OpenAI Codex

Codex is OpenAI’s coding agent, accessed from inside ChatGPT. It works nothing like a completion tool. You connect it to a GitHub repository and describe a task in a sentence — fix a bug, add a set of tests, refactor a module a particular way — and it checks out the code in an isolated cloud environment, makes the changes, runs the tests, and hands back a diff along with a log of what it did.

It is asynchronous and parallel: dispatch several tasks, go do something else, then review each result and open pull requests for the ones you like. It feels less like an assistant typing alongside you and more like a few junior colleagues working on their own. Beyond the cloud version, OpenAI also offers the open-source Codex CLI and editor extensions for working in a local codebase.

The limits are just as clear. The cloud environment needs your dependencies and tests set up properly, or it cannot run anything. Given a vague task, it will confidently deliver a change that heads in the wrong direction. And for users in mainland China, simply getting access is the first hurdle.

Key features

• Cloud task execution: checks out the repo in an isolated container, edits code, runs commands and tests.
• Parallel tasks: dispatch several tasks that run independently of each other.
• Change review: each task returns a diff and execution log; ask follow-ups or open a pull request directly.
• Codebase Q&A: have it read the repository and explain a module or call chain without changing anything.
• AGENTS.md conventions: a file in the repo tells it how to run tests and which standards to follow.
• Local tools: the open-source Codex CLI and editor extensions work directly in a local codebase.

How to use

1. Sign in to ChatGPT and open the Codex page.
2. Connect your GitHub account and grant access to the repositories you need.
3. Configure the environment for the repo so dependencies install and tests run.
4. Describe tasks in specific, verifiable terms; you can dispatch several at once.
5. Review each result and its log, open a pull request when satisfied, and do your usual code review before merging.

Best for

• A backlog of small jobs: minor bugs, missing tests, dependency bumps, all handled in parallel.
• Projects with good test coverage: it can verify its own work, so results are more reliable.
• Unfamiliar repositories: have it read the code and explain the structure before deciding what to change.
• Teams that collaborate through pull requests: its output fits naturally into review.

Strengths and limitations

Strengths

• Asynchronous and parallel, so several tasks move forward without tying up your editor.
• Runs tests in an isolated environment and checks its own work before handing it over.
• Output is a standard diff or pull request, easy to fit into existing workflows.

Limitations

• Mainland China is not a supported region for OpenAI; direct access does not work.
• The cloud environment needs setting up first, which is costly for projects with complex dependencies.
• Vague task descriptions produce changes that look complete but go the wrong way.
• Mainly offered with paid subscriptions, and usage limits bite under heavy use.