Codex with ChatGPT, a public GitHub project, routes planning and code review to the ChatGPT web app while retaining Codex as the agent that edits files, runs shell commands and executes tests. For developers, the practical aim is to use an existing ChatGPT web subscription for reasoning without replacing the Codex-based execution harness.
A read-only handoff between the two agents
The project describes a local C2C Bridge between ChatGPT and a workspace. Codex and ChatGPT exchange small structured state messages for the plan, execution and review loop, while ChatGPT retrieves repository context through nine read-only MCP tools, including file reads, workspace search, Git status and diffs, and test-status and execution-output records. The README says full file bodies, diffs and logs are not placed in those control-plane messages. See the project README.
That separation is central to its security design. The repository says the bridge has no write, delete, shell or commit tools, and that sensitive paths such as .env files, keys, SSH material and credentials are denied by default. It also describes workspace-scoped access, OAuth 2.1 protection and one-time pairing codes for the publicly reachable MCP endpoint. These are project claims rather than an independent security assessment. The stated security model is documented in the README.
Setup still carries operational cost
Installation is packaged as a Codex Skill, with a manual route that copies the skill into Codex’s skills directory and asks Codex to perform first-time setup. The project requires Git, Node.js 20 or later and cloudflared for its public connection; it says users may need to sign in to ChatGPT and, for an optional stable hostname, authorize Cloudflare and use a domain already managed there. The installation and hostname flow are specified by the project.
The default connection uses a temporary Cloudflare URL, according to the README. When that address changes after a bridge restart, the project says Codex repairs the ChatGPT connector; a stable Cloudflare hostname is presented as an optional way to avoid that repair cycle. The project describes the temporary and stable tunnel options here.
The unresolved trade-off
Analysis: This is a narrow integration pattern, not evidence that splitting planning from execution produces better code. Its appeal is architectural: it gives a web-based planner access to selected live workspace evidence while keeping side effects in the coding harness. But the arrangement also adds a tunnel, OAuth pairing, connector configuration and another failure boundary to a development workflow. The read-only policy can limit what ChatGPT can do directly, yet developers still need to decide whether the allowed source context is appropriate to expose through the bridge. The project is explicitly an unofficial community effort, not an OpenAI-endorsed Codex integration. Its status and disclaimer are in the README.
Top comments (0)