MCP client times out when launching Docker MCP Toolkit gateway (full catalog, Windows) #5444

Summary

Codex CLI (0.47.0) fails to attach to Docker’s MCP Toolkit when the gateway
is launched with the full Docker catalog. Codex spawns docker mcp gateway run
via stdio, but the MCP handshake never completes: the CLI waits for a JSON-
RPC response that never arrives and exits after ~1 s with “MCP client for
MCP_DOCKER failed to start: request timed out”. The gateway itself finishes
initialisation (~1.7 s) and prints “Start stdio server”, so the process is
healthy—the timeout happens on Codex’s side.

Environment

• Codex CLI: codex --version → 0.47.0 (on Windows 11, PowerShell)

• Docker Desktop: (include your version if you have it handy)

• Docker MCP Toolkit CLI (bundled with Docker Desktop)

• Codex config entry:

  [mcp_servers.MCP_DOCKER]
  command = "docker"
  args = ["mcp", "gateway", "run"]

Steps to reproduce

1. Ensure Docker Desktop’s MCP catalog is fresh (default registry.yaml with
   all servers enabled).
2. Open one PowerShell terminal and run:
   tools.

Actual behaviour

Codex tears the connection down after roughly one second:

■ MCP client for MCP_DOCKER failed to start: request timed out

Relevant log excerpt from C:\Users<user>.codex\log\codex-tui.log:

2025-10-21T09:53:10.916906Z INFO aggregated 0 tools from 0 servers
2025-10-21T09:53:10.916945Z ERROR MCP client for MCP_DOCKER failed to
start: request timed out

Meanwhile the gateway that Codex launched reports:

• Watching for configuration updates...

Initialized in 1.7119549s
Start stdio server

Additional notes

• If I manually run docker mcp gateway run beforehand, it reaches “Start
  stdio server” in ~1.7 s (with images cached). Codex still times out when it
  launches another instance itself.
• Trying the HTTP/SSE transport (codex mcp add --url http://127.0.0.1:5000)
  fails earlier with 405 Method Not Allowed because the Docker gateway only
  allows GET on /sse; the Codex client POSTs the handshake payload. So the
  stdio path is the only available option.
• Reducing the enabled servers (e.g. only time, memory, SQLite) makes the
  handshake succeed—suggesting Codex’s timeout is simply too aggressive for
  the full catalog initialisation.

Please let me know if you need the entire log file or a recording; happy to
provide any extra diagnostics.