Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

Codex computer use: letting the Codex app operate your desktop

Unofficial community guide for Codex computer use. Not affiliated with OpenAI. All trademarks belong to their owners.

This guide covers Codex computer use, the feature of the Codex desktop app that lets the agent see and operate graphical applications on macOS and Windows. It draws on OpenAI's developer documentation, a long-form write-up by Jason Liu that separates the three ways Codex can use a computer, and a Lenny's Newsletter episode by Claire Vo showing day-to-day use cases. If a detail is not in those sources, I say "check the docs" instead of inventing it.

Reached for computer use because you wanted the agent to click through a site builder and produce a page? Try Begin.sh - turn a prompt or a URL into a downloadable static site or Expo app. No clicking, no hosting, just a zip.

What it is

Computer Use is the broadest of three surfaces the Codex app has for interacting with software. According to Liu's write-up, it lets Codex work with windows, menus, keyboard input and the clipboard in apps you approve, on macOS and Windows. The other two surfaces are the Codex Chrome extension (for multiple tabs and authenticated sessions) and the in-app browser. They overlap enough to be confusing, which is why most of this guide is about choosing between them.

The trade-off is speed versus reach. A structured plugin or MCP tool calls an API directly; Computer Use has to look at the interface, decide where to click, wait for the app to respond and inspect the next state. That visual loop is slow, but it means Codex can operate apps that expose no useful API at all: Spotify, Xcode, System Settings, an iOS simulator, or the iPhone Mirroring app, to take the examples from the source. On macOS the source notes it can work in approved apps in the background while you keep using the rest of the machine.

How to get started

  1. Install the Codex desktop app. Both sources agree the desktop app is a prerequisite; Vo also installs the official Codex Chrome extension for browser tasks.
  2. In Codex, open Settings, then Computer Use, and click Install. This is the install path given in Liu's write-up.
  3. Approve only the apps you want the agent to touch. Computer Use operates in "apps you approve", so the approval list is your main control.
  4. Trigger it by mentioning @Computer in a thread, or by explicitly asking Codex to use Computer Use. The source expects models to pick the surface on their own more often over time.
  5. Read the official Computer Use documentation and the ChatGPT learn page on computer use for the current platform requirements.

Pricing and limits

None of the sources state a separate price for Computer Use; it is a feature of the Codex app, which is tied to your ChatGPT plan. Liu notes that Computer Use became available in the EU and UK in an update to the article. Check the Codex documentation for current availability and plan requirements.

Practical notes and gotchas

  • Prefer a plugin or MCP when one exists. Liu's rule: a Slack plugin retrieves a thread more precisely than clicking around Slack, and a GitHub plugin produces actions that are easier to inspect than driving the website. Visual control is most useful at the boundary where a structured tool stops.
  • Use the Chrome extension for anything that needs several tabs or an existing login. Vo's setup is desktop app plus Chrome extension; Liu's section on @Chrome is titled "for multiple tabs and auth".
  • Give it one clear app or flow at a time. Computer Use is the widest trust boundary of the three surfaces. Keep sensitive apps closed when they are not part of the task and review permission prompts.
  • Stay present for anything financial, account-related, payment, credential, privacy or system-security related. That list is quoted from the source and is a good hard rule.
  • Use it as the last mile in a structured workflow. Liu's example: a Slack integration could read feedback but not upload a file, so Computer Use clicked "Add file" for that single step.
  • Expect slowness, and design around it. Polling tasks (checking a support chat every five minutes, then every minute once an agent appears) are a good fit precisely because they are slow and boring for a human.

Comparison

Codex Computer Use Codex Chrome extension Structured plugin or MCP Begin.sh
What it drives Native apps on macOS and Windows via windows, menus, keyboard, clipboard Browser tabs, including authenticated sessions APIs directly Nothing; it generates a site or app from a prompt or URL
Speed Slowest: visual look-decide-click loop Faster than native control per the sources Fastest: direct calls One generation, then download
Works without an API Yes Yes No Not applicable
Trust boundary Broadest of the three Narrower Narrowest, actions are inspectable No access to your machine

FAQ

What are the three ways Codex can use a computer? Computer Use, the Chrome extension, and the in-app browser. Liu's article is organized around exactly this split.

How do I turn it on? Settings, then Computer Use, then Install, inside the Codex app. Then mention @Computer in a thread.

Does it work while I am using the machine? On macOS the source says it can operate approved apps in the background while you use the rest of your computer.

Is it available in Europe? Liu's update says Computer Use is now available in the EU and UK.

Should I use it for everything? No. Both sources prefer structured tools when they exist and reserve Computer Use for GUI-only flows, native apps, system settings, or a missing action in an otherwise useful integration.

When you want the artifact, not the automation

A common reason people reach for Codex computer use is to drive a website builder or a design tool until a page exists. That is a slow visual loop over a UI that was never designed for an agent. If the goal is the page itself, skip the clicking: Try Begin.sh - turn a prompt or a URL into a downloadable static site or Expo app. Describe the site or paste a URL to clone, download the zip and host it wherever you like. There is no hosting, backend or auth included, which keeps it honest for static and prototype work. Keep Computer Use for the apps that genuinely have no other door.

Sources

  • Codex computer use documentation: learn.chatgpt.com/docs/computer-use and developers.openai.com/codex/app/computer-use
  • Three Ways Codex Can Use a Computer, Jason Liu: x.com/jxnlco/article/2066970432855581052
  • Computer and browser use in Codex, Claire Vo on Lenny's Newsletter: lennysnewsletter.com/p/computer-and-browser-use-in-codex

Last reviewed: 2026-09-22

About

Guide to Codex computer use: how the Codex app drives macOS and Windows apps, when to use it, setup and safety notes.

Topics

Resources

Stars

8 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors