MCP failures are now sorted into causes like server down, login problem or bad config for reporting.
What's wrong with this entry?
Failures from MCP servers are now classified as downstream_error, downstream_unreachable, user_auth, user_config or other and attached to tool-error telemetry, so it is possible to tell a server that is down from a login or config problem. Nothing you see changes.
- Classification uses error classes, JSON-RPC and HTTP codes, and a curated set of network error codes including ECONNREFUSED, ERR_PROXY_TUNNEL and ESOCKETTIMEDOUT, following up to four chained causes.
- The classification is also attached to "MCP server is not connected" and MCP tool timeout errors.
- MCP connect telemetry gained flags for a suppressed listen and a denylisted negotiation.
downstream_unreachable, mcpErrorSource
Strings lifted out of the shipped bundle, so the claim above can be checked against them.