codex-chatgpt-web: A ChatGPT Web Bridge for Codex That Runs ChatGPT as a Native Model

Coding agents ship locked to whichever model their maker picked. Codex runs OpenAI's own models by default, and switching to something else usually means switching apps entirely. A new open-source project, codex-chatgpt-web, takes a different approach: a ChatGPT web bridge for Codex that slots ChatGPT, including a Plus or Pro subscription, directly into Codex's native model picker. As of today, the repo sits at 865 stars and 102 forks, up from roughly 750 stars two and a half weeks after it was created on 2026-07-26.

What it actually does

The pitch is simple: keep using Codex's task history, tool harness, streaming UI, and approvals, but let one of the selectable models be ChatGPT itself, running on whatever plan you already pay for. Free and Go accounts get a model called Luna. Plus accounts get Instant through High reasoning tiers. Pro accounts additionally unlock Extra High and Pro. Each entry in the picker maps to a fixed ChatGPT mode, so toggling Codex's effort or speed settings does not silently swap the underlying model.

For anyone already juggling Claude Code, Codex, and a ChatGPT subscription, this closes a real gap: it lets a specific model's strengths get used without leaving the coding agent that's already wired into the project.

How the bridge actually works

This is the part worth reading carefully before installing anything. codex-chatgpt-web is not an API wrapper. It is browser automation, built on Playwright, that drives a real, signed-in ChatGPT session inside a private embedded browser.

Setup works like this: the launcher opens a dedicated system Chrome or Chromium window for sign-in, including passkey support. Once it detects an authenticated ChatGPT composer, it captures that session, imports an allowlisted slice of it into its own embedded-browser profile, and independently verifies the composer is authenticated there too. From then on, every model turn opens a fresh ChatGPT Temporary Chat, gets the current compiled Codex task context (text and images) fed into it, and streams the visible reasoning, tool activity, and Markdown output back into the Codex task.

Temporary Chat is a privacy mode, not an anonymity layer or local inference. Prompts still get processed by OpenAI under the account's normal settings, and the project's own README says so plainly rather than glossing over it.

Browser-only versus full harness

There are two operating modes. Browser-only just routes model turns through ChatGPT with no local tool access, and Codex shows a warning in that state. Full harness is more ambitious: it connects ChatGPT's tool calls back to the actual Codex task's filesystem, shell, and configured tools over MCP, using OpenAI's own tunnel-client. The tunnel is outbound only, so it does not open an inbound port or need router configuration.

Full mode requires creating a new ChatGPT connector, naming it exactly "Codex Native2," and setting its permissions to "Allow all actions." The README is unusually specific here: reusing or renaming an older connector, or picking the lower "Allow low-risk actions" tier, silently blocks commands and patches from reaching Codex. That level of precision in the setup instructions is a decent signal the maintainer has actually hit these failure modes.

Pro is the one exception across both modes. ChatGPT's Pro tier does not currently expose the custom MCP connector the bridge needs, so Pro can read the compiled Codex context and reason over it, including its native web search, but it cannot call local tools. Everything runs subject to what the README calls a "measured browser ceiling," with automatic compaction when a task's context runs past what the ChatGPT composer can hold in one turn.

The catch: this is unofficial, and it can break

Be direct about this: codex-chatgpt-web works by automating ChatGPT's web interface, not by calling a supported API. The project says so itself, describing it as "unofficial browser automation, not an OpenAI API," and warns that UI changes on OpenAI's side can break its selectors. To its credit, the failure mode is explicit rather than silent. A drifted selector produces an error, not a quiet fallback to a different model or a garbled response, which is the right way to fail but doesn't remove the underlying fragility. Anything OpenAI changes in the ChatGPT composer, sign-in flow, or Temporary Chat behavior can knock this out until the maintainer patches it.

There is also a standing terms-of-use question that no amount of good engineering resolves. Driving a real ChatGPT session through browser automation, rather than the API, sits in a gray area that OpenAI could tighten at any point. The README addresses this directly: users "remain responsible for complying with applicable OpenAI terms and workspace policies," and the project states it does not bypass authentication or access controls. That framing is honest, but it does not eliminate the risk that an account gets flagged if usage patterns look automated. Anyone running this against a primary ChatGPT account, especially a paid Pro subscription tied to other work, should treat that as a real possibility rather than a footnote.

One more practical point: browser session state is a sensitive login artifact. The project's own docs flag that the local listener is reachable by other processes running as the same user, and explicitly warn against sharing the launcher profile.

Setup and platform support

Installation is a one-line script for macOS or Linux, a PowerShell command for Windows, or building from source with Bun 1.3.14. Release packages target macOS 13 and up on arm64 and x64, Windows x64, and Linux x64. The maintainer notes the browser flow has been manually exercised end-to-end only on macOS and Windows 11 so far, with all three platforms gated in CI. It is MIT-licensed, written in TypeScript, and has 16 open issues at the time of writing, one of which specifically tracks the ChatGPT composer's context-window ceiling versus the underlying model's real limits.

Who this is actually for

This fits a narrow but real use case: developers already paying for ChatGPT Plus or Pro who want that model's strengths available inside Codex without paying twice, once for a subscription and again for API tokens. It is not a fit for anyone who needs guaranteed uptime, a stable integration surface, or a setup with no exposure to OpenAI's discretion over automated access. Given how new the project is, created less than three weeks ago and still gaining stars quickly, it is worth watching how it holds up as ChatGPT's web interface changes and as the terms-of-use question either gets addressed or doesn't.