The short answer
Codex is the coding agent from OpenAI, the company behind ChatGPT. You give it an instruction in plain language, and it reads your project’s code, edits files, runs commands and tests, and shows you what changed. Codex works in the terminal, as an extension in your editor, in ChatGPT in the browser and in the ChatGPT desktop app.
OpenAI released Codex in April 2025 as Codex CLI, an open-source program for the terminal. The cloud version, the editor extension and the desktop app followed. Millions of developers now use it.
Mind the name. In 2021 OpenAI also had a language model called Codex, the model behind the first version of GitHub Copilot. That model has been retired. Today’s Codex is a different product that happens to share the name.
What Codex does
Give Codex a task like “add a date filter to the orders overview” and it breaks it into small steps. It finds the files involved, writes the change, runs the tests and fixes whatever fails. You choose where that happens.
Locally in your terminal
Codex CLI runs on your own machine, in your project folder. You see every step as it happens and can step in at any point. The CLI's code is open source.
In your editor
With the extension, Codex works in VS Code and editors built on it, such as Cursor and Windsurf. There are also integrations for JetBrains and Xcode.
Tasks in the cloud
From ChatGPT you give Codex access to your GitHub repository. Each task then runs on an OpenAI server, in its own isolated environment with a copy of your code. That lets you run several tasks at once and pick up each result as a pull request.
Reviewing pull requests
Codex can read through a pull request on GitHub and leave comments on what it spots, such as a bug or a forgotten edge case.
Conventions in AGENTS.md
An AGENTS.md file in your project holds the conventions: which commands exist, which style you follow, what it must never do. Codex reads that file for every task, and many other tools read the same file.
On your own machine, Codex runs in a sandbox by default. You decide how much it may do without asking: read only, edit files in your project, or more. That combination of model, tools and limits is called an AI harness. What that means is in what is an AI harness.
What does Codex cost?
Codex comes with ChatGPT, and you sign in with your ChatGPT account. The free plan and Go let you try it in a limited form, in the desktop app and for small tasks. From Plus up, Codex works everywhere: the terminal, your editor, the browser and the app, including tasks in the cloud. The more expensive the plan, the more you can do before you hit a usage limit. Teams and businesses have separate plans with more admin control.
You can also use Codex with an OpenAI API key. Then you pay per use, which adds up fast with many large tasks. With an API key you work in the terminal and the editor, without the cloud features such as pull request review. Current prices and limits are on OpenAI’s website.
Codex, Claude Code and Cursor compared
Codex and Claude Code are the closest of the three: both are agents you hand a task to. Cursor is an editor you type in yourself, with an agent built in.
| Codex | Claude Code | Cursor | |
|---|---|---|---|
| Made by | OpenAI, the company behind ChatGPT. | Anthropic, the company behind Claude. | Anysphere, part of SpaceX since 2026. |
| What it is | A coding agent that carries out tasks on its own, locally or in the cloud. | A coding agent that carries out tasks on its own in your project. | A code editor with AI built in, based on VS Code. |
| Where you work | Terminal, editor, ChatGPT in the browser or the desktop app. | Terminal, VS Code, JetBrains, desktop app or browser. | In the editor itself. |
| Models | OpenAI's models, plus an open model when run locally. | Anthropic's Claude models. | A choice of models from several providers. |
| Conventions in | AGENTS.md | CLAUDE.md | Rules in your project |
| Plan | Limited on the free plan, in full from ChatGPT Plus, or pay per use through the API. | In the paid Claude plans, or pay per use through the API. | Its own subscription, with a free version. |
Between Codex and Claude Code, the biggest differences are the models and which subscription you may already have. If your team uses ChatGPT, Codex is the shortest route. If you already work with Claude, Claude Code is the obvious pick. A well set-up project makes more difference than the choice between the two. More on the other two is in what is Claude Code and what is Cursor. Where else you can run OpenAI’s models is in the best AI harness for each model.
How a senior team works with Codex
Whichever agent you use, the way of working that gets good results is the same. What matters is how you brief it and what you do with what comes back.
A good AGENTS.md
List the test commands, the folder structure, the conventions and what is off limits. Every mistake the agent makes twice becomes a line in that file.
Small tasks, in parallel
In the cloud, Codex can run several tasks at once. That works best with small, separate tasks that don't touch each other: a bug, a test, a screen.
Tests the agent runs itself
An agent that can test its own work finds its own mistakes. Make sure the tests also run in the cloud environment.
Limits set up front
On your own machine, give Codex no more permissions than the task needs. Full access is convenient, and it also makes every slip immediately real.
Read every change
What Codex produces goes through the same review as a colleague's work. On security, payments and customer data a senior developer looks extra hard.
Handing off ten tasks at once is easy. Reviewing ten changes at once, and getting it right, is the actual work.
That is where the gain from an agent tops out: at the speed at which someone experienced can judge what comes back. What that looks like across a whole project, from first setup to production, is in how to develop software with AI.
Where Codex is a weaker fit
Codex is built for people who can read code. Without that knowledge you can’t judge whether a change is safe, and small mistakes pile up unnoticed. If you want to build something without programming knowledge, a tool like Lovable is an easier start.
For work where you want to steer every line yourself, an agent is a detour too. An editor like Cursor, where you type and the AI completes, fits better. And on a large, old codebase without tests, Codex needs a lot of guidance: get the tests and an AGENTS.md in place first. Which tool fits which situation is in the best AI harness for coding.
Conclusion: OpenAI’s agent, under the same rules
Codex is OpenAI’s answer to the question of how AI can do programming work on its own. It runs in the terminal, your editor or the cloud, and suits teams that already work with ChatGPT. What produces good work is the same as with any agent: a clearly set-up project, small tasks and review by someone who knows where things can go wrong.
Codex = doing
Tasks you hand to an agent, locally or in the cloud, several at once if you like.
Senior developer = judging
The brief, the limits and the review. That is where quality comes from, and it stays with a person.
What is Codex?
Is Codex free?
What is the difference between Codex and ChatGPT?
What is the difference between Codex and Claude Code?
Is Codex open source?
Is Codex the same as the old Codex model?
Build faster with AI without lowering the bar?
Our senior developers work with coding agents every day. Bring them onto your project, or have your team learn to use them well in a workshop.