Codex is an agentic coding tool from OpenAI. It reads your codebase, edits files across it, runs commands and opens pull requests — from a CLI, an IDE extension, the ChatGPT app, or an isolated cloud sandbox.
It is built for delegation. Specify a task properly, hand it off, stop thinking about it, and review a finished branch later. For a migration or a test suite that is close to the ideal shape of the work.
What comes back is a branch. That is exactly right when the definition of done is in the tests, and it is the whole problem when the thing you need to judge is an interface — because a branch is not something you can look at.
What it is
Codex is an agent rather than an assistant: you describe an outcome, and it decides which files to open, what to change and when it has finished. The unit of work is a task.
Its centre of gravity is the cloud sandbox. You hand over a well-specified job, it runs in an isolated environment without your machine on, and you collect a branch or a pull request when it is done. Everything about the tool's shape follows from that: it is fast, it expects review at the end rather than supervision throughout, and it is at its best on work you would genuinely rather not watch.
Where it runs
| Codex | |
|---|---|
| Made by | OpenAI |
| Models | The GPT-5.6 family, including the Terra and Luna variants |
| Surfaces | CLI, IDE extension, the ChatGPT desktop app, the web and a cloud sandbox |
| Access | A ChatGPT subscription, or an OpenAI API key |
| Project instructions | AGENTS.md, an open convention other tools also read |
| Extending it | MCP servers, IDE and browser integration |
| Unattended work | Cloud tasks in an isolated sandbox, returned as a branch or PR |
Telling it your rules
Codex reads AGENTS.md, walking from the repository root down to your working directory and layering each file it finds, with a size cap on the combined result. That layering is genuinely useful on a monorepo: a root file carries the house style, and a package can narrow it without repeating it.
AGENTS.md is an open convention that several other tools now read, which makes it the more portable of the two formats. If you also run Claude Code you maintain CLAUDE.md alongside it — most people keep the substance in one and have the other defer to it.
What it is good at
- tasks specified well enough that you would rather not supervise them
- work you want back as a finished branch to review, not as a conversation
- the same mechanical change landing across many services
- teams already working inside ChatGPT every day
- anything where speed matters more than a narrated plan
- interface work, where the diff answers a different question than the one you have
- problems with no obvious cause yet, where you want to interrupt a wrong assumption early
- loosely specified tasks — delegation assumes the spec can carry the intent
- anything whose success is judged by looking at it
Reviewing what it ships
Delegation assumes you can evaluate the result when it arrives. On interface work you cannot — the task completes, the tests pass, the diff is clean, and the page is subtly wrong in a way nothing in that report could have told you.
Reading the diff harder does not help, because a diff tells you what changed in the source while what you need to know is what changed on the screen. That is the specific gap Clodex closes: it runs Codex against your real project, renders the result as a live page beside the conversation, and lets you fix what is wrong by selecting it. The visual editor guide covers it in full, including what it does not do.
Guides
Review the page, not the diff.
Run Codex on a live preview of your project, and fix what is wrong by pointing at it.
- AGENTS.md guide — OpenAI's specification for Codex project instructions and how they layer.
- Clodex security — where credentials live, what the agent can reach, and how to report a vulnerability.