ChatGPT's browser extension can bring an open Chrome tab into a desktop conversation. That is useful for a narrow question about a page already on your screen, provided you check the answer against the page itself. This guide follows OpenAI's documentation as read on October 18, 2026. BIG CHANGE did not install the extension or test an answer.

Our earlier Agents API browser-session guide covers a developer workflow in a hosted browser. The desktop Chrome procedure here uses an open tab in the reader's browser profile.

The big change

  • What changed: OpenAI documents a way to use a tab in your regular Chrome profile as context for a ChatGPT Work or Codex task. Selecting @Chrome chooses the browser; mentioning one open tab separately identifies the page you want to discuss.
  • Why it matters: Someone reading a long page can ask for one stated fact and inspect the supporting passage without copying the whole page into a prompt. The answer still needs human checking, especially on a signed-in or changing website.
  • What to watch: Browser availability varies with rollout and workspace settings. Website permissions govern where ChatGPT can work, while the extension's browser permissions are broader. Check both before using sensitive pages.

Set up Chrome for a desktop chat

You need the ChatGPT desktop app, an available ChatGPT Work or Codex chat, Chrome, and the extension in the Chrome profile you intend to use. OpenAI says browser availability can depend on rollout and workspace settings, so these instructions do not establish that any particular account has access. The browser extension guide did not show a version identifier when checked on October 18, 2026.

  1. Update or install the ChatGPT desktop app. In Settings > Computer Use, select Chrome and follow any prompt to install the required plugin.
  2. Select Install beside Chrome, install the extension from the opened store page, and review Chrome's permission request.
  3. Return to Computer Use and check that Chrome says Manage. Use the Chrome profile in which you installed and enabled the extension.

These are documented setup steps, not steps BIG CHANGE performed.

The extension can work with sites where you are signed in. Before using it for a private page, consider whether that page belongs in the chat context. OpenAI says the installation prompt can include permission to read and change website data, browsing history and bookmarks, manage downloads, and view and manage tab groups. A narrow question in your prompt does not narrow those extension permissions or make the task read-only.

Ask one question about one tab

  1. Open the page in Chrome and choose one factual question the page should answer explicitly.
  2. Start a new ChatGPT Work or Codex chat. In the @ menu, choose Chrome to select the browser.
  3. Separately mention the specific open tab to provide that page as context. Then ask the question.

OpenAI documents the browser choice and tab mention as separate actions; it does not describe a universal tab-picker layout across app builds.

After selecting the browser and mentioning the tab, a bounded request could read:

From this one mentioned Chrome tab, answer this question: [insert one factual question that the page itself addresses]. Give the page title and URL, identify the passage that supports your answer, and say when the page does not establish the answer. Do not use other tabs or take an action on the website.

Replace the bracketed text with your actual question. The wording sets the task you are requesting; it is not a technical access restriction or a guarantee that the answer will follow it. OpenAI's Computer Use guidance recommends scoped tasks and reviewing permission prompts because computer use can affect state outside a project workspace.

The expected result is a short answer, the identified page title and address, and a passage you can locate on that page. This is an expected format, not an observed response. If the answer needs a date, name or number absent from the page, it should say the page does not establish it.

Compare the answer with the page

  1. Look at the open tab yourself. Compare the answer's title and URL with the page and address bar.
  2. Find the cited passage on that same page. Check whether it supports every factual part of the answer, including qualifiers, dates and units.
  3. If the page supports the whole answer, use that answer with its source. A matching title alone is insufficient; the passage must support the claim.

OpenAI's general ChatGPT guidance tells readers to review results and correct them; the page check turns that advice into a specific acceptance test for this task.

If the passage is missing, contradicts the answer or belongs to another page, do not reuse the answer as a fact. Ask ChatGPT to locate the exact supporting text from the mentioned tab or to state that the page does not answer the question. Recheck the page yourself. For a current claim, also confirm that the page remains the right and current source; a page can change after a chat response. The successful outcome is traceable evidence, not simply a fluent answer.

Decide what the website may share

OpenAI says ChatGPT asks before interacting with a new website by default, with the prompt based on its host. The choices are Allow once for one use, Allow for this site for repeat use without another prompt, Allow for all sites to stop site prompts, or Decline. Choose according to the page and task. Manage in Computer Use lets you review allowed and blocked domains; the browser toggle separately controls whether Chrome appears in the @ menu.

OpenAI says browser content such as page text, screenshots, tool calls and summaries is stored when it becomes part of ChatGPT context. Browser history has a separate request, and relevant entries can enter context if you allow it. Treat a page's instructions as untrusted material and review the site and any new permission prompt. Avoid putting secrets or highly sensitive material into a browser task unless it is necessary and you can review the interaction. The manual source check helps catch unsupported claims; it does not change the extension's permissions or the site's access setting.

If Chrome is missing or the answer cannot be checked

If ChatGPT cannot use the page, first open Settings > Computer Use > Manage beside Chrome and check whether its host is on the blocklist. OpenAI says a blocked host is one ChatGPT should not use; removing it lets ChatGPT ask again. Review the website before changing that choice. If Chrome itself does not appear in the @ menu, confirm its toggle is on under Settings > Computer Use. If setup still says Install, follow the extension setup again.

For a connection problem that remains, OpenAI's recovery sequence says to update the desktop app, restart Chrome, confirm that Computer Use shows Manage, check the active Chrome profile, try a new Work or Codex chat, restart the desktop app, and reinstall through Computer Use if it still cannot connect. Its final support route is /feedback with the chat ID. Do not interpret a failed connection as evidence that a page contains or lacks the requested fact.

If the connection works but no verifiable passage appears, narrow the question to text you can see, or stop with “not established by this page.” The procedure cannot confirm how well the extension answers in a particular account. No hands-on browser run or answer-quality test was performed for this guide.

OpenAI's pricing page describes plan and model usage limits, which vary by plan and task. The checked pages did not state a separate numeric extension charge. Check your plan's current usage information before a long browser task rather than assuming a particular allowance.

Sources & further reading

  • OpenAI, Browser extension. The primary setup and behavior source: browser selection, tab mentions, site prompts, extension permissions, data handling and connection recovery. Read October 18, 2026; the live page showed no version identifier. It documents capability, not a BIG CHANGE test or universal rollout.
  • OpenAI, Use ChatGPT. Explains supplying context and reviewing and correcting a result. Read October 18, 2026. The title, URL and passage check here is BIG CHANGE's proposed reader procedure.
  • OpenAI, Computer Use. Establishes desktop app control, scoped tasks and permission review. Read October 18, 2026. It does not certify that this prompt prevents browser actions.
  • OpenAI, Pricing. Describes plan and model usage limits. Read October 18, 2026. No separate extension price was established from these checked pages.