Re-run MCP login and verify server identity after tool failures
You will confirm which server build your client actually runs, then retest one simple call. This shows whether the failure sits in the server or in your client integration.
- Capture errorSave the exact error text and tool name
- Verify serverConfirm which package or binary is running
- Restart connectionRestart the server connection and check the token
- RetestRun the same trivial input and compare clients
Why it matters
A tool can fail after working for days with no change on your side. In one report, create_or_update_file returned a schema error for new and edited files while the Claude Desktop config and GitHub token stayed the same, and the error was explained as a mismatch between the server's function schema and GitHub's API requirements (the GitHub MCP issue (opens external site)). In another report, a server worked in MCP Inspector but failed in Cursor Agent mode (the Cursor issue (opens external site)). Checking which build runs and retesting one call tells you whether the fault is in the server or in the client.
Steps
- Write down the exact error text and the tool name before you change anything.
- Confirm which server package or binary your client launches. The Cursor report carries a server-filesystem label, but its maintainer notice and comments are about server-github, so labels can be wrong (the Cursor issue (opens external site)).
- If you run the GitHub server, move to the maintained github/github-mcp-server project. Development has moved there, and pull requests to the original repo are no longer merged (the GitHub MCP issue (opens external site)).
- Restart the server connection in your client, and confirm the GitHub token it uses is still valid and has access to the repository. This is general practice for narrowing down credential problems.
- Run one trivial call, such as a small single-file commit or a read-only lookup. Use identical input on every attempt so the results can be compared.
- Run the same call from a second client, such as MCP Inspector. If it works there and fails in your main client, the fault is most likely in that client's integration (the Cursor issue (opens external site)).
- If a fallback tool is available, use it while you wait. A commenter reported that Claude could work around the broken file call by using push_files (the GitHub MCP issue (opens external site)). Retest the original tool after switching builds.
Common mistakes
- Assuming a client restart is the same as a fresh server connection. Restart the connection explicitly and check the result.
- Trusting the server name shown in the client UI without checking which package actually runs.
- Changing the test input between attempts, which makes it hard to tell whether anything changed.
- Filing a bug against a repo that no longer accepts fixes for that server.
Checklist
0 of 8 done