Codex: Beginner

Beginner

Codex is OpenAI’s agentic coding agent. It runs in your terminal and your IDE, reads and edits files in your project, runs commands in a sandbox, and works through multi-step tasks, pausing for your approval along the way. It ships as a CLI, an IDE extension, and a cloud service that all share one backend. This guide gets you from install to a productive first task. For the bigger picture, see How AI coding tools work.

What you need

• A terminal, or VS Code, Cursor, Windsurf, or a JetBrains IDE for the extension
• Node.js (for the CLI install via npm), or Homebrew on macOS
• A ChatGPT plan (Plus, Pro, Business, Edu, or Enterprise; Free and Go work at low limits) or an OpenAI API key

Install

1. Install the Codex CLI (or the IDE extension):

   npm

   npm install -g @openai/codex

   Homebrew

   brew install codex

   IDE extension

   Install the Codex extension from the VS Code, Cursor, Windsurf, or JetBrains marketplace.

2. Go to a project and start it:

   cd my-project
   codex

3. On first run, sign in with your ChatGPT account or paste an API key.

Tip

Start Codex from the root of your project. The current directory is its workspace: it decides which files Codex can see and edit. This is the context window in action, see The context window.

Your first task

Talk to it in plain English. A few good first prompts:

What does this project do? Give me a short overview.

Find where the page title is set and change it to "Hello".

Codex explores the files, shows the edits it plans to make, runs any commands in a sandbox, and asks for approval before it changes anything.

Approvals and the sandbox

Codex doesn’t edit files or run commands behind your back. Two controls decide how much freedom it has:

• Approval mode: how often it stops to ask.
• The sandbox: an OS-level boundary (macOS Seatbelt, a Linux sandbox, or the native Windows sandbox / WSL2) that limits what a command can touch.

Approval mode	What it does
Auto (default)	Edits inside the workspace and asks before anything riskier
Read-only	Looks but doesn’t change files or run commands
Full Access	Acts without stopping to ask

Switch mode at any time with /permissions.

Caution

Full Access turns off the approval gate. Keep it for throwaway containers, never on a machine that holds secrets or production credentials. Reading what Codex proposes before you approve is your real safety net, see Verification.

The commands you’ll use first

Command	What it does
/init	Draft a starter AGENTS.md so Codex knows your project
/permissions	Switch approval mode mid-session
codex resume	Continue your previous session

You’re done when…

• ✅ You can start Codex in a project (CLI or IDE)
• ✅ You’ve approved at least one file change
• ✅ You know how to switch approval modes and resume a session

Next step

Ready for a real workflow: a steering file, the right context, and the window that limits it all?

Codex: Intermediate Steer Codex with AGENTS.md, feed it the right context, and manage the window.

Foundations New to the ideas behind these tools? Start with how AI coding tools work.