Dispatching GPT via Codex MCP

Overview

You dispatch a task to GPT running as a Codex agent. It has its own sandbox and blank context — it sees nothing of your session, so you construct exactly what it needs. The value is that it is a different model: second opinion, adversarial review, or a parallel executor.

Core principle: One codex call = one independent GPT agentic turn. Reuse the threadId to continue; review its diffs before trusting them.

Prerequisite — set it up on first use. This skill needs the tools mcp__gpt-codex__codex and mcp__gpt-codex__codex-reply. If they're absent, provision them now, then tell the user to restart Claude Code (MCP tools only load at startup):

1. Codex CLI + auth — run codex login status. If Codex isn't installed or isn't logged in, tell the user to run npm i -g @openai/codex && codex login (interactive browser auth — only they can complete it).
2. Register the server yourself (user-level):
   • macOS/Linux: claude mcp add gpt-codex -s user -- codex mcp-server -c sandbox_mode=workspace-write -c approval_policy=never
   • Windows (run from PowerShell, or prefix MSYS_NO_PATHCONV=1 in Git Bash — else /c mangles to C:/): claude mcp add gpt-codex -s user -- cmd /c codex mcp-server -c sandbox_mode=workspace-write -c approval_policy=never
   • Equivalent: copy the bundled references/mcp.json (macOS/Linux) or references/mcp.windows.json (Windows) into the project root as .mcp.json.
3. Tell the user: server registered — restart Claude Code, then ask again. Verify with claude mcp list → gpt-codex: connected.

When to Use