What is OpenAI's Codex app for Mac, and how do you manage several coding agents at once?

The Codex app, released by OpenAI for macOS on February 2, 2026, is a desktop "command center" where you can run several AI coding agents side by side, each in its own thread and its own copy of your code, then review their changes before keeping them. It adds reusable skills, scheduled Automations and a choice of personalities, and for a limited time it is open even to ChatGPT Free and Go users.

What is OpenAI's Codex app for Mac, and how do you manage several…

Yesterday, February 2, 2026, OpenAI released the Codex app for macOS. Codex is OpenAI's coding agent: an AI that does not just suggest a line of code but takes a task ("add a login page," "fix this failing test"), reads your project, edits files, runs commands and reports back. Until now you could use it in a terminal, inside a code editor, or on the web. The new app gives it a home on the desktop, designed for a different way of working.

OpenAI's reasoning is that developers increasingly run several agents at once, on tasks that can last hours or even days. A terminal window or an editor sidebar is built for one conversation at a time. The app is meant to be the place where you hand out work, watch progress and check results across many agents.

What OpenAI released

The app signs in with your ChatGPT account. Its main parts:

Feature	What it does
Threads by project	Each agent works in its own thread, grouped under the project it belongs to
Review in place	You read an agent's changes, comment on them or open them in your editor
Worktrees	Each agent edits a separate copy of your code, so agents don't collide
Skills	Packaged instructions and scripts that teach Codex a tool or workflow
Automations	Tasks that run on a schedule and drop results in a review queue
Personalities	A brief, practical style or a chattier, more empathetic one

It also picks up your history and settings from the Codex command-line tool and editor extension, so existing projects appear right away.

SiliconANGLE, reporting on launch day, describes a minimal chat box similar to ChatGPT, with the generated code appearing step by step and a sidebar of shortcuts to project files. It adds that the app runs on GPT-5.2-Codex, OpenAI's current coding model, which can take prompts of up to 400,000 tokens. A token is a small chunk of text, often part of a word.

How it works: parallel agents, safely

The key idea is isolation. A Git worktree is a second working folder linked to the same code repository, on its own branch. Give three agents three worktrees and each one changes its own copy. You compare their results, keep the best, and your main folder stays untouched until you decide. Think of three assistants each editing a photocopy of a contract rather than all scribbling on the original.

Skills extend what Codex can do. A skill is a folder holding instructions, reference files and small scripts. OpenAI ships a library that includes skills for turning Figma designs into interface code, managing tasks in Linear, deploying sites to hosts such as Cloudflare, Netlify, Render and Vercel, generating images, and creating PDF, spreadsheet and Word files. You can ask for a skill by name, or let Codex choose one that fits. A skill you create in the app also works in the command-line tool and the editor extension, and a team can store skills in its repository to share them.

To show what this allows, OpenAI had Codex build a 3D kart-racing game from one detailed prompt plus a series of generic follow-ups such as "play the game and fix what's missing." Using a game-development skill and an image skill, Codex worked through more than 7 million tokens, acting as designer, programmer and tester.

Automations are scheduled jobs: instructions, plus optional skills, that run at times you choose. OpenAI says its own teams use them for daily issue triage, summaries of failed test runs and release briefs. Results wait in a review queue rather than changing anything silently. For now they run on your Mac; OpenAI says cloud triggers are coming.

Security defaults are strict. An agent may edit only files in the folder or branch it was given and use cached web search. Anything needing more power, such as network access, requires your permission, though teams can write rules that allow certain commands automatically. This sandbox, a fenced-off area the agent can't leave, uses the same open-source system as the Codex command-line tool.

Who can use it, and why it matters

Codex is included in ChatGPT Plus, Pro, Business, Enterprise and Edu, with extra credits for heavy use. For a limited time OpenAI is also opening it to Free and Go users and doubling rate limits on paid plans, everywhere Codex runs. A Windows version is planned but has no date.

According to OpenAI, use of Codex has doubled since GPT-5.2-Codex arrived in mid-December, with more than a million developers using it in the past month. SiliconANGLE places the launch in a close race with Anthropic's Claude Code, reported to have passed $1 billion in annualized revenue.

The bigger shift is in the role of the person. With one agent, you pair-program. With several, you become closer to a manager: you split work into clear tasks, assign them, and review what comes back. That is a skill in itself, and the app is designed around it.

Try it: Pick a small project and write two different instructions for the same improvement, for example "make the home page load faster by compressing images" and "make it faster by removing unused code." Run each in its own thread and compare the two results before keeping one.

How do you start your first project in the Codex app?

Download the app from OpenAI's Codex page, sign in with your ChatGPT account and open a project folder that is already under Git version control, so every change can be undone. If you have used the Codex command-line tool, your earlier sessions will already be there.

Begin with a contained task and spell out the finish line:

Prompt

In this project, add a "Contact" page with a form for name, email
and message. Match the style of the existing pages. Don't add new
libraries. When done, list the files you changed and how to test it.

Review the changes in the thread, ask for fixes in plain language, and only then merge them into your main branch. Once that feels comfortable, try two agents on separate tasks, then a skill, and finally a simple daily Automation such as a summary of open issues.