Codex with ChatGPT v0.1.2 is an unofficial, MIT-licensed bridge that turns a ChatGPT web subscription into the planner while Codex keeps execution. For individual developers and small teams on non-sensitive code who want planning and review off paid Codex tokens, it is worth a controlled pilot. Treat it as pre-1.0 community software, audit what ChatGPT can read, and expect tunnel and browser-automation fixes to continue.

This is a source and documentation review of the GitHub repository, its release notes, and its open issues at the 9-10 September 2026 snapshot. No local installation or runtime test was performed, so claims from the project are attributed to its maintainers and contributors rather than independently verified here. The decision this review supports is whether you should adopt this bridge now to move planning and review off Codex tokens, and what you must verify before relying on it for real work.

Key facts at a glance

What Codex with ChatGPT does and why it exists

Codex with ChatGPT (C2C) connects a ChatGPT web subscription to a Codex coding session so the two models split a job. According to the README, ChatGPT owns planning, review, and debugging strategy, while Codex keeps full ownership of execution: editing files, running shell commands and tests, and managing Git. The maintainers’ stated motivation is that ChatGPT Plus or Pro web quota sits idle while Codex or API tokens are spent on planning and review; moving that reasoning to the web subscription is the point of the project.

Technically, it is not an API bridge. The README and release notes state there is no API key, no reverse proxy, and no cookie or session extraction. Instead, a local C2C bridge runs an HTTP server on loopback and exposes nine read-only MCP tools for workspace inspection, Git state, test status, and execution output. It reaches ChatGPT through a Cloudflare Quick or Named Tunnel. Access is protected with OAuth 2.1 (PKCE, dynamic client registration, rotating refresh tokens) and one-time pairing codes with a short TTL. The README claims the repository is never uploaded wholesale: ChatGPT pulls only what it needs, and sensitive file types such as keys, .env files, and credentials are denied by default, with .c2cignore available for extra rules.

Those are design claims from the repository, not conclusions from an independent run. The project is very young: the repository was created on 28 August 2026 and the first public release followed on 30 August. The README describes C2C as an unofficial community project not affiliated with or endorsed by OpenAI.

What v0.1.2 changed

The release under review, v0.1.2, is a reliability release. Its notes describe three themes: structured MCP outputs, safer Windows Quick Tunnel startup, and quieter background behavior.

• Structured output. The read-only MCP tools now advertise output schemas and return matching structuredContent, while keeping the previous text JSON for compatibility. Execution records are validated on write, malformed historical records are skipped on read, workspace metadata is sanitized, and git diff pagination documentation now matches the real hasMore and nextOffset fields.
• Safer Windows tunnel startup. Quick Tunnel startup now fails closed: it only reports success after the public /health endpoint confirms a healthy C2C bridge, rejects invalid or reserved api.trycloudflare.com URLs, and treats early cloudflared exits, bad health responses, and timeouts as failures. A custom cloudflared path can be set with C2C_CLOUDFLARED_PATH.
• Quieter Windows operation. Release notes say bridge and tunnel daemons no longer open or flash console windows during normal use.
• Smarter diagnostics. The doctor command no longer assumes an inconclusive local probe means the bridge process is dead, so it avoids starting a second bridge or deleting the ChatGPT connector unnecessarily.

The release notes report 164 tests passing after the structured-output hardening, along with a successful typecheck and build. This review did not run those tests. The notes also state that no new MCP permissions were introduced and no reverse proxy, cookie, session-extraction, or credential-sharing behavior was added.

Version cadence is itself a maintenance signal: v0.1.0 shipped on 30 August 2026, v0.1.1 on 31 August, and v0.1.2 on 4 September. At this stage the project is changing almost daily, and the installed Skill checks GitHub once a day and updates itself per the README.

Who should evaluate it now

A pilot makes sense if you already pay for ChatGPT web, use Codex often enough that planning and review consume real quota, work on code you can afford to expose to an AI web app, and run on a supported platform with Codex’s in-app browser available. Single developers, consultants, and small teams testing agent workflows on their own machines fit best. The most natural early use is letting ChatGPT review actual git diffs and test output instead of asking someone to paste context manually.

Whether C2C saves money depends on your usage split, which the evidence does not measure. If your ChatGPT web quota is small relative to your Codex or API spend, the bridge changes where reasoning happens but not necessarily what it costs you.

What to check before production

Treat C2C as promising but not production-default. The concerns are not about code quality alone; they combine early-stage tooling, a public tunnel into a workspace, and automation of the ChatGPT web interface.

Data exposure. The MCP layer is read-only, but read-only still means file contents and git diffs leave your machine over a Cloudflare tunnel to OpenAI’s web app. Before connecting proprietary code, read the project’s security documentation, list every path ChatGPT can reach, tighten .c2cignore beyond the defaults, and decide whether web-app automation is acceptable under OpenAI’s terms and your own data policy.

Platform coverage. The v0.1.0 and v0.1.1 release notes state that the fully automated workflow targets environments where Codex has access to the ChatGPT in-app browser, and that combinations such as Linux plus the VS Code Codex extension are not yet guaranteed to support the complete browser-driven workflow. Verify your exact setup before relying on the one-paste install.

Connection reliability. The repository had 17 open issues at the 9 September 2026 snapshot, several of them about exactly the failures that hurt daily use:

• Issue #424 reports that after a Codex restart on macOS the previous temporary Quick Tunnel address stops working, and C2C fails to accept a newly issued tunnel as ready even though cloudflared connects (status reports fetch failed and a 45-second timeout).
• Issue #417 reports Named Tunnel provisioning can succeed while an existing stable hostname still routes to a different tunnel ID, leaving the public endpoint returning Cloudflare 530 / error 1033.
• Issue #224 reports a recovered temporary address whose ChatGPT connector state no longer matches, so workspace reads fail.
• Issue #419 reports a local checkpoint can wait for a ChatGPT message that was never submitted after project navigation hits a retry-only page.
• Issue #193 reports that the same ChatGPT account on two computers for the same workspace cannot create a second connector because the default connector name already exists.

Not every report identifies its version, and v0.1.2 targets some of these areas, so treat them as patterns to test in your environment rather than confirmed defects in the current release.

Change velocity and updates. An auto-updating Skill is convenient for a personal laptop and a governance problem for a team. If you depend on C2C, pin a version, review release notes before updating, and keep a tested rollback path. The project’s adoption (3,694 stars in under two weeks) signals interest, not maturity.

Where custom development and maintenance will be needed

Most of the work needed to run this beyond one laptop is integration and hardening, not feature writing:

• Stable connectivity. The default is a temporary Cloudflare URL that changes when the bridge restarts, with an optional named tunnel if you own a domain on Cloudflare. That lifecycle is where the open issues concentrate, and production use will need monitoring, restart handling, and a documented recovery procedure.
• Team support. One connector per workspace with no per-machine naming means the same workspace used from multiple computers with one ChatGPT account is not cleanly supported (issue #193). A shared always-on endpoint is mentioned as an option in the discussion, not shipped.
• Browser-automation drift. The workflow drives the ChatGPT web app: login, developer-mode toggles, connector creation, project navigation. When OpenAI changes that interface, parts of C2C will need fixing by whoever maintains your installation.
• Independent security review. The OAuth, pairing, path-containment, and sensitive-file design is described in detail, but only the maintainers and contributors claim it behaves as documented. A production decision warrants code review of the bridge, the tunnel configuration, and the file policy before real repositories are connected.

Who should wait

• Organizations with proprietary code, regulated data, or contractual restrictions on code leaving their environment, until they independently verify what ChatGPT can read and over what network path.
• Teams that need shared, always-on agent infrastructure with a support commitment rather than per-machine tooling.
• Developers on platform combinations the maintainers have not guaranteed, notably Linux with the VS Code Codex extension.
• Anyone with occasional Codex usage rather than steady daily spend; the setup and maintenance overhead only pays off if the workflow is used regularly.

Alternatives and evaluation questions

Alternatives depend on where your cost actually sits. If planning is the expensive part, compare C2C with plain Codex sessions that plan themselves, doing design and review in ChatGPT web yourself and handing Codex concise instructions, or an API-based setup that routes planning to the model you choose under your own control. If you need a documented, auditable connection from a language model to a workspace, a purpose-built bridge you control may be simpler than tracking an evolving community project.

Before installing anything, get answers to these questions:

• Which files could ChatGPT read in my repository, and is that acceptable outside my network?
• Where are credentials stored, and who can rotate or revoke them? The README says the OS app-state directory, which you should confirm in your environment.
• What happens to a running task when the tunnel drops, the machine sleeps, or the ChatGPT session changes? Open issues suggest reconnection is not seamless yet.
• Does my usage actually match the project’s assumption: generous ChatGPT web quota and expensive Codex or API planning?
• Is automated browser control of ChatGPT acceptable under my plan and my company’s policy?
• What is my exit path if development stalls or OpenAI changes the web app?

Verdict

Codex with ChatGPT v0.1.2 is worth a controlled pilot for individual developers and small teams that already pay for ChatGPT web, work on non-sensitive code, and can tolerate pre-1.0 connection behavior. It is not yet a default production tool: the security model needs independent review, tunnel and connector failures remain open at this snapshot, and platform support is narrower than the one-paste install suggests. Reconsider it for wider use once those issues settle and you have measured whether it actually lowers your billed usage.

Help with the decision and rollout

If you are deciding whether a bridge like this belongs in your workflow, I can help with the parts community software leaves to you: a pre-adoption review of exactly what ChatGPT could read in your workspace, a version-pinned deployment with a stable tunnel and rollback plan, or maintenance when OpenAI’s web app changes and breaks the automation. To do that I would need access to the repository and the environment you intend to connect, plus details of your Codex and ChatGPT setup; I cannot promise a specific cost or reliability outcome before seeing those. This kind of review, hardening, and AI agent integration is part of my software development services, and you can contact me about your setup.

Sources

• XiaoDuoYa/codex-with-chatgpt repository (GitHub)
• v0.1.2 release notes (GitHub)
• Issue #417: Named Tunnel provisioning can succeed while an existing hostname points to a different tunnel ID
• Issue #424: Quick Tunnel reconnect fails after Codex restart although cloudflared succeeds
• Issue #193: support per-computer connector names for the same workspace