Codex

Codex is OpenAI’s answer to the same problem Claude Code solves. A chat window cannot see your project, so you spend your time copying code out of it and pasting errors back in.

What it actually does

You install it, open a terminal in your project folder, and type codex. From there it reads your files, edits them, runs commands, and sees what breaks. The loop is the point. You describe an outcome rather than a snippet, and it keeps going until the outcome exists or it gets stuck.

The part people miss

Codex is two things wearing one name. There is the agent you type at, and there are the models underneath it. Other tools will happily call those models on your behalf, which means one subscription can power more than the one terminal you bought it for.

That is worth knowing before you pay for a second AI plan you may not need.

How I use it

Two jobs. It is my second coding agent, and its models are what my scheduled agent calls when a job fires, so the agent work and the coding work come off the same plan. My current setup shows how that split is wired, and why the thing that is awake all day is not the thing doing the thinking.

It also generates still images when I need them, which is not what you would expect from something with “code” in the name.

Where it fits

If you already pay for ChatGPT, this is the cheapest way to get an agent onto your real files, because you are not adding a bill.

Running two coding agents is a luxury rather than a requirement. They fail differently, though, and when one of them is confidently wrong the other tends to notice.

Honest limitations

It has the failure mode every coding agent has. It is most confident in exactly the situations where it is guessing, so read the diff (the list of changes) before you accept it.

It also wants a terminal and a real project folder. If you do not have either yet, a normal chat window is genuinely fine and you have lost nothing.

What a plan includes here changes regularly, so confirm the current terms before assuming one subscription covers both the terminal agent and whatever else you point at its models.

Next step

If you would rather see edits land in a file than scroll past in a terminal, Cursor is the same job with more visibility. Claude Code is the closest direct alternative, and running one of each is the arrangement my setup ended up with.