StdioClientTransport.close() does not kill the process tree, leaving orphan processes #2023

Problem

When StdioClientTransport.close() is called, only the direct child process is signaled. If the MCP server was started via a wrapper command (e.g. npx, uvx, python -m), the wrapper's child processes (the actual server) become orphans and continue running indefinitely.

Root cause

The close() method uses processToClose.kill('SIGTERM') / processToClose.kill('SIGKILL'), which calls Node's ChildProcess.kill(). This only signals the direct child PID — not its descendants.

This affects all platforms:

• macOS/Linux: No process group kill (process.kill(-pid, signal)) is used, so children of the direct process are not signaled.
• Windows: ChildProcess.kill() does not kill the process tree at all. Only taskkill /T /F /PID <pid> can do that.

Reproduction

1. Start an MCP server via a wrapper (e.g. npx @some/mcp-server)
2. Call transport.close()
3. Observe that npx is killed but the underlying node process remains running

Possible solutions

Option A: Kill the process tree in close() — use pgrep -P (Unix) or taskkill /T /F (Windows) to enumerate and signal all descendants before/instead of ChildProcess.kill().

Option B: Spawn with detached: true and use process.kill(-pid, signal) to kill the process group. However, this has trade-offs: if the parent crashes, detached children become permanent orphans with no automatic cleanup.

Option C: Expose a hook or option so consumers can provide their own process cleanup logic.