Codex ChatGPT Subscription: Codex-chatgpt-web and Codex-with-chatgpt

**Both projects solve one pain: your ChatGPT subscription’s web quota sits idle while Codex burns expensive API/Codex tokens. But they approach it from different angles. WKVTXNTOKEN0X substitutes the ChatGPT Web (up to Pro) as the native model in the Codex picker. WKVTXNTOKEN1X makes ChatGPT a brain-planner, and Codex leaves the executor, linking them via read-only MCP. Both are unofficial browser automation under the MIT license; you are responsible for complying with OpenAI rules.

• **The general idea is to use an already paid ChatGPT subscription rather than paying for API tokens.
• codex-chatgpt-web: ChatGPT Web as a Codex model (Luna / Instant / Medium / High / Extra High / Pro).
• codex-with-chatgpt: "ChatGPT thinks Codex does" - scheduling and review on the ChatGPT side.
• Both: informal, MIT, working through your web session; UI ChatGPT changes - the bridge may break.

## The problem is why it even appeared

Codex can log in to your ChatGPT plan - it's official. But subscriptions have usage limits, and heavy models and large context through paid APIs cost money. The web version of ChatGPT that you are already paying for is often idle.

Hence the idea: to use this idle web quota in the Codex work cycle. Then the fork begins – what exactly** should ChatGPT do? And here the two projects differ fundamentally.

## Two different approaches

This is a key idea that is easy to confuse: projects are similar in name, but do different things.

• codex-chatgpt-web - model substitution ChatGPT Web appears right in the picker of Codex models, and the entire course of the task is driven through the built-in browser to ChatGPT. It's like you've chosen another model.
• *codex-with-chatgpt - the division of labor. Codex remains the executor (corrects files, runs tests), and ChatGPT connects as a “brain”: plans and makes an independent review, reading the code through a secure read-only channel.

The following is detailed for each, on data from their README.

## codex-chatgpt-web (miuyyy): ChatGPT Web as native codex models

Positioning directly from the repository: “Use ChatGPT Web (including Pro) as native Codex models.” You change the model range, you keep your workflow.

How it works:

• Model in native picker Codex. Free/Go-accounts see the item "ChatGPT Web - Luna". Accounts with a reasoning level selector receive Instant, Medium, High, Extra High and Pro – as far as subscription allows.
• ** Built-in browser.** Cross-platform launcher (macOS arm64/x64, Windows x64, Linux x64) has input, built-in private ChatGPT browser, model configuration and health checks. No API key is needed for model moves.
• The bridge sends the compiled context of the Codex problem to the task-bound Temporary Chat, attaches pictures and streams back reasoning, tool activity, and Markdown to the same task.
• Two modes. Browser-only - model only, local Codex tools are not available (Codex will show a warning). Full harness - through the official openai/tunnel-client ChatGPT gets access to the tools of the current task (files, shell, pictures, apruvas). Outgoing tunnel: does not open the incoming port and does not require a pass.
• Fail-closed. Missing models/tools or changed ChatGPT UIs are a clear error, not a quiet switch of route or features.

Installation - one team downloads the desktop launcher:

# macOS / Linux
curl -fsSL https://github.com/miuuyy/codex-chatgpt-web/releases/latest/download/install-launcher.sh | sh

Next in the application: sign in to ChatGPT in the built-in browser → smoke-test → “Install models” → restart Codex and select the model “ChatGPT Web — ...”. There is also a launch from the source on Bun 1.4.0.

An honest frame from the author himself: this is unofficial browser automation, not the OpenAI API; Temporary Chat is a privacy mode, not anonymity and not local inference (prompts are still handled by OpenAI). The disclaimer explicitly states that the tool ** does not bypass authentication and access control**, and the responsibility for compliance with the Terms of Use is on the user. MIT license.

## codex-with-chatgpt (XiaoDuoYa): ChatGPT - brain, Codex - hands

ChatGPT Thinks – Codex Does. The idea is different: not to replace the model, but to transfer planning and revision to ChatGPT, leaving the execution to Codex.

How it works:

• Read-only MCP bridge. Repository never loads in its entirety. ChatGPT, through a secure OAuth connection, reads exactly what it needs on request with eight read-only tools: workspace_info, list_directory, read_file, search_workspace, git_status, git_diff, test_status, WKNTXNTOKEN7X.
• Two communication plans.* Control plane (Computer Use) are tiny structured state messages of [C2C]: INIT → PLAN → EXECUTED → REVIEW → DONE, with no diffs, logs, or file bodies. Data plane (MCP) - ChatGPT itself pulls up the desired.
• Independent review. After Codex has completed the task, ChatGPT watches the actual git diff and MCP test recordings – rather than taking the word “all tests passed.”.
• Security model. Read-only by construction: there are no write/delete/shell/commit tools on the server. One Workspace – One Border (Canonical Realpaths vs. Simlink Escape and ../). Sensitive files (.env*, keys, SSH, creds) are locked by default, plus its .c2cignore. Public MCP-endpoint requires OAuth 2.1 (PKCE S256, dynamic customer registration, rotation of refresh-tokens); without a token - 401, someone else's workspace - 403. The model never sees long-lived secrets - only a one-time pairing code (TTL 5 minutes, 5 attempts).
• Stack and launch. c2c CLI + Codex Skill, tunnel - Cloudflare Quick Tunnel. Requirements: Node.js ≥ 20, git, cloudflared (set automatically). There is a “one-paragraph installation”: copy the ready-made prompt to the agent, and Codex itself unfolds everything. 76 tests in CI. MIT license.

It's essentially a neatly designed bridge, where safety is part of the architecture, not a prefix. The author also emphasizes that the project is unofficial and not associated with OpenAI.

## How different are they

## What to choose

• Want to just "another model" in Codex and ready for browser automation - see codex-chatgpt-web. Especially if you have a Pro and want its level of reasoning in the terminal.
• Want a division of labor with a focus on security (ChatGPT plans and reviews, code does not leave the machine, all read-only) - see codex-with-chatgpt.
• Not sure – start with the one whose installation model you understand better: launcher application (first) versus CLI + Skill (second).

## Risks and honest reservations

• Both are unofficial. Neither is made by OpenAI. Both automate your ChatGPT web session, and both explicitly ask to comply with Terms of Use.
• Fragility. ** This is browser automation: updating the ChatGPT interface can break selectors. The plus is that both projects were designed fail-closed - when a failure is a clear error, not a quiet substitution.
• Session security. Browser status and local listeners are sensitive artifacts. Do not run on someone else's car, do not share a profile. The second project has a stricter security model (read-only, OAuth, block of secrets).
• Not bypass protection. codex-chatgpt-web explicitly states that it does not bypass authentication and access control; codex-with-chatgpt is read-only and does not unload the repository. These are not tools for breaking limits head-on, but ways to tap into a paid web quota.
• Product - with caution. For stable automation and CI, the official API is more reliable. These bridges are for personal work and experimentation.

## Frequent questions

Are both repositories alive? ** Yeah. Both are available on GitHub, give away HEAD, and contain full READMEs with installation, architecture, and disclamers. Both are licensed by MIT.

Are they the same thing? ** No. codex-chatgpt-web substitutes ChatGPT as a model in Codex. codex-with-chatgpt makes ChatGPT a scheduler-reviewer, and executes Codex. Different architecture and different scenarios.

Do you need an API key? ** For model moves in codex-chatgpt-web - no (work goes through the built-in browser). In codex-with-chatgpt, code data does not go through the API, but through read-only MCP; keys and reverse proxy are not used.

**Can they ban the account? ** No one has direct guarantees. Both projects are informal and ask for compliance with OpenAI rules. Use your account consciously and read their disclaimers.

What's safer for code? ** codex-with-chatgpt: read-only by design, repository not unloaded, secrets (.env, keys) blocked, access under OAuth 2.1.

## Conclusion

codex-chatgpt-web and codex-with-chatgpt are two working answers to the same question “how not to burn the API when a web subscription is idle”, but with different philosophies. The first turns ChatGPT Web into a native Codex model and takes over the entire move through the built-in browser. The second builds a neat division of labor: ChatGPT thinks and reviews through secure read-only MCP, Codex executes. Both are unofficial Browser Automation under MIT, both are honest about the Terms of Use.

The choice is according to the scenario: you need a model in the Codex picker - the first; you need a safe ChatGPT brain + Codex hands - the second. Both should be tried as an experiment, and not as a foundation of a product infrastructure.

## What to read next

• /goal в Codex и Claude Code: почему промпт должен быть коротким, а роадмап — подробнымXX
• Spec-driven development: как писать ТЗ, чтобы ИИ-агент не додумывалXX
• Ox Alpha оказался GLM-5.3-Flash: как стелс-модель Z.ai возглавила чарты OpenRouterXX