How to Set Up OpenAI Codex

Codex is OpenAI’s coding agent. It can read a codebase, edit files, run commands, review changes, and handle longer tasks in the cloud.

Unlike an ordinary chat, Codex can:

• Work straight inside your project folder
• Run your test suite and tools
• Show diffs before you accept changes
• Follow project rules set in AGENTS.md
• Use skills, plugins, and MCP servers
• Review GitHub pull requests
• Run work in cloud environments
• Run tasks on a schedule

ChatGPT desktop app

Codex ships inside the ChatGPT desktop app for macOS, Windows, and Linux.

Codex CLI

A terminal client for local work, scripts, and CI pipelines.

IDE extension

Works in VS Code, Cursor, and Windsurf. Xcode and JetBrains IDEs get their own integrations.

Codex Cloud

Runs tasks remotely so you can pick them up on the web, mobile, or desktop.

GitHub and Slack

Ask for reviews and tasks from the tools your team already uses.

Free and Go

GPT-6 Luna at Standard speed in the desktop app, depending on rollout.

Plus ($20/month)

Codex on the web, CLI, IDE extension, and iOS, plus cloud features such as automatic code review and Slack.

Pro (from $100/month)

Tiers at $100, $200, or $500 per month for heavier workloads.

Business

$20 per user per month billed annually for 2+ users, or $25 billed monthly, with admin controls included.

Enterprise & Edu

Advanced security, compliance, and admin features. Talk to sales.

API key

Runs Codex in the CLI, SDK, or IDE extension at API pricing. Cloud features are not included.

ChatGPT Work and Codex draw from the same usage limits on your plan.

Pick the smallest plan that covers the surfaces you actually use.

Download the ChatGPT desktop app for your operating system and sign in with your ChatGPT account.

To begin coding:

• Open your project folder
• Pick Codex from the ChatGPT dropdown
• Describe the task

Start here if the terminal is not your daily home.

On macOS or Linux, run this:

curl -fsSL https://chatgpt.com/codex/install.sh | sh

On Windows, open a fresh PowerShell window and run:

powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

Or install through a package manager:

npm

npm install -g @openai/codex

Homebrew

brew install --cask codex

In your terminal, open a project folder and run:

codex

On first launch, pick Sign in with ChatGPT to draw on your plan’s included usage.

You can sign in with an OpenAI API key instead, billed at API rates. Some features, such as cloud-based integrations, do not work with an API key.

Never paste an API key into a prompt, a committed file, or a shared document.

Install the Codex extension in VS Code, Cursor, or Windsurf, then open the Codex sidebar to:

• Reference open files and selected code
• Review edits beside your code
• Send longer tasks to the cloud
• Keep follow-ups in one chat

No icon? Open the Command Palette and run Codex: Open Codex Sidebar.

GPT-6.1 Sol

The recommended pick for complex coding and agentic work. Close to Astra performance at a lower cost.

GPT-6 Luna

Best for focused, repeatable tasks.

GPT-6 Astra

OpenAI’s most capable model, built for the hardest problems.

Switch models with /model in the CLI, or launch with:

codex --model gpt-6.1-sol

GPT-5.5 left Codex for ChatGPT sign-ins on October 14, 2026. Swap out gpt-5.5 wherever it appears in your settings, scripts, or scheduled tasks.

Higher reasoning effort can lift results on complex tasks, but it runs slower and burns more of your limits.

• Begin at the default effort
• Raise it for planning, debugging, and architecture work
• Lower it for quick edits and repetitive chores

In the CLI, /model also adjusts reasoning effort. Ultra mode goes further and splits large tasks across subagents.

Two layers keep Codex safe:

Sandbox mode

What Codex can technically do, such as where it may write and whether it can reach the network.

Approval policy

When Codex has to stop and check with you before acting.

On your machine, Codex runs in an OS-enforced sandbox, and network access is off by default.

The Auto preset lets Codex read, edit, and run commands inside your working folder, and it asks first before stepping outside it or touching the network:

codex --sandbox workspace-write --ask-for-approval on-request

Run /permissions to drop to read-only when you just want to plan or ask questions.

Codex reads AGENTS.md files before it starts any work. Run /init to generate a starter file for your project.

Cover:

• What the project does
• How to install dependencies
• How to run tests and linting
• Code style rules
• Folders to leave alone
• Commands that need approval
• What counts as done

Keep personal defaults in ~/.codex/AGENTS.md. Project files add to and override them, and the combined size caps at 32 KiB by default.

Short, specific rules beat long explanations.

Your local settings live in ~/.codex/config.toml. Use it to store:

• Your default model
• Sandbox and approval defaults
• Network access settings
• MCP servers
• Trust levels per project

To allow network access in the default sandbox, turn it on explicitly:

[sandbox_workspace_write] network_access = true

The old untrusted approval policy is retired. Delete it from older configs, or Codex may fail to start.

A skill is a folder holding a SKILL.md file, plus optional scripts and references. SKILL.md must include a name and a description.

Project skills

.agents/skills inside your repository

Personal skills

~/.agents/skills inside your home folder

Type $ or run /skills to mention a skill. To build one with guidance, invoke $skill-creator.

MCP servers let Codex reach external tools and data, such as documentation, tickets, or databases.

codex mcp add

Run /mcp inside a session to see which servers are connected.

Connect only the servers a task requires, and check what each one can do.

Let Codex review your work before it ships:

• Run /review on uncommitted changes or against a base branch
• Mention @codex review on a GitHub pull request
• Enable automatic reviews for connected repositories

Codex returns prioritized findings and leaves your working tree untouched, so you choose what to fix.

Cloud tasks run remotely, so they keep working while you focus elsewhere.

In a new task, pick Work in, then Cloud, open Select environment, and choose Create environment. Codex inspects your repository and helps you prepare and test the setup.

In the terminal, codex cloud lets you browse chats, submit work, and apply results locally.

codex

Open an interactive session

codex exec

Run a task without prompts in scripts or CI

codex resume

Pick up a recent chat

codex --search

Turn on live web search for current information

codex --image

Attach a screenshot or diagram

/init

Scaffold an AGENTS.md file

/review

Review changes

/status

Check context usage and rate limits

Begin with something low-risk:

Walk me through this project. Explain its structure, how to run it, and how the tests work. Do not change any files.

Next, try a small change you can review:

Write tests for [one function]. Run the test suite and show me the diff before you finish.

Commit to Git before and after every task so you can always roll back.

• Commit before each task
• Set one clear goal per task
• Request a plan for large changes
• Review every diff
• Run tests yourself before you merge
• Keep AGENTS.md current
• Open a new chat for each new topic

• Leave the sandbox on for everyday work
• Enable network access only when a task needs it
• Keep secrets out of AGENTS.md and prompts
• Vet MCP servers and plugins before installing them
• Check your plan’s data settings before touching client code
• Treat website and file content as data, never as instructions

GPT-6 Astra also includes safety monitoring that can pause a task when it detects potentially unsafe behavior. Read the findings before you resume.

The sandbox and approvals protect you. Careful review protects your users.

Plan

Pick a plan that covers your surfaces.

Install

Desktop app, CLI, or IDE extension.

Sign in

ChatGPT account or API key.

Model

GPT-6.1 Sol for complex work, Luna for focused tasks.

Permissions

Workspace-write with approvals on request.

Instructions

An AGENTS.md file with project rules.

Extend

Skills and MCP only where they help.

Review

Review code before anything ships.

Paste this into a new repository once Codex is installed:

Act as my Codex setup assistant for this repository. Inspect the project without changing anything. Then propose an AGENTS.md that covers what the project does, how to install dependencies, how to run tests and linting, code style rules, folders to avoid, and which actions must ask for approval. Recommend sandbox and approval settings for daily work, list any MCP servers or skills that would truly help, and explain why. Wait for my approval before you create or change any files.