The browser that lives inside ChatGPT

For two years the pattern was the same: work in your browser, copy something, paste it into ChatGPT, read the answer, return to the browser, and apply it by hand. AI lived in a separate tab, and you were the cable between the two windows.

The built-in browser reverses that pattern. The AI app stops being a destination and becomes the place where the work happens. Sheets, Notion, an ads dashboard, App Store Connect, or the page you are building can all open inside the app, with a persistent chat underneath that already knows what you are looking at.

That is the strategic shift. OpenAI is not alone: Anthropic is moving in the same direction with Claude. Both are turning their products into something closer to an operating environment - a place where you work, not merely a tool you consult.

What changes in practice

• From describing to pointing. Instead of explaining which landing-page element feels wrong, select it and write “this.”
• From asking to delegating. The model can do more than comment on a page: it can navigate, click, and type.
• From sessions to continuity. The chat does not reset when you change tabs. It keeps the path you have taken.

It is a Chromium-based browser embedded in the ChatGPT desktop application for macOS and Windows. It has its own profile, tabs, and address bar, paired with a chat panel that can access what is being rendered.

Three common misconceptions

One. “It is the same as searching the web.” It is not. Web search brings back text from results. Here, the model sees the rendered page, including pages behind a session you opened yourself.

Two. “It cannot sign in, so it cannot use my tools.” The documentation says the cloud browser cannot sign in for you. In the desktop browser, you sign in manually once and the session remains available. That is why it can work with Notion, Sheets, or App Store Connect.

Three. “It is only for developers.” It also appears in Codex mode, but annotating a design, auditing a campaign, or working through a spreadsheet does not require code.

The mode matters

The sidebar lets you switch between ChatGPT (“Create, learn, and explore”) and Codex (“Build, debug, and ship”). The browser exists in both. In Codex it also carries the context of your open project or folder, which is what can turn an annotation on a page into a code change.

• Tab bar. Multiple sites can stay open, and the chat can follow the work across them.
• Address bar. Enter a URL or search normally.
• Annotating. Click an element or drag over an area to select it visually.
• Selection. The blue outline becomes the thing you mean by “this.”
• “Do anything” composer. The persistent chat at the bottom already sees the page.
• Model and effort selector. Change the model and reasoning level without leaving the browser.
• Layout controls. Collapse the chat, minimize the panel, or toggle the sidebar.

The most underrated control

It is the button that hides the chat panel. With the panel open you feel like you are “using ChatGPT.” With it collapsed, you are simply browsing and the chat becomes a quiet line at the bottom. That reframing is how the feature becomes a habit: it stops feeling like an AI session and starts feeling like your browser can answer back.

The built-in browser keeps its own profile, separate from the Chrome, Safari, or Arc profile you use every day. Nothing is imported automatically.

What that means

• Your open tabs do not appear. The profile starts empty.
• Saved passwords are unavailable. Sign in manually to the few tools you actually need.
• Extensions do not install themselves. The browser profile has to be configured separately.
• Sessions persist. Sign in once and the session remains.
• Downloads go to the system's default folder unless you change it in Settings.

The privacy reading

The separation works both ways. Activity inside ChatGPT's browser stays in that profile and does not touch your personal browsing profile. The app is not reading your general browser history. The model can, however, see the pages you open inside its browser while the chat is active. Treat it as a workspace shared with AI, not a vault.

1 · Page context

The chat sees what is rendered in the active tab. Ask it to summarize a table, identify what is missing on a screen, or translate a section without pasting anything. It is the simplest capability and probably the one you will use most often.

2 · Visual annotation and Adjust

Turn on Annotating, click an element or drag around an area, and leave the comment in place. It feels like reviewing in Figma, but on the real page. Adjust adds precise controls for typography, text size, spacing, and color so you can specify values instead of negotiating through adjectives.

This block competes with the headline. Lower the weight to 500,
increase line height to 1.6, and use the secondary gray instead
of pure black.

3 · Computer Use: let the agent operate it

With the Browser plugin installed, mention @Browser and ChatGPT Work or Codex can operate the browser: open pages, click, type, inspect rendered state, take screenshots, and verify its own result. Verification is the important part. The agent should look again instead of merely saying it is done.

@Browser open the campaigns dashboard, filter the last 14 days,
and build a table with the five campaigns with the worst CPA.
Show me a screenshot before changing any budget.

4 · Developer Mode (CDP)

Developer Mode can expose Chrome DevTools Protocol so ChatGPT can inspect the DOM, applied styles, console, network traffic, and performance. It is the difference between “the page feels slow” and “this request blocks rendering for 800 ms.” Organizations can restrict the option through policy.

ChatGPT has two browser surfaces: the one embedded in the desktop app, which this guide covers, and a cloud-operated browser used from the web. Many “it will not let me sign in” complaints come from confusing them.

What the system does for you

• Requests permission before entering a new site when the agent navigates on its own.
• Requests confirmation for sensitive actions such as submitting information, purchasing, changing permissions, or deleting data.
• Does not automate file uploads. That is a deliberate limit.
• Keeps Developer Mode opt-in, and organizations can block it through policy.

What you need to do

Nobody changes tools by reading a document. The change happens through one useful repetition each day.

How to know it worked

The signal is not that you use ChatGPT more. It is that you stopped taking screenshots to explain the screen. If you are still capturing everything, you are using the old chat through a new window.

Context stopped being expensive. The implementation details in this guide may change with the next version. The direction will not: AI tools are moving underneath the work instead of sitting beside it. The sooner you adapt your workflow, the sooner you recover that margin.