XiaoDuoYa/codex-with-chatgpt

ChatGPT thinks. Codex works. Use ChatGPT as the planning brain while keeping the Codex harness.

Codex with ChatGPT – What it is

Codex with ChatGPT is a bridge that lets the official ChatGPT web UI act as the thinking (planning, reviewing) part of a coding‑assistant workflow while the Codex agent does the actual code execution. It does not require an OpenAI API key – the ChatGPT side runs in a normal browser session, and a tiny read‑only MCP (Microsoft Code Protocol‑style) bridge forwards only the few lines of code that the model needs to see.

Why it exists

ChatGPT Plus/Pro web subscriptions give you a lot of unused quota, while Codex‑based agents consume scarce API tokens for planning and review.
This tool swaps the “thinking” to the paid‑for web UI and keeps the execution on Codex, saving API usage.
It works without uploading your repository: ChatGPT reads files through a secure, OAuth‑protected, read‑only bridge.

How you get it running (one‑paste install)

Prerequisites – git, Node ≥ 20, and cloudflared (installed automatically via Homebrew on macOS or winget on Windows).
Clone & build – git clone https://github.com/XiaoDuoYa/codex-with-chatgpt ~/codex-with-chatgpt && cd ~/codex-with-chatgpt && corepack pnpm install && corepack pnpm build.
Install the Codex Skill – copy skill/SKILL.md into your Codex skill directory ($CODEX_HOME/skills/codex-with-chatgpt/).
First‑time setup – run c2c setup. A built‑in browser opens, you log into ChatGPT, pair the session with a short code, and (optionally) link a Cloudflare tunnel for a stable hostname.
Use – tell Codex “Set up Codex with ChatGPT” and then “Use Codex with ChatGPT to implement …”.

The whole flow is designed to be automated; a non‑technical user can just paste the block above and let the agent do the rest.

Core architecture (simplified)

ChatGPT Web UI  ←→  C2C Bridge (loopback HTTP + Cloudflare tunnel)
   ▲                     ▲
   │   read‑only MCP (10 tools)   │
   │                     │
   └─> Local workspace ←─ Codex harness (executes, runs tests)

Control plane – tiny [C2C] messages (INIT → PLAN → EXECUTED → REVIEW → DONE).
Data plane (MCP) – read‑only tools such as list_directory, read_file, git_status, test_status, etc.
Security – OAuth 2.1 with PKCE, one‑time pairing code, path containment, and a default deny list for secrets (.env*, keys, etc.).

What you can do with it

Plan – ask ChatGPT to design a feature; it can inspect the current repo via the read‑only tools.
Execute – Codex runs the generated code, commits, runs tests.
Review – after execution, ChatGPT reads the actual git diff and test results to give an independent review.
Asset handling – read‑only image/video inspection; you can import verified assets with c2c asset import.

Limitations & status

Read‑only: the bridge never writes to your files; all modifications come from Codex.
Unofficial: community project, not endorsed by OpenAI.
V1: end‑to‑end verified, but still early – expect occasional bugs; the skill auto‑updates daily.

License

MIT – you can freely use, modify, and redistribute.

In short: Codex with ChatGPT lets you keep the cheap, quota‑free ChatGPT web UI as the brain of your coding assistant while Codex does the heavy lifting, all behind a secure, read‑only bridge that protects your code and credentials.*