Codex with ChatGPT Splits Planning From Execution

Codex with ChatGPT is a GitHub project that separates the planning and code review processes from the execution of code. This is achieved by using ChatGPT for reasoning, while Codex handles file editing, shell commands, and test execution. The project aims to leverage an existing ChatGPT web subscription for its reasoning capabilities without replacing Codex's execution harness.

A read-only handoff is established between the two agents, with Codex and ChatGPT exchanging structured state messages for the plan, execution, and review loop. ChatGPT retrieves repository context through nine read-only MCP tools, such as file reads, workspace search, Git status and diffs, test-status, and execution-output records. The separation of planning and execution is crucial for the security design of the project.

The bridge has no write, delete, shell, or commit tools, and sensitive paths like .env files, keys, SSH material, and credentials are denied by default. Workspace-scoped access, OAuth 2.1 protection, and one-time pairing codes are used for the publicly reachable MCP endpoint to enhance security. The README documents the project's claims regarding security.

Setting up the integration requires operational costs, as it is packaged as a Codex Skill. Users must manually copy the skill into Codex's skills directory and initiate the first-time setup. The project requires Git, Node.js 20 or later, and cloudflared for its public connection. Users may need to sign in to ChatGPT and authorize Cloudflare for an optional stable hostname, which can help avoid the need to repair the ChatGPT connector after each bridge restart.

While the integration offers architectural benefits by allowing a web-based planner to access selected live workspace evidence while keeping side effects within the coding harness, it also introduces additional components like a tunnel, OAuth pairing, and connector configuration. These additions create another failure boundary in a development workflow.

The read-only policy may limit ChatGPT's direct capabilities, but developers must still determine if the allowed source context is appropriate to expose through the bridge. The project is an unofficial community effort and not endorsed by OpenAI.