---
name: Claude + Codex Model Router
category: coding
tools_required: [Claude Code, ChatGPT account (Free works) or OpenAI API key, Node.js 18.18+]
cost: uses-existing-subs
time_to_set_up: 10 minutes
version: 1.0
source: adapted from viral thread; setup verified against github.com/openai/codex-plugin-cc (official OpenAI plugin), July 2026
tested: no
---

# Claude + Codex Model Router

## What it does
Runs two AI coding agents in one terminal: Claude Code plans, reviews, and orchestrates; Codex executes heavy implementation. You spend your premium Claude tokens on thinking, not typing — and get a second model catching the first one's blind spots.

## What you need
- Claude Code installed and working
- A ChatGPT account (Free tier works) or OpenAI API key
- Node.js 18.18 or later
- Note: delegated work counts against your Codex usage limits

## Setup (one time)

1. Inside Claude Code, run:
```
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup
```
2. `/codex:setup` checks everything. If Codex CLI is missing it offers to install it via npm. If not signed in, run `!codex login` (ChatGPT account or API key).
3. Optional — pick Codex's default model per project by creating `.codex/config.toml` in the repo root:
```
model = "gpt-5.4-mini"
model_reasoning_effort = "high"
```

## The workflow (daily use)

Paste this once at the start of a session (or save it as a Claude Code skill so one command activates it):

```
You are the orchestrator. Division of labor:

YOU (Claude): planning, repo understanding, architecture decisions,
task decomposition, and final review of all delegated work.

CODEX (via /codex:rescue): heavy implementation, debugging, test
fixing, refactoring, multi-file edits.

Rules:
- Keep each Codex task focused and specific — one job per delegation.
- Run long Codex tasks with --background and keep working.
- When Codex finishes, inspect the result yourself before accepting.
  Do not blindly trust its output.
- For risky changes, run /codex:review as an adversarial second
  opinion before merging.
```

The rhythm: Claude plans → Codex builds → Claude reviews → you approve.

## Maintenance habits
- **Adversarial review before merges:** `/codex:review --base main --background` gives a skeptical second-model pass on branch changes.
- **Context hygiene:** long sessions degrade. Before clearing a bloated session, have Claude write a handoff note (task state, decisions made, what's left) to a file, then start fresh and load it.
- **Check jobs:** `/codex:status` shows running/recent Codex jobs, `/codex:result` fetches finished output.

## What we stripped
The source thread claimed "$1.86/task", "3x lower cost than Fable 5", "catches every bug", "60%+ fewer tokens", "never hit limits again". None of it is verifiable and some of it is impossible (nothing catches every bug). The workflow stands without the fantasy numbers — test it and record your own.