What Codex Can Actually Do With Computer Use and Browser Use

For months, Codex was good for writing codes. You typed a prompt, it handed back a function or a pull request, and the clicking around your screen stayed your job.

That changed recently. OpenAI turned Codex into something that can see a screen, move a cursor, and act inside your browser the way you do. It is not just a coding tool anymore. It is a general work agent that happens to be good at code too.

Non-developers now make up roughly 20% of weekly Codex users, with adoption across business roles growing nearly three times faster than in software engineering. People are delegating real administrative, testing, and operational work directly to AI agents. Let’s understand how Codex computer use and browser use works, and what you can do with it.

One Topic: Codex Can Use Your Browser and Computer, Now

The Three Ways Codex Touches Your Screen

Codex now works through 3 control modes, and knowing which one is running matters more than most people realize.

The In-App Browser is a sandboxed window inside the app. It does not know your logins or cookies, so it is built for local previews and public pages. It is the same feature I covered when Codex first added Sites, plugins, and annotations, just more capable now. So you can login and it will take it further.

Browser Use, powered by the official Chrome extension, rides inside your actual, signed-in Chrome profile. It can open Gmail, LinkedIn, or an internal dashboard without you typing a password.

Codex Computer Use is the biggest jump. It reads your whole screen like a person would, works out where to click, and runs real desktop apps like Excel, Figma, or a mirrored iPhone window.

One point clears up much of the confusion: the built in Browser is not available through Codex CLI or IDE. Use the ChatGPT desktop app for these workflows.

My rule is simple: use a plugin when one exists, Chrome for authenticated browser work, Browser for local or isolated web work, and Computer Use only when the task must cross into a mac app.

7 Real Example You Can Do Codex Computer Use and Browser Use

1. Run full QA sweeps. You can run Codex through your site’s entire signup flow. It can catch eleven real bugs, including a broken submit button, and log each one into a Google Sheet with steps to reproduce.
2. Test as a fake user. Give Codex a persona (a skeptical buyer, a confused first-timer) and have it try to complete a task as that person would. It will find confusing copy and dead ends that standard testing misses.
3. Clear out a message inbox. Through the Chrome extension, Codex can read your incoming Facebook, Linkedin (or any other where you don’t have clear APIs) messages, sort them by priority, and draft replies. You just review and hit send.
4. Shop inside a budget. Tell Codex to buy a lens filter or a camera bag based on your camera or the flight you’re travelling on. It will filter a store, fill the cart, pause at a CAPTCHA, and pick back up once you solve it.
5. Cross-SaaS Executive Briefings: Every morning, background scripts can launch Codex to pull pending tasks from Slack, project updates from Notion, and pipeline numbers from Salesforce, consolidating everything into a clean daily Slack briefing.
6. Visual Element DOM Editing: Web designers can open a local site preview inside Codex, draw a box around a misplaced button, and request a visual change. Codex links that exact visual region to the underlying CSS and React code and updates it automatically.
7. Write a caption: Point to a folder where you edited images in Lightroom, ask it to select one, and write a caption in your voice based on the elements and metadata. It will be ready to post on social media.

None of these needed a developer sitting in the loop. That is the real shift.

The Guardrail You Cannot Skip

Every example above still had a human checkpoint: approving a message, solving a CAPTCHA, reviewing a bug log. That is by design. Codex runs in a sandbox by default, and full, unrestricted access is meant for isolated test machines, never your main computer. Keep the checkpoint. It is the difference between an assistant and a liability.

How to Try This Today

1. Open the Codex desktop app and go to Settings, then Plugins.
2. Install the Chrome extension for anything that needs your logged-in accounts or Computer Use plugin to control your mac.
3. Start small: fill one form or check one dashboard.
4. Watch where it pauses. That is where your own review should sit going forward.
5. Once you trust one workflow, save it as a repeatable automation.

For a different take on this same hand-off, see how Hermes Desktop Agent approaches it.

Interested in travel or photography, read last week’s LensLetter newsletter about how AI can help plan your landscape photography travel.

Read last week’s JustDraft about GPT-5.6-Sol and Other New Models.

Two Quotes to Inspire

Strategy is not choosing what to do next. It is choosing what you will finally stop doing yourself.

Leadership in the age of AI agents is no longer about managing manual workflows, but about designing clear boundaries for autonomous execution.

One Prompt to Steal

Copy this into Codex with the Chrome extension turned on:

Open [your website URL] as a first-time visitor who has never used a product like this before. Try to complete [specific goal, like "sign up and book a demo"]. At every step, note anything confusing, slow, or unclear. When you're done, list the three biggest friction points in order of how much they would make a real user quit.

Swap in your own site and goal. It catches friction a person too close to the product never notices.