Codex with ChatGPT: a Codex skill and local MCP bridge that lets the ChatGPT web app plan and review while Codex executes

The pitch is idle web quota instead of API tokens, and it works by driving the ChatGPT web app

The problem statement is that the web quota of a paid ChatGPT subscription sits idle while the coding agent burns scarce API and Codex tokens on planning and review. The project moves the thinking to the subscription you already pay for and leaves execution to Codex. It says there are no API keys and no reverse proxy, only the official web interface plus a read-only MCP bridge.

The consequence is a dependency on a user interface, not an API. The control plane runs through Computer Use, so Codex operates the ChatGPT web app on your behalf. Changes to that page can break the flow, which explains why the README opens with a prominent note: when something goes wrong, first ask Codex to update Codex with ChatGPT and retry, because updating resolves most known issues. Whether automating the web app suits your account's terms is a question you should check yourself.

The control plane exchanges tiny state messages, while the data plane serves ten read-only tools

Two channels carry the work. The control plane uses Computer Use, and Codex and ChatGPT exchange small structured messages tagged C2C that move through the states INIT, PLAN, EXECUTED, REVIEW and DONE. Messages are under one kilobyte, and no diffs, logs or file bodies are pasted in them. The data plane is MCP: ChatGPT pulls what it needs through ten tools, namely workspace_info, list_directory, read_file, search_workspace, git_status, git_diff, test_status, execution_summary, execution_output and read_image.

All ten are read-only. The result is that ChatGPT can look but not change anything, and after Codex executes, ChatGPT inspects the actual git diff and test records, so it does not rely on a claim that all tests passed. The limit is that ChatGPT sees whatever those tools expose from your workspace on demand.

Your repository is not uploaded, but ChatGPT reads lines through a tunnel you expose

The README states that the repository is never uploaded: ChatGPT reads exactly the lines it needs over a secure, OAuth-protected, read-only MCP connection to your current workspace. The bridge is a loopback-only HTTP server using OAuth 2.1 and a one-time pairing code, reached through a Cloudflare tunnel, and credentials stay in the operating system's app state directory instead of the project.

The distinction matters. Not uploaded does not mean not transmitted: the lines ChatGPT requests do travel to ChatGPT, and the bridge is reachable from the internet through the tunnel, protected by OAuth and pairing. If your code is confidential, check your organisation's rules for using ChatGPT with it. The examples directory includes a c2cignore example file, which is where you list paths the bridge should not serve.

The default public address is a temporary Cloudflare URL that changes on restart

The default address is a temporary Cloudflare URL. It changes whenever the bridge restarts, and Codex repairs the situation by deleting that workspace's connector in ChatGPT and adding it again. The optional fix is a stable hostname such as c2c-project.your-domain.com, which needs a Cloudflare account and a domain already on Cloudflare, and opens a browser once so you can authorise it. After that, the ChatGPT connector keeps working across restarts.

If you skip it or the login fails, Codex stays on the temporary address with the same features but a slower repair. So the cost of the free path is repeated reconnection after every restart. If you restart your machine often, set up the stable hostname once. The installer also needs the cloudflared program, which the one-paste install prompt installs alongside git and Node.js.

Installation is a pasted prompt or a short manual path, and it needs Node 20

There are two install routes. The one-paste route gives a non-technical user a paragraph to hand to Codex, which checks for git and Node.js 20 or newer, installs what is missing with Homebrew on macOS or winget on Windows, clones the repository to a folder in the home directory, builds with corepack pnpm install and corepack pnpm build, copies the skill file into the Codex skills folder and runs first-time configuration. The manual route is three steps: copy the skill folder into the Codex home skills directory, tell Codex to set up Codex with ChatGPT, then ask it to implement things.

A successful setup shows a checklist ending in Ready:

Codex with ChatGPT

✓ Project detected
✓ Workspace Bridge started
✓ Secure connection established
✓ ChatGPT connected
✓ File read test passed

Ready.

The project clarifies its installation scope: it does not publish or install a Codex Web GPT, launcher or model-catalog entry. It builds the checkout, installs the skill and runs c2c setup. Codex home is CODEX_HOME when set and non-empty, otherwise the .codex folder in your home directory.

A new workspace needs a ChatGPT Project once, with project-only memory

The only steps that may need you are logging into ChatGPT and, for a stable hostname, logging into Cloudflare once. A new workspace also asks you to create a ChatGPT Project, called a collection, once. You pick project-only memory and name it after the workspace. If the sidebar has no Projects row, you hover over Chats, open the menu and choose Organize by project. Codex then saves the collection link and starts chats from that page.

Existing workspaces that already have a C2C chat stay on the old one-conversation style until you ask to switch. The practical consequence is one ChatGPT Project per workspace, so planning history is kept apart per repository. The skill also checks GitHub once a day and updates itself when a new version is released, so behaviour can change without a manual step.

Version 0.1.x ships quickly, and package.json lags the newest tag

The recent releases are v0.1.4 on 2026-10-01, v0.1.3 on 2026-09-11 and v0.1.2 on 2026-09-04, with titles such as Executor Records, Tunnel DNS and Media Workflows, Connector Recovery with hidden git_status and HTTP/2, and Structured MCP with safer tunnels and Windows polish. The last push was on 2026-10-01. Yet package.json still reads version 0.1.3, one behind the newest release.

The zero major version and the daily self-update mean you are tracking early software. The package exposes a command named c2c, needs Node 20 or newer, and uses pnpm 11.24.0 with express, commander, zod and the MCP SDK, the last of which is listed as latest and not pinned. Read each release note before relying on a workflow, and keep your own copy of anything you cannot afford to have change.

Editorial conclusion

Codex with ChatGPT is worth trying if you pay for ChatGPT, already work in Codex, and want the web app to do planning and review while Codex edits files. It is not for anyone who needs a supported, API-based setup, because it depends on driving the ChatGPT web interface and on a public tunnel to your machine. Before using it, review what the ten read-only tools can see in your repository, set an ignore file, and expect to update it when the ChatGPT interface changes.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. XiaoDuoYa/codex-with-chatgpt on GitHub