Codex with ChatGPT is an open source bridge for developers who want ChatGPT to plan and review code work while Codex keeps control of execution. The project targets a practical mismatch: many users already pay for ChatGPT Plus or Pro in the web app, but their coding harness still spends scarce API or Codex quota on planning, review, and context reading. Codex with ChatGPT moves that thinking loop into the official ChatGPT web UI and gives it a narrow, read-only view of the local workspace.
The tool has two planes. The control plane uses small structured Computer Use messages so Codex and ChatGPT can coordinate states such as init, plan, executed, review, and done. The data plane is a read-only MCP server that exposes workspace information, directory listing, file reading, search, git status, git diff, test status, and execution summaries. ChatGPT can inspect the files and diffs it needs, but it does not get write, delete, shell, or commit tools. Codex remains the actor that changes files, runs tests, and applies fixes.
Setup is designed for non-experts. The README includes a one-paste instruction that asks Codex to check for git, Node.js 20 or newer, and cloudflared, clone the repository, run `corepack pnpm install` and `corepack pnpm build`, install the Codex skill, and walk through the first-time connector pairing in the built-in browser. The CLI includes commands such as `c2c setup`, `c2c status`, `c2c doctor`, `c2c pair`, `c2c unpair`, `c2c logs`, and `c2c stop`.
Security is the main reason this is worth a dedicated listing. The MCP server is read-only by construction, tokens are bound to one workspace, path containment blocks symlink and traversal escapes, sensitive files such as `.env` and private keys are denied by default, and public access requires OAuth 2.1 with PKCE plus a short-lived pairing code. ChatGPT sees only the workspace data it requests through those tools, not long-lived credentials.
Codex with ChatGPT is best for builders who trust Codex for implementation but prefer ChatGPT’s web subscription for planning, architecture review, or final diff critique. It is not an official OpenAI product and it does not replace API-based coding agents. It is a community MIT-licensed harness for splitting reasoning and execution across two tools while keeping repository writes on the Codex side. The tradeoff is operational complexity: teams must run the bridge, tunnel, OAuth pairing, and skill installation correctly, then keep the connector limited to the intended workspace.