Claude Code /mcp reconnect all: what it does for a Sume MCP server

/mcp reconnect all retries every MCP server that failed to connect or needs authentication, from the interactive terminal, and the changelog lists it in v2.1.284 (2026-09-28). For a Sume server it is safe at any moment, because Sume's hosted endpoint keeps no protocol session: a reconnect does not lose in-flight work, since the work lives in job ids.

The Claude Code facts come from its changelog and MCP documentation, read 2026-09-29. The statement about Sume's server comes from its MCP code and docs: each request is handled on its own and the response sets no session id.

What did the recent Claude Code releases change for MCP?

Several entries touch MCP behavior. The table lists the ones relevant to a remote HTTP server.

Why does a stateless server make reconnects easy?

With no session to restore, a fresh connection can serve the next call with the same credentials.

What a reconnect cannot know is what your agent had started. Submitted video jobs keep running on Sume; the agent recovers them by job id with jobs_status, jobs_wait or jobs_result.

What should I do after a reconnect mid-run?

First ask the model to call mcp_health to confirm the endpoint and auth. Then have it wait again on the job ids it already holds. A wait_slice_expired result only means the slice ended and the jobs are still running.

Do not resubmit a paid create call because a reconnect happened. If the model must retry a create whose result it never saw, reuse the same idempotency_key.

When is /mcp reconnect all not the fix?

Reconnect all retries a server that needs authentication but does not sign you in; run claude mcp login sume to complete OAuth. An OAuth session with only mcp:read will also connect fine and still refuse paid tools with insufficient_scope. That is a scope question, not a connection one.