Pick the Browser Before the Task

Claude can drive a browser for me in two ways, and this week I learned they aren’t interchangeable. A survey at the end of a long web form loaded its last page and went white. No button, no error. Claude had been running it in the built-in browser. In Chrome, the same page would have worked.

Claude in Chrome is an extension that runs inside my own Chrome, with my logins and my Bitwarden. The built-in browser is a separate pane in the Claude desktop app, with sign-ins of its own. I’d been treating them as one tool with two doors. They aren’t.

The Insight

The white screen came down to one pop-up. When the survey finished, it opened a small return page meant to tell the window that launched it “done,” then close. The built-in browser opened that page in the same tab instead. So the page that was supposed to report back had nothing left to report to.

Nothing was broken. The site did its part, and the browser handled one thing its own way. That one difference killed the step. The rest of the week showed the same pattern in smaller ways, and each browser ended up with its own short list:

• Claude in Chrome handles real pop-ups, uses my existing sign-ins and password manager, and attaches files itself. Each chat gets its own tab group, so two tasks can run side by side. The costs: it works in my own Chrome, screenshots can fail while another chat’s tab is in front, it couldn’t reach the upload button inside one embedded form, and with two Chrome extensions connected, Claude has to pick mine by its ID every time.
• The built-in browser stays out of my Chrome entirely, shows in the Claude app’s side panel so I can watch it work, and remembers a sign-in after the first time. The costs: pop-ups open in the same tab, and it can’t attach files, so every upload means me dragging the file in by hand.

The Fix

The survey was a small repair. The answers were saved; only the handoff failed. With my OK, Claude went back to that step and sent the site the same “complete” signal the pop-up would have sent.

The bigger fix is choosing before the task starts. Chrome is my default, because most of what I hand Claude needs a login I already have or a file to attach, and Chrome handles both without me.

The built-in browser earns its place on busy days. Two chats sharing one Chrome is exactly where screenshots start failing, so when two tasks run at once, I split them. The heavier one gets Chrome: anything with pop-ups, a survey, or a lot of uploads. The simpler one, plain forms and few files, goes to the built-in browser, where I sign in once and watch it in the side panel.

The Hard Truth

An agent is only as good as the window it works in. I spent the first part of that afternoon assuming the site was broken. It wasn’t. The browser was the wrong one for that page.

So the question before any browser task isn’t which tool is smarter. It’s which one behaves like a real Chrome for this site. Pop-ups, uploads, and logins answer that faster than anything else. The white screen was the browser telling me I’d picked wrong.