AI Agents · Codex · Quick Guide

Three Ways Codex Can Use a Computer

Computer Use, Chrome, and the in-app browser. Three surfaces, one rule for picking the right one.

Jason Liu gave Codex one job: sit in an Amazon support queue and get the refund for a stolen package. Check the chat every five minutes, switch to every minute when a human shows up, do whatever it takes. Then he went to take a shower. He came back to a finished refund.

That’s the shift. Not a chatbot answering a question, but software that opens your apps and gets the work done. Codex does it three ways, and they overlap just enough to be confusing. Here’s each one, when to reach for it, and the rule that keeps them straight.

The rule: prefer a structured tool whenever one exists. A plugin, an MCP server, a CLI, a connector. A Slack plugin pulls a thread more cleanly than clicking around Slack; a GitHub plugin gives you actions you can actually inspect. (Haven’t wired AI into your stack yet? Start with the five connectors worth setting up today.) Visual control is for the boundary where structured tools stop.

1. Computer Use (@Computer): the whole desktop

This is the powerful one. Codex sees and operates your screen (windows, menus, keyboard, clipboard) on macOS and Windows, in the apps you approve. It’s the broadest surface and the slowest, because it has to look, click, then check the result. But it runs anything, including the apps with no API and no plugin. And on macOS it works in the background: you’ll open an app and realize Codex has been quietly grinding through a workflow the whole time.

For marketers, that’s the messy long tail no integration ever covers:

  • Pull spend out of a legacy DSP or local-TV buying tool whose only “export” is a clunky desktop dashboard.
  • Batch-export 40 creative variations from a design app that has no API.
  • Reconcile last month’s numbers across QuickBooks, a Google Sheet, and three ad dashboards that don’t talk to each other.
  • Run a competitor’s iOS app in the simulator and screenshot their full onboarding funnel.
  • Chain several apps at once: grab data from a desktop analytics tool, drop it in Sheets, and post the summary to Slack, all while you’re in a meeting.

Turn it on: Settings → Computer Use → Install. Trigger with @Computer. It’s also the broadest trust boundary of the three, so give it one clear app or task at a time and stay present for anything touching money, accounts, or credentials.

2. Chrome extension (@Chrome): your signed-in browser

Codex drives your real Chrome: your cookies, your profile, your logged-in tabs. It treats the job as a browser workflow, not a list of screen clicks. The edge is multi-tab: read context in one, compare it in a second, act in a third.

This is where the weekly reporting chore goes to die. Google Ads in one tab, Meta Ads Manager in another, LinkedIn Campaign Manager in a third, GA4 in a fourth, all done. Ask for this week’s spend, CPA, and ROAS in one summary, then go do something that actually moves the number as it can be slow.

  • Use it for: cross-channel ad reporting, prospecting and ABM (one founder had it click into every event registrant’s LinkedIn to qualify them), auditing competitor pages while signed in, cleaning a CRM view.
  • Turn it on: Plugins → add Chrome, follow the setup. Trigger with @Chrome.
  • Mind the boundary: sites read Codex’s clicks as you. Let it research and draft. Make it wait for your review before it sends, buys, posts, or submits. (Here’s the line not to cross with AI and your brand.)

3. In-app browser (@Browser): the build-and-debug loop

A browser that lives inside the Codex thread. It’s isolated: no cookies, no profile, no extensions. Wrong for logins, right for building. You and Codex see the same rendered page, so it can edit code, reload, screenshot, and repeat.

For marketers, this is your landing-page and microsite room. Hand Codex a campaign brief, have it build the page, open it, then click straight on the hero: “this headline is buried, make it the first thing.” It edits and reloads. The page becomes the brief, which feels a lot closer to working next to a designer than trading screenshots.

  • Use it for: building landing pages and one-page client reports, QA’ing your site’s mobile layout, reproducing a checkout bug, previewing an email template’s responsive behavior.
  • Turn it on: Plugins → add Browser. Trigger with @Browser.

Bonus: Appshots

Not a fourth way to act, a way to point. On a Mac, tap CMD+CMD to hand Codex the front window (an error, a competitor’s ad, a design) as context, with no control granted.

Quick pick

Browser task that needs your login? @Chrome. Building or debugging a page? @Browser. An app with no API, or a job across the whole desktop? @Computer. And if a clean plugin or connector can do it, use that first.

Resources

Official (OpenAI):

Related from MarketingAlec:

Keep going.

I write about using AI for real work, minus the hype.

More at marketingalec.com →