🤝 ask-codex

Make OpenAI Codex a teammate inside Claude Code.

Delegate reviews, plans, refactors, and implementations to Codex over the local first-party codex MCP server — with the model + reasoning effort matched to the task, and a subagent that keeps multi-turn loops out of your main context.

#claude-code · #codex · #mcp · #ai-pair-programming · #code-review · #developer-tools

---

Why

Claude Code is great. So is Codex. They're better together — one drafts, the other checks; one plans, the other stress-tests. ask-codex teaches Claude when to bring Codex in, which Codex model + effort fits the job, and how to phrase the hand-off so the answer comes back tight instead of as a wall of tokens.

Everything runs locally on your own Codex/ChatGPT seat through OpenAI's first-party codex mcp-server. There is no third-party bridge and no key sharing.

What you get

Piece	What it does
ask-codex skill	Fires on "ask codex", "second opinion", "let codex review/plan/refactor" — and proactively offers Codex on high-stakes calls (architecture, irreversible changes, stubborn bugs, security-critical diffs). Never a silent call: it offers, you decide.
codex-teammate subagent	Runs multi-turn Codex loops (implement → self-review → fix) in an isolated context and returns only the outcome.
scripts/sync-models.py	Regenerates the model list from Codex's own cache so the skill never suggests a model that doesn't exist or an effort a model can't do.

Requirements

• Claude Code
• OpenAI Codex CLI (codex), signed in to your seat
• Python 3.9+ (only for the optional sync-models.py maintenance script)

Install

1. Register Codex as an MCP server for Claude Code

ask-codex talks to Codex through the MCP server that ships with the Codex CLI:

claude mcp add codex -- codex mcp-server

(On Windows, if codex is a shim: claude mcp add codex -- cmd /c codex mcp-server.)

Verify it's connected with /mcp inside Claude Code — you should see the codex server exposing codex and codex-reply tools.

2. Add the plugin

From inside Claude Code:

/plugin marketplace add emirdegirmenci/ask-codex
/plugin install ask-codex@ask-codex

That's it. The skill and the codex-teammate subagent are now available.

A git release tag is not required to install — a Claude Code marketplace is just a repo with .claude-plugin/marketplace.json. Tags/releases (like v1.0.0) are provided for version tracking and discoverability.

Use it

Just talk to Claude Code:

• "Ask Codex to review my uncommitted diff." → read-only review, findings ranked by severity.
• "Get a second opinion on this plan from Codex." → verdict + reasoning in ≤10 lines.
• "Let Codex refactor parser.py to remove the duplication." → Codex writes, returns the diff.
• "Have Codex implement the retry logic and self-review it." → multi-turn loop via the subagent.

On a heavy call you didn't ask about, Claude will offer:

"This is an architecture decision with real blast radius — want a Codex second opinion too? (gpt-5.6-sol/high, read-only.)"

You say yes or no. It never spends your Codex seat silently.

Picking the model + effort

The skill matches the model to the weight of the work:

Task	Model	Effort
Hard / architecture / deep debugging / whole-repo	gpt-5.6-sol	high (→ max/ultra for the hardest)
Everyday coding, balanced	gpt-5.6-terra	medium
Fast, simple, repetitive	gpt-5.6-luna	medium

The full, always-accurate list lives in skills/ask-codex/MODELS.md — it's generated from the Codex CLI's own model cache, so it can't drift into suggesting models that aren't on your seat.

Keeping the model list fresh

If OpenAI ships new Codex models (or your seat changes), refresh the list:

python scripts/sync-models.py           # rewrite MODELS.md from your Codex cache
python scripts/sync-models.py --check    # CI-friendly: exit 1 if MODELS.md is stale

It reads $CODEX_HOME/models_cache.json (default ~/.codex/models_cache.json). If that cache isn't present, the committed MODELS.md is kept as the shipped fallback.

How it stays cheap

Delegation only saves tokens if the hand-off is thin both ways. The skill enforces four habits: pass paths + cwd (Codex reads the repo itself, Claude never loads the files), bound the reply ("diff only", "findings only"), prefer read-only for review/plan/consult, and run heavy multi-turn work in the codex-teammate subagent so the chatter never hits your main context.

Trust boundary

read-only inspects; workspace-write lets Codex edit files in cwd; danger-full-access removes the sandbox and is never used unless you explicitly ask. Central policy lives in your $CODEX_HOME/config.toml.

License

MIT © Emir Degirmenci

---

_{Not affiliated with OpenAI or Anthropic. "Codex" and "Claude" are trademarks of their respective owners.}