Claude Code MCP server shows connected but tools fail after a 404
Claude Code 2.1.283 fixed a brief 404 from a stateless remote MCP server leaving it unusable while shown connected. What to do for Sume's endpoint.
Claude Code 2.1.283 fixed a case where a brief HTTP 404 from a stateless remote MCP server, for example a proxy in the middle of a redeploy, left that server unusable for the rest of the session while it still showed as connected. On 2.1.283 or later the session recovers; on an older build, reconnect the server. Your Sume jobs are not affected either way.
The fix is from the Claude Code changelog entry for September 25, 2026, read 2026-10-01. Sume's hosted endpoint is a single URL, https://mcp.sume.com/mcp, per the MCP overview.
What did the bug look like?
The server list said connected, but tool calls kept failing. The trigger was one short-lived 404 from the server side of a stateless remote MCP server. Because such a server holds no session between requests, the next request should have worked; the client had instead marked the server unusable. The changelog describes the cause as a proxy redeploy example, not a Sume event.
How do I tell a brief 404 from a wrong URL?
A wrong URL fails every time; a brief 404 fails once and then the endpoint answers again. The quickstart gives one correct URL, so a persistent 404 points at your entry.
| Symptom | Likely cause | Next step |
|---|---|---|
| 404 on every call from the first call | URL differs from https://mcp.sume.com/mcp | Fix the entry and reconnect |
| Worked, then one failure, then tools fail while connected | The pre-2.1.283 stuck state | Update Claude Code, run /mcp reconnect all |
| Tool call times out, no 404 | Long wait, not a routing error | Re-issue jobs_wait with the same ids |
Do I lose a running Sume job when this happens?
No. A job is created on Sume and keeps its id; the jobs page says to store the id so work can be recovered after process restarts, and not to resubmit a paid request because a local process failed. After you reconnect, call jobs_list or jobs_status with the id.
Which command reconnects every server at once?
/mcp reconnect all, added in Claude Code 2.1.284, retries every MCP server that failed to connect or needs authentication. Run it in the interactive terminal. If the Sume entry then asks you to sign in, run claude mcp login sume as the quickstart shows.
Sources
Related posts
More in Integrations
- Switch Opus 5.5 and Sonnet 5.5 mid-session: Sume jobs keep running
Claude Code 2.1.287 fixed /model and opusplan switches rewriting earlier MCP tool announcements. Sume jobs live on the server, so a new model can pick them up.
- Claude Managed Agents vault static_bearer for Sume tools
For Sume's hosted MCP in a Claude Managed Agents vault, use a static_bearer credential keyed to https://mcp.sume.com/mcp; OAuth tokens last an hour.
- Sonnet 5.5 and a 10-minute render: about 12 jobs_wait calls
A Sume job that takes ten minutes needs roughly 11 to 12 jobs_wait calls at the 55 and 50 second slices. How to keep a Sonnet 5.5 tool loop that short.
- Code by Zapier 225 requests per 10 seconds and Sume 429s
Zapier lists 225 requests per 10 seconds for Code by Zapier on Pro and Team. Sume has its own 429 rate_limited: honor retry-after and send an Idempotency-Key.
Written by Sume