ChatGPT Computer Use in Chrome: How to Prompt the Browser Agent

OpenAI just put a browser agent inside the browser most of us already use. On October 15, the ChatGPT account announced "the power of computer use" for Chrome: in Chat, ChatGPT can look across your open tabs and gather info; in Work, it can control your Chrome tabs — sending messages, filling out forms, downloading files — and even organize and group the pile of tabs you have open. The extension listing shows 7,000,000 users and an update dated October 15, 2026.

The underreported half of that thread matters more for prompting: the full power of ChatGPT now lives inside a Chrome tab — your entire chat history, with Chat questions or Work tasks running side-by-side with whatever site you're on.

What Actually Shipped

OpenAI's docs describe three computer-use surfaces you invoke with an @-mention: @Computer for desktop apps, @Browser for the built-in browser in the desktop app, and @Chrome for your regular browser — including sites where you're already signed in. The Chrome plugin can start a new tab, read an internal dashboard, use page controls, or fill out a form.

Why this is a different product from what came before: ChatGPT's cloud browser — the one that runs tasks in the background on web and mobile — cannot access your local files, open tabs, extensions, saved passwords, sign in to sites, or complete payments. The Chrome extension is the opposite design: it works because it borrows your existing profile, signed-in session, and open tabs. That's also why prompting it well matters. You've handed a language model your logged-in self.

If you've followed this space, the shape of the race changed too. Claude's tooling puts a browser inside a coding agent, Gemini made computer use a native tool in a fast, cheap model, and OpenAI now distributes one through an extension already installed by millions. Same capability, three very different places to live.

A day earlier, OpenAI also rolled out ChatGPT Space, introduced at DevDay on September 29: shared living documents that teammates and ChatGPT edit together, on Pro, Business, and Enterprise plans. It's a separate product from the browser agent.

Setup, in Four Steps

From the official walkthrough:

1. Open the ChatGPT desktop app, choose Work, open Plugins.
2. Install the Chrome plugin.
3. Install the ChatGPT extension from the Chrome Web Store and approve the permission prompts.
4. Confirm the side chat loads — on macOS you can open it beside any page with Cmd+Shift+.

Two settings are worth five minutes before your first real task: Settings > Computer use > Manage next to Chrome, where you review which websites ChatGPT may touch; and the site-permission prompts you'll get per site. If you're new to the app, our Getting Started with ChatGPT guide covers the rest of the basics.

The Prompt Structure That Works

A browser agent doesn't respond to "optimize my inbox." It responds to a bounded task with explicit edges. OpenAI's launch thread shows comparing couch listings and explaining an article in plain English; the official walkthrough adds a week-over-week dashboard comparison. Every prompt that survives contact with a real browser does five things:

1. Names the surface when it matters — @Chrome versus the built-in browser, because they have different permissions and different browser state.
2. States one goal, not a vibe. "Find the invoice" beats "tidy up my email."
3. Scopes the tabs: "the three tabs I have open," "this dashboard," "only the Gmail tab."
4. Sets stop conditions: what the agent must not do — send, submit, buy, delete.
5. Demands a checkpoint before irreversible actions and a named output format.

The five-part structure generalizes what Anthropic documented for GUI agents in our computer-use prompting guide — describe targets concretely, sequence actions, define recovery — and pairs with the human-in-the-loop patterns for deciding which actions need a human yes.

Eight Copy-Paste Prompts

1. Cross-tab comparison
@Chrome, compare the three product tabs I have open in a table:
price, dimensions, delivery time, return policy. Don't buy anything.

2. Article breakdown
Explain this article in plain English. What's the main argument,
and what evidence supports it?

3. Dashboard readout
@Chrome, look at the dashboard I have open. Compare this week with
last week and explain the biggest changes.

4. Form filling with a gate
@Chrome, fill the application form in this tab using the details
below. Show me each field before submitting, and wait for my OK.

5. Signed-in retrieval with a stop condition
@Chrome, open Gmail, find this month's invoice from Acme, and
download the PDF. Do not reply to, archive, or delete anything.

6. Tab triage
@Chrome, sort my open tabs into three groups — read-later,
reference, close — and list what's in each before you move any tab.

7. Draft, don't send
@Chrome, open the message from Sarah and draft a reply about the
demo. Leave it unsent — I'll review the wording first.

8. Structured extraction
Read the pricing pages I have open and build a table: plan name,
monthly cost, seat limit, API access. Cite the page each cell
came from.

Prompts 1, 2, and 3 are OpenAI's own published examples, lightly expanded. Prompts 4, 5, 6, and 7 add the gates that make signed-in browser control safe to use daily. Prompt 8 turns "look at this" into reusable output.

What It Can't Do — Yet

• Not every browser gets the full experience. The extension targets Chrome; OpenAI's browser-support doc lists Edge, Brave, Opera, and Vivaldi as working, but side chat is supported everywhere except Opera, where you start tasks from the desktop app.
• File uploads need an extra permission — allow the extension to access file URLs before asking ChatGPT to attach a local file to a web form.
• Page content is untrusted by design. OpenAI's own guidance: treat what's on the page as untrusted context and review the site before allowing ChatGPT to continue. A page can lie; a prompt with a stop condition is your backstop.
• Availability is region-gated. OpenAI says Computer Use is available "in supported regions" on macOS and Windows in ChatGPT Work and Codex — and hasn't published the list for the Chrome tab features.
• Plan requirements for the Chrome update are unstated. Space is confirmed for Pro, Business, and Enterprise only; neither the announcement nor the docs state plan gates for the extension's tab controls, so treat Free/Plus availability as unconfirmed until OpenAI says so.

Where This Leaves You

On a Mac, the agent is one Cmd+Shift+. away. Start with the read-only prompts (1, 2, 3, 8) for a week — they teach you where the agent misreads your intent without risking a mis-sent message. Add the gated ones (4, 5, 6, 7) only with the confirmation steps left in. The stop condition you write in the prompt is the only permission system you fully control.