Unclear It is not clear which HTTP servers are covered, or whether this tracking is active for anyone.
What
MCP servers are outside programs that give Claude Code extra tools. Each call to one of those tools now has its own watchdog, a check that follows whether the call is still making progress.
- Tracking: for HTTP MCP servers that match a URL check, sending a request is wrapped so the watchdog learns when the request was sent, when it was taken, and when its answer stream ended. Incoming responses mark the request as answered.
- Abort instead of hang: if the answer stream ends mid-call with no answer, the tool call aborts with a
downstream_unreachableerror rather than waiting forever. - The call also records
claudecode/toolUseId.
Before, the only watchdog watched for transport errors and gave up with "response presumed lost" after 90 seconds.
Why
A tool call whose connection dies should now fail promptly with an error instead of leaving your session stuck. When slow MCP calls time out may also differ from before.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
It is not clear which HTTP servers are covered, or whether this tracking is active for anyone.