MCP servers are add-ons that give Claude Code extra tools. When Claude Code is driven through --sdk-url, their startup handling changed:
A sync-kind update is waited on only up to a limit: MCP_CONNECT_TIMEOUT_MS, 5000 milliseconds by default, or longer when relay-marked.
If the limit is reached, Claude Code records tengu_sdk_mcp_update_wait_timeout and moves on.
Servers already waited on at entry are skipped in the first-message wait.
Servers started since the process was kicked off are now waited for before the turn goes ahead.
Session-start hooks (commands you configure to run when a session begins) are awaited for a claimed process before its input is drained.
Why
A slow MCP server can no longer hold up the first message indefinitely. Servers that start late are still ready before they are needed.
How sure we are
One source agreesOne thing we can check says the same as this entry.
Anthropic's release notes agreeFixed MCP_CONNECTION_NONBLOCKING=0 giving up on claude.ai connectors after 1s instead of honoring MCP_CONNECT_TIMEOUT_MS
