Corone Vault

Codex (ChatGPT) MCP integration

2 min read

Codex MCP — Claude → ChatGPT/GPT-5 bridge

Installed 2026-05-03. Lets Claude invoke ChatGPT (GPT-5/5.5) for second opinions, parallel work, cross-AI delegation.

Setup

  • Codex CLI: /usr/bin/codex v0.128.0 (apt package)
  • Auth: ~/.codex/auth.json (ChatGPT Plus account, paired via codex login)
  • MCP transport: stdio
  • Server name: codex in ~/.claude/mcp_servers.json

Verifying

codex --version            # → codex-cli 0.128.0
codex login --status       # → logged in as <email>

If MCP codex is offline:

codex login                # follow browser flow
# then restart Claude Code MCP server

Usage patterns

Quick second opinion

mcp__codex__codex_chat(prompt: "Critique this approach: ...", model: "gpt-5")

Long-form research (GPT can browse)

mcp__codex__codex_research(query: "Latest Tailwind 4 alpha modifier behavior with CSS vars")

Code review by GPT-5

Hand a diff to GPT for parallel review while Claude handles other work.

When to invoke

  • User explicitly asks “ask Codex” / “ask ChatGPT” / “second opinion”
  • Stuck on a problem, need a different model perspective
  • Need GPT browsing for very recent docs (post-cutoff)
  • Long parallel research while continuing main thread

When NOT to invoke

  • Simple questions Claude can answer directly
  • Anything with secrets in the prompt (token-based MCP, but no extra need to leak)
  • When user wants Claude’s opinion specifically

Reverse direction

Codex side does NOT have Claude MCP installed. Bridge is one-way: Claude → Codex.

  • ~/.claude/skills/codex/ — also exposes the bridge as a slash command
  • See skills-inventory

Failure modes

  • Codex auth expired → codex login again
  • Network rejection → check ~/.codex/config.toml proxy settings
  • MCP not loaded → restart Claude Code session

Graph View