Codex is OpenAI's agent for the terminal. It reads a repository, proposes changes, runs commands in a sandbox and reports back. It ships with a ChatGPT subscription, so if you pay for Plus or Pro you already have it.

How I use it

I use it as the second reviewer. When a build is stuck and the first agent keeps circling the same wrong idea, I hand the same problem to Codex with a short brief and let it diagnose cold. It has a different model behind it, and a different model has different blind spots. Maybe one in three times it finds the thing. That ratio is worth a subscription I had anyway.

The second use is images. The same CLI can call OpenAI's image model, which is useful when I need eight variations of a product scene in one go and do not want to open a web app. I cap the credits per task before I start, because a loose image batch can eat a day's allowance in ten minutes.

What surprised me is how good it is at reading code it did not write. Ask it to explain a module and it gives a calm, accurate summary. What annoys me is the metering. Since the switch to credits, a long agentic session is a line item, and the limits are harder to predict than a flat plan. The sandbox rules also shift between releases, and twice a script that worked on Monday needed a flag on Friday.

Who it is for

If you already pay for ChatGPT, turn it on; it costs nothing extra to try. If you are choosing one coding agent and paying for one subscription, I would not make it this one. It shines as the checker, not the builder. Skip it if you do not have a primary agent yet; get that first, then add this when you want a second set of eyes.

Related tools