier: initialize and call a tool presenting token A, then present token B in the same session (same client, like after a refresh) - get_access_token() inside the tool still returns token A.

    One thing worth flagging: on current main (v2) the same scenario already returns the refreshed token, because the new dispatcher propagates each message's sender context into the handler. So this is live on v1.x, which matches the reports here (1.16, 1.19).

    Where I would fix it on v1.x: AuthContextMiddleware sets auth_context_var per HTTP request, but handlers run in the session's long-lived task, so they keep seeing the context of whichever request opened the session. _handle_request in the lowlevel server already re-applies the per-request request_ctx from the message metadata for that exact reason, and the HTTP transports attach the current request to that metadata - its scope carries this request's authenticated user. So I would apply auth_context_var from that scope for the duration of the handler and reset it afterwards, same pattern, no public API change. Both HTTP transports attach the request, so both get it; stdio has no request attached and stays untouched. And since the value is applied per request at dispatch, parallel sessions cannot see each other's tokens.

    I know #1998 took a similar route a while back (thanks @bgaidioz, the repro repo helped pin this down) but it has been sitting in draft, so I will put up a fresh PR against v1.x with regression tests, unless you would rather fold this into the #1743 rework instead.

  8. Ptraaah commented on Oct 31, 2026

    @Ptraaah

    Quick follow-up on this: I opened PR #3794 with the fix, but it was closed automatically within a couple of minutes because I am not assigned to this issue. The change itself is ready and tested locally (new regression tests fail without the fix and pass with it, details are in the PR). Could a maintainer assign me to this issue so I can reopen it? Happy to adjust the approach first if you would rather see it done differently.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P1Significant bug affecting many users, highly requested featureauthIssues and PRs related to Authentication / OAuthbugSomething isn't workingready for workEnough information for someone to start working on

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions