Run OpenAI's Codex (GPT-5.4) directly from Claude Code — as a collaborator, not a competitor.
Claude Code is a strong orchestrator. Codex is a strong autonomous executor — and it comes with up to 8 parallel sub-agents built in. This skill puts all of that at Claude Code's disposal.
Once installed, Claude Code gains access to Codex's entire execution layer: its sub-agent pool, its skill system, its multi-turn memory. A single /codex call from your Claude Code session can fan out into 8 concurrent workers, each going 16 levels deep, and funnel the results back into your conversation. Claude Code doesn't just call Codex — it commands Codex's full infrastructure.
No tab-switching. No copy-pasting between terminals. One prompt, two engines, and a pool of sub-agents you didn't have before.
- One command —
/codex <task>dispatches work to Codex and surfaces results in your Claude Code session - Sticky sessions — pick up where you left off; Codex remembers context across invocations
- Skill passthrough — invoke Codex's own skills (
$polymarket,$ezfile, etc.) without leaving Claude Code - Multi-agent fan-out — Codex automatically parallelizes across up to 8 sub-agents when the task calls for it
Without this skill, Claude Code works alone. With it, Claude Code has a bench.
When you send a task to Codex via /codex, Codex can internally spin up up to 8 parallel sub-agents with a recursion depth of 16. Those sub-agents are now effectively working for Claude Code. You describe what you need; Claude Code decides what to handle itself and what to ship to Codex; Codex's sub-agents do the work; results flow back into your session.
You don't manage the parallelism. Codex decides when to fan out based on the task. Ask it to "research 5 competitors and compare their pricing" and it may spawn a sub-agent per competitor, run them concurrently, and merge into a single deliverable — all from one /codex call.
This is especially powerful for:
- Broad research — scanning multiple sources, APIs, or datasets simultaneously
- Code generation at scale — building multiple modules or services in parallel
- Analysis pipelines — running independent computations and synthesizing results
- Anything embarrassingly parallel — if the subtasks don't depend on each other, Codex can run them all at once
The real power shows up when you use Claude Code as the planner and Codex as the executor. You can build a plan in Claude Code that deliberately farms out parts of the work to Codex:
- Use Claude Code to break down a project into components
- Handle the parts Claude Code is best at (architecture decisions, code review, orchestration)
- Dispatch the heavy autonomous work to Codex via
/codex - Review what Codex returns, iterate, and repeat
Think of it like a senior engineer (Claude Code) delegating tasks to a capable team (Codex + its sub-agents) and reviewing the output.
Codex has its own skill system at ~/.codex/skills/. You invoke them by prefixing the skill name with $ inside a /codex call. The skill name gets passed through verbatim — Codex handles the routing internally.
/codex $skill-name <arguments>
If you've built custom Codex skills (or installed community ones), they're immediately available through this bridge. No extra configuration on the Claude Code side.
Here are some real-world examples to help you get started:
/codex scaffold a FastAPI app with JWT auth, SQLAlchemy models, and Alembic migrations
Codex builds the full project structure autonomously and returns the result.
/codex research the top 5 LLM inference providers, compare their pricing per million tokens,
latency benchmarks, and supported model families. Return a markdown comparison table.
Codex may spawn a sub-agent per provider to research in parallel, then merge into a single table.
/codex $polymarket what are the highest-volume markets this week?
Routes directly to Codex's Polymarket skill for market analysis.
/codex build a CLI tool that fetches weather data from OpenWeatherMap
...Codex returns v1...
/codex now add a --forecast flag that shows 5-day outlook and write tests
Picks up the same session — Codex has full context from the first turn.
You (to Claude Code): I need to build a dashboard that shows real-time crypto prices,
a news feed, and portfolio tracking. Plan this out.
Claude Code: Here's the plan:
1. React app with three panels (prices, news, portfolio)
2. WebSocket connection to Binance for live prices
3. News aggregation from multiple RSS feeds
4. Portfolio state in localStorage with import/export
I'll set up the project structure and component architecture.
Let's have Codex build the WebSocket integration and news aggregator.
You: sounds good, farm out the websocket and news parts
You (to Codex): /codex build a WebSocket client module that connects to Binance's
stream API for BTC, ETH, SOL price feeds. Also build a news aggregator
that pulls from CoinDesk, The Block, and Decrypt RSS feeds. Return both
as standalone TypeScript modules with types exported.
Claude Code owns the architecture; Codex builds the modules. You review and integrate.
/codex status
Polls the active run and shows current state, progress, or final results.
/codex new analyze this repo's test coverage and suggest improvements
Clears the sticky session and starts a brand new Codex thread.
See CLAUDE.md for installation instructions — designed for Claude Code to walk you through setup.