According to OpenAI, you install Codex with a single command: on Mac and Linux with curl -fsSL https://chatgpt.com/codex/install.sh | sh, on Windows in a new PowerShell window with powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex". You can also use npm install -g @openai/codex, or brew install --cask codex on a Mac. Then run codex in your project folder, choose "Sign in with ChatGPT" and sign in through the browser. With ChatGPT sign-in, OpenAI lists the command line from Plus; the pricing page names Free and Go only for the desktop app. Alternatively you sign in with an API key and pay per use. To use Codex as a second reviewer next to Claude Code, you have two options: /review in a second terminal, or OpenAI's official plugin inside Claude Code with /codex:review. According to the docs, neither review changes your files.
I work with Claude Code every day and run 14 AI employees on it. I use Codex regularly for correction loops: Claude Code builds something, Codex checks the result with fresh eyes, or the other way round. The installation steps in this article come from OpenAI's official docs and README, sources as of November 1, 2026, not from a test of my own. Everything about Claude is collected on my Claude page.
What you need before you install
Three things decide whether the installation goes without detours.
A suitable plan or an API key. OpenAI's pricing page lists Codex for Free (0 dollars) and Go (8 dollars) only as "GPT-6 Luna at Standard speed in the desktop app, subject to rollout". Web, command line and editor extension start at Plus for 20 dollars a month. The README of the Codex repository says the same for the command line: with ChatGPT sign-in you use Codex "as part of your Plus, Pro, Business, Edu, or Enterprise plan". Alternatively you sign in with an API key; OpenAI then bills at API rates, and cloud features are not available.
A Git repository for the review. Codex itself runs in any folder. According to the docs, however, the /review command only appears when the folder is inside a Git repository, and codex exec requires one by default as well. If your texts or workflows are not in Git yet, a git init in the folder is enough.
An up-to-date operating system. For Windows, OpenAI names Windows 11 as the recommended baseline. A fully updated Windows 10 from version 1809 is "best effort" according to OpenAI; older builds are not recommended. On Windows, winget should also be available.
Installing Codex on a Mac
On a Mac you have three options, all from the official docs:
# OpenAI's install script
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# or via Homebrew
brew install --cask codex
# or via npm
npm install -g @openai/codex
Updating works the same way: run the script again, use brew upgrade --cask codex with Homebrew, or repeat the npm install. According to the README, the script downloads from releases.openai.com by default and falls back to GitHub Releases if something is unavailable there. If you prefer a ready-made file, the latest GitHub release has codex-aarch64-apple-darwin.tar.gz for Apple Silicon and codex-x86_64-apple-darwin.tar.gz for older Intel Macs. On November 1, 2026, the latest release was version 0.166.0 from October 29, 2026.
If you don't want a terminal: Codex is also part of the ChatGPT desktop app for macOS, Windows and Linux. According to the quickstart, you select "Codex" in the dropdown at the top. For the reviewer role next to Claude Code, though, the command line is the more direct route, because both tools then work in the same folder.
Installing Codex on Windows
On Windows, open a new PowerShell window and run:
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"
The same command updates Codex later. The npm install works on Windows just as on a Mac.
You no longer need WSL for this. According to OpenAI, Codex runs natively in PowerShell with its own Windows sandbox that limits file and network access. OpenAI describes three variants. The recommended one is mxc, Windows' own process isolation, without administrator setup and without extra Windows accounts; Microsoft introduced it with Windows 11 24H2 (build 26100.9278) and 25H2 (build 26200.9278) and, according to the docs, is rolling it out gradually. According to the docs, the desktop app picks it by itself; on the command line you turn it on with prefer_mxc = true in the [features] section of config.toml, otherwise the older selection applies. If it is missing, elevated is the preferred fallback, which requires a one-time administrator-approved setup; commands in the sandbox still run without administrator rights. The third variant, unelevated, has weaker network isolation and is meant only as a last resort.
WSL remains an option if your tools already live on Linux. OpenAI's wording: "Use the native Windows sandbox by default." OpenAI recommends WSL2 only if you need Linux tooling or your repositories already live there. Codex has not supported WSL1 since version 0.115.
Installing Codex on Linux
On Linux you use the same script as on a Mac, or npm:
curl -fsSL https://chatgpt.com/codex/install.sh | sh
If you prefer a ready-made file, the latest GitHub release has codex-x86_64-unknown-linux-musl.tar.gz and codex-aarch64-unknown-linux-musl.tar.gz, according to the README. After extracting, rename the file to codex. The ChatGPT desktop app with Codex is available for Linux as a preview, according to the docs for Ubuntu 24.04 and 26.04, Debian 13, Fedora 43 and 44, and Arch Linux, each for x64 and ARM64.
Signing in: ChatGPT account or API key
The first time you run codex in a project folder, Codex asks you to sign in. Choose "Sign in with ChatGPT", a browser window opens, and after you sign in the browser hands the credentials back to Codex. Later you reach the same flow with codex login. codex login status shows which method is active.
Two notes from the docs that I consider important. First, Codex stores your login as a plaintext file at ~/.codex/auth.json or in your operating system's credential store. OpenAI's advice: "treat ~/.codex/auth.json like a password". With cli_auth_credentials_store = "keyring" in config.toml you force the credential store. Second, your sign-in method decides which data rules apply: with a ChatGPT account, your ChatGPT workspace settings apply; with an API key, those of your API organization. If you use Codex with client data, clarify this first.
Setting up Codex as a second reviewer next to Claude Code
For me, that is a fixed rule: whoever creates something does not correct it. The checker does not know the path to the result and therefore only reads what is there. Codex suits this because it works in the same folder as Claude Code. There are two ways, and both review read-only.
Option 1: Codex in a second terminal. Claude Code works in one window; you start Codex in the same folder in a second one. There you type /review and choose one of the four scopes from the docs: against a base branch, uncommitted changes, a specific commit, or custom review instructions. OpenAI describes the result like this: Codex "reports prioritized, actionable findings without changing your working tree". Without a dialog, use codex review --uncommitted. If the review should use a different model from your session, set review_model in config.toml.
Option 2: Codex inside Claude Code. OpenAI publishes a plugin for Claude Code at openai/codex-plugin-cc. According to the README, the plugin needs Node.js 18.18 or later. If Codex is missing and npm is available, it offers to install it. You install it in a Claude Code session with these commands from the README:
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup
/codex:setup checks whether Codex is installed and signed in. After that, /codex:review reviews your current changes, "read-only" according to the README and with the same quality as /review in Codex. /codex:adversarial-review also questions the decisions behind a change and takes a focus of your own as text. The plugin uses your locally installed Codex with the same sign-in and configuration. As a requirement, the README still names a ChatGPT subscription "incl. Free" or an API key; for the command line, the current pricing page is what counts. Two caveats belong here: according to GitHub, the last change in the repository dates from July 8, 2026, and the model examples in the README still name older models such as Codex-Spark, which according to the changelog has not been available since September 14, 2026. And the optional "review gate", which has Codex check every Claude response automatically, can create a long loop between the two tools and drain usage limits quickly, according to the README. How to check plugins before installing them is covered in Claude Code Plugins: Install, Build, Trust.
So Codex knows your rules. According to the docs, Codex reads AGENTS.md, not CLAUDE.md. The simplest route is to create an AGENTS.md and have your CLAUDE.md import it; I explained this in CLAUDE.md vs AGENTS.md: What Survives. New is the /import command in the Codex command line: according to the docs, it brings instruction files, settings, skills, MCP servers, hooks and subagents from Claude Code into Codex's formats and "doesn't change or delete your existing agent setup". A pure reviewer doesn't need this. A short AGENTS.md that says what to check against is enough there.
Your first review task in five steps
This is what a first review task looks like when Claude Code has just changed something:
- Save the state. Don't let Claude Code commit anything you still want to review, or note the commit. OpenAI recommends Git checkpoints before and after every task anyway, so you can go back.
- Start the reviewer.
codex in a second terminal in the same folder, or /codex:review in Claude Code.
- Choose the scope. "Review uncommitted changes" for a fresh change, the comparison against the base branch for a whole piece of work.
- Set the criteria. With "Custom review instructions" you say what matters. An example of how I would word such an instruction: "Check the changed files against the sources in the
daten/ folder. Report every number without evidence, every link without a target and every claim that goes beyond the source. Don't change anything."
- Hand back the findings. The list goes to the session that built the change. The reviewer does not fix things itself, otherwise nobody reviews the fix.
For reviews without a dialog, for example in a script, there is codex exec. According to the docs it runs "in a read-only sandbox" by default and prints only the final answer to standard output, so you can write it straight to a file with -o findings.md.
What a review costs and where the limits are
Every review counts against your Codex usage, and according to the README that includes the plugin. For Plus and Standard Business, OpenAI gives estimates, not fixed limits: 15 to 160 local messages per five hours with GPT-6.1 Sol, 350 to 3,000 with GPT-6 Luna. Pro comes at 100, 200 or 500 dollars a month; according to the pricing page, Pro plans "currently" have no five-hour limit, but weekly limits may apply. Business costs 20 dollars per user per month on an annual contract with at least two users, or 25 dollars monthly. /status shows your current state in a running session. What each plan gives you in usage and how it compares to Claude Code is worked through in Codex Pricing and Usage Limits 2026.
The limit I consider most important is not a technical one: a second tool only reviews against what it knows. Without sources and clear criteria, Codex cannot check whether a number is correct. How I build reviewer roles into a setup of several agents is shown in Orchestrating Multiple AI Agents. What Anthropic itself offers for reviews is in Claude Code Review: Cost, Setup, Limits.
Frequently asked questions
Is Codex free?
Partly. According to the pricing page, Free and Go include Codex only in the desktop app with GPT-6 Luna, subject to rollout. OpenAI lists command line, editor extension and web from Plus at 20 dollars a month. With an API key you pay per use instead.
Does Codex run on Windows without WSL?
Yes. According to OpenAI, Codex runs natively in PowerShell with its own Windows sandbox, recommended on Windows 11. WSL2 remains an option if your tools run on Linux; WSL1 has not been supported since version 0.115.
How do I update Codex?
The same way you installed it: run the install script again, brew upgrade --cask codex with Homebrew, or npm install -g @openai/codex with npm.
Does Codex read my CLAUDE.md?
Not by itself. According to the docs, Codex reads AGENTS.md. You can list other file names under project_doc_fallback_filenames in config.toml, and Codex then treats them as instruction files. You can create an AGENTS.md and import it in your CLAUDE.md, or use /import to bring your Claude Code setup into Codex without changing the original.
Does Codex change my files when it reviews?
No, as long as you review and don't ask for fixes. /review reports findings "without changing your working tree" according to the docs, /codex:review in the plugin is "read-only" according to the README, and codex exec starts in a read-only sandbox by default. Only when you ask Codex to fix findings do your normal write permissions apply.
Where to go from here
Install Codex with the command for your system, sign in, and have it review a small, finished change that Claude Code built. If you haven't set up Claude Code itself yet, start with Install Claude Code: Windows, Mac, Linux; which learning paths come after that is covered in Claude Code Course: Options and Learning Path. Which of the two tools should be the home for your work is what I compare in Claude Code, Codex or Copilot: Which One?. How others build their reviewer roles is what I show in my community AI Employee Framework.
