TOOLS & SKILLS
Get Started With A Codex Agent 🛠️
The 15-minute setup to get OpenAI's Codex agent running on your Mac, paired with your phone, and shipping real work for you.
OpenAI just put Codex on your phone. That means you can now have an AI agent running on your Mac for hours while you're walking around with your iPhone, checking in on its progress and approving its work. It's the closest thing yet to "real assistant that does actual work."
But most people get stuck before they even start. The install is confusing. The plan tier requirements are unclear. The first prompts feel like dev-tool territory if you don't write code. This guide fixes that.
By the end you'll have Codex installed, your first project running, your phone paired with your Mac, and three real prompts you can use today. Even if you've never opened a terminal in your life.
What You'll Need
A paid ChatGPT plan. Codex is included with every paid ChatGPT subscription, no separate sign-up needed. The recommended starting tier is Plus at $20/month. It gives you full Codex access on Web, CLI, IDE, and iOS, with enough capacity (15 to 80 GPT-5.5 messages per 5-hour window) to run a few multi-step tasks per session. Free and Go ($8) work for trying it out, but you'll hit limits fast on real work.
A Mac or Windows computer. The Codex desktop app runs on macOS (Apple Silicon or Intel) and Windows (via the Microsoft Store). Linux support is coming. If you're on Windows, the Codex CLI works fine too.
A folder for your project. Codex operates inside a folder you pick. Make a new one on your Desktop called "Codex" before you start. That's where your work will live.
About 15 minutes for the full setup including pairing your phone.
5 Steps To Your First Agent
Go to chatgpt.com and make sure you're on a paid plan (Plus is the sweet spot). Then go to developers.openai.com/codex/quickstart and download the Codex desktop app for your operating system. Install it and launch it.
When Codex opens, it asks you to sign in with your ChatGPT account. Click sign in, complete the OAuth flow in your browser, and come back to the app. Then it asks you to pick a project folder. Select the Codex folder you made on your Desktop earlier. That's now Codex's workspace. It will only edit files inside that folder by default.
Type into the chat box. OpenAI's literal suggested first prompt is "Tell me about this project." If your folder is empty, ask it something practical instead. Drop a CSV, a resume, or any document into your project folder, then ask Codex to do something useful with it.
This is the part that changes how you use Codex. On the desktop app, open the mobile pairing screen (it shows a QR code). On your iPhone or Android, open the latest ChatGPT app, tap into Codex, and scan the QR code. Done. Under a minute.
Now you can start tasks from your phone, get push notifications when Codex finishes or needs your approval, review the diffs and terminal output, and approve commands. All while your Mac quietly does the work at home.
Codex has 3 approval modes. Pick the one that matches how cautious you want to be:
Read-only: Codex can read your files and answer questions. It asks before editing anything, running commands, or accessing the network. Safest for beginners trying it out.
Auto (default): Codex can read and edit files inside your workspace and run local commands. It asks before editing outside the workspace or accessing the network. This is the right setting for daily use.
Full Access (YOLO mode): No sandboxing, no approval prompts. OpenAI flags this as "elevated risk and not recommended." Stay out of this unless you really know what you're doing.
5 Mistakes Beginners Make
Where to learn more
OpenAI's official docs are at developers.openai.com/codex. The quickstart, pricing, sandboxing, and use-cases pages are the highest-signal reading. Skip the third-party tutorials until you've spent a week with Codex yourself.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.