Claude Now Has Its Own Browser, and It Is Not Yours
Claude Cowork on the desktop opens a browser in a side panel to read pages, click, and fill forms. It cannot see your tabs, bookmarks, or passwords, and banking and email sites stay out.
Anthropic has put a browser inside Claude Cowork, the desktop app version of Claude. When a task needs a website, a browser opens in the side panel and Claude navigates it: reading pages, clicking, typing, filling forms. It is rolling out this week to Pro, Max, and Team plans on macOS, Windows, and Linux, and Enterprise administrators can switch it on now.
Until now, letting Claude use the web in Cowork meant handing it your own browser through the Claude in Chrome extension. The new one is separate. Claude cannot see your tabs, your bookmarks, or your saved passwords. If you want it signed in somewhere, you move logins over one site at a time from Chrome, Edge, or Firefox, and banking, email, and single sign on sites are excluded unless you deliberately include them.
Anthropic frames the two options by what the task needs. The built in browser is for work you hand off and walk away from: collecting this month’s invoices from a supplier portal, pulling numbers out of a dashboard, gathering research. Claude in Chrome is for the page already open in front of you, with the accounts you are already signed into. If you use the extension already, it stays your default; otherwise the built in browser takes over. You can switch under Settings, Cowork, Preferred browser.
What’s actually going on here: an agent that can only read the open web is limited to what has an API or a public page. Most of the tedious work people actually want automated sits behind a login on a portal nobody ever built an integration for. Giving the agent its own browser is the cheapest way past that, and it removes the awkward step where you had to install an extension and grant it access to your entire browsing session.
The tradeoff is one Anthropic names plainly in its own post. A browser agent is exposed to prompt injection, where instructions hidden inside a web page try to hijack what the agent does next. The built in browser runs the same safeguards as Claude in Chrome, including checks that compare Claude’s actions against what you actually asked for, and Anthropic says those measures reduce the risk but cannot remove it. Their advice is to start on sites you trust. That is honest, and it is also a real limit worth respecting.
What this means for you: if you are curious about AI but have never let one act on your behalf, this is a lower stakes way in than an extension that sees your whole browser, and you can start with something reversible like collecting links or reading a public dashboard. If you already run agents, the useful detail is the split: your browser for the page in front of you, Claude’s browser for the errand you delegate. One caveat before you get excited. The browser lives in the desktop app, so driving it from your phone or the web only works while your computer is open and online.
Sources
Google's Video Model Gets a Cheap Draft Mode, and the Price List Is the Story
Gemini Omni 1.1 Flash adds 40-second scene extension, start and end frames, and 4K upscaling. The interesting part is a 360p preview at a third of the cost.
