OpenAI Codex: what it is and what you can do with it

Most AI coding tools want you to sit next to them while they work. OpenAI Codex takes a different approach: you describe a task, hand it over, and the agent goes off to complete it, writing code, running tests and reporting back when it is done. It is less a tool you operate and more a colleague you delegate to. This post covers what it is, what you can do with it, and how to get started.

What is Codex?

Codex (openai.com/codex) is OpenAI's coding agent, powered by models trained specifically for software engineering work. It comes in several forms that all connect to the same brain:

Codex in the cloud. Tasks run in isolated cloud environments connected to your GitHub repository. You can fire off several tasks in parallel: one fixes a bug, another writes tests, a third drafts a feature, and you review the results as pull requests.
The Codex CLI. An open source command line agent that runs on your own machine, inside your own project. You type a request in the terminal, it reads your code, proposes changes and can execute commands, with you approving along the way.
The IDE extension. Codex inside VS Code and similar editors, so the agent works in the place you already write code.
Codex in ChatGPT. If you have a paid ChatGPT plan, Codex is reachable from the apps you already use, including handing tasks over from your phone.

Where it shines: existing code

Tools like Lovable and Bolt are at their best starting from zero. Codex is at its best when there is already a project. It is built to navigate large codebases, understand how things connect, and make targeted changes without breaking the rest. That makes it the natural choice for the second phase of any project: the bug fixing, the refactoring, the "add this feature without touching that one" work that makes up most of real software development.

Typical tasks people hand to Codex:

Fix this bug. Point it at the failing behavior or paste the error. It traces the cause through the codebase and proposes a fix, often with a test that proves it.
Review my changes. Codex is a strong code reviewer that catches issues humans skim past.
Write tests for this module. Tedious for humans, perfect for an agent.
Explain this repository. Drop it into an unfamiliar project and ask how things work.
Build this well described feature. The more precisely you describe it, the better the pull request that comes back.

The delegation mindset

Codex rewards a different working style than chat tools. Instead of a long back and forth, you invest in the task description up front: what to change, what not to touch, how to verify success. Then you let it run. The quality of what comes back tracks the quality of what you wrote, which is excellent training for thinking like an engineer even if you never write a line yourself. Always review the result before merging; an agent that runs unattended deserves the same scrutiny as a new team member.

What does it cost?

Codex is included with paid ChatGPT plans, with usage limits that scale with the tier. The CLI is open source and can also be used with API credits if you prefer paying per use. Limits and bundles shift regularly, so check the official page for the current setup before committing.

Honest about the limitations

Codex assumes a codebase and some comfort with developer concepts like repositories, branches and pull requests. If you have never touched GitHub, start with a browser builder first and come back. The agent is strong but not infallible: it can misunderstand intent, and long autonomous runs amplify a misunderstanding into a lot of wrong code. Keep tasks scoped, be explicit about constraints, and treat its output as a draft from a fast, talented, slightly overconfident colleague.

How to get started today

Check openai.com/codex for the access route that fits you: ChatGPT plan, CLI install, or the IDE extension.
Connect a repository, or open a local project with the CLI.
Start with a read only task: ask it to explain the project or review recent changes.
Then hand it one small, well described fix and review the result.
Scale up to parallel tasks once you trust the workflow.

An example task to hand over:

In this repository, the contact form does not validate email addresses.
Add validation: show an inline error under the field for invalid input,
and block submission until it is fixed.
Do not change the styling system or any other form behavior.
Add a unit test that covers a valid and an invalid address.

The shift Codex represents is subtle but big: from writing code with help, to deciding what should be built and letting an agent do the building. Learning to delegate well is becoming a skill of its own, and Codex is one of the best places to practice it.

Prices, limits and features change fast in the AI world. Always check the official page of the tool for the current state of things.