OpenAI Codex CLI
by OpenAI
Getting-started guide
OpenAI's coding assistant for the command line: your machine, one project folder, paid for by your ChatGPT sign-in.
Current prices: Free · See all plans
Last reviewed 2026-09-08 · for OpenAI Codex CLI
You open a project folder on your computer and start Codex in a terminal (a text window where you type instructions instead of clicking). Describe the change you want in plain English. It reads the files, shows you the edits and commands it proposes, and waits for your approval. You read the result and say what to fix.
Codex CLI is OpenAI's coding agent — AI that takes actions on its own rather than only answering — running on your own machine: it opens your real files and runs commands rather than handing you code to paste.
It is one front door onto one product: Codex also lives in the ChatGPT app, a VS Code extension and a desktop app. The CLI is the one you drive where the code sits.
The name is recycled, so older search results mislead: the 2021 "Codex" was a model OpenAI sold through its API and retired in March 2023. Today's is a different thing, a coding agent from 2025.
The entry requirement is code already on your disk and comfort opening a command line. Never used one? A browser-based coding assistant is the better first step; this will still be here.
You do not buy Codex; you sign in with a ChatGPT account and your plan pays for it. OpenAI's pricing page lists it on every plan including the free one, while its own CLI documentation still says Plus or above — so expect to need a paid plan (Go at $8/month, or Plus at $20/month). Billing your own OpenAI developer account per use is the other route.
Usage runs as rolling five-hour and weekly windows rather than a balance you can watch, and Plus and Pro can buy extra credits instead of upgrading.
Official source: OpenAI's official Codex CLI documentation
It installs in one command: use OpenAI's installer, the official Homebrew cask or npm package, or the GitHub releases page — lookalikes exist. On Windows the fence that keeps Codex inside your folder behaves most predictably under WSL2.
Open it in a project that is saved and tracked, with nothing uncommitted — that tracking is how you read back and undo every line it changed.
/permissions sets what it may do: read-only, a workspace default that edits inside the folder you opened and asks first about anything else, or full access, which removes both the fence and the questions — one keystroke, and it stays granted.
Your code goes to OpenAI's servers to be worked on — the mechanism, not a leak, but the line to check before a client's private project.
codex exec works inside a script, but nobody is approving anythingBefore changing anything: list the commands this project uses to build, test and run, and name the files you would touch to add one new setting. Do not edit anything.
Why this one: Codex picks its opening permissions by checking whether the folder is under version control, so your first session may already be allowed to write. Asking a question with /permissions on read-only shows which preset you landed in.
/init
Why this one: Codex reads AGENTS.md before doing any work, merging a global one in ~/.codex with the project's. It is a shared convention with a published spec, not a Codex-only file, so rules you write once travel to other agents.
Tweak: It writes a draft — add your exact test command and anything it must never touch.
Add [the feature] using only what is already in this project — do not add a dependency. If you need something that is not here, stop and tell me what and why.
Why this one: Codex's fence has network access turned off by default, so this matches what it can actually do: a package it cannot download becomes a question, not an install you find later.
Want the full picture first? See the OpenAI Codex CLI overview for pricing, the free tier and related tools.