openclaw - 💡(How to fix) Fix MCP stdio server processes become orphans after gateway restart [1 comments, 2 participants]

Describe the bug

When the gateway restarts (via openclaw gateway restart), MCP servers configured in mcp.servers that use stdio transport are left running as orphan processes. The new gateway instance forks fresh MCP server processes, but the old ones from the previous gateway instance are never cleaned up.

Over a day of normal operation (with periodic gateway restarts for config changes), I accumulated 41 orphan gbrain serve processes, each holding a PGLite database lock. This caused new MCP connections to time out (30s) because they couldn't acquire the database lock.

Config

{
  "mcp": {
    "servers": {
      "gbrain": {
        "command": "gbrain",
        "args": ["serve"]
      }
    }
  }
}

Expected behavior

When the gateway shuts down or restarts:

1. All child MCP server processes should be terminated (SIGTERM → SIGKILL after timeout)
2. On startup, the gateway should detect and clean up orphan processes from a previous instance (e.g., by PID file or process group)

Additionally, it would be great to support an idleTimeoutMs option per MCP server:

{
  "mcp": {
    "servers": {
      "gbrain": {
        "command": "gbrain",
        "args": ["serve"],
        "idleTimeoutMs": 300000
      }
    }
  }
}

This way, MCP servers that haven't received requests for N minutes would be automatically terminated, and re-spawned on demand.

Logs

[bundle-mcp] failed to start server "gbrain" (gbrain serve): Error: MCP server connection timed out after 30000ms

This error repeats for every agent session trying to connect, because the PGLite database is locked by orphan processes.

Environment

• OpenClaw 2026.4.22
• macOS 15.x (arm64)
• Node v24.13.1
• 14 Feishu bot accounts, each bound to a different agent
• gbrain 0.16.4 as MCP server (PGLite backend)

Workaround

Cron job that kills stale gbrain serve processes every 5 minutes:

*/5 * * * * /path/to/gbrain-guardian.sh