Anthropic’s new browser tool doesn’t actually run a browser
The New Stack Amanda Caswell
Anthropic added Browser Use for Claude, but it doesn’t run the browser itself. It reads page structure instead of screen pixels, which should make clicks less clumsy.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has shipped a new Browser Use tool for Claude, and the clever bit is what it doesn’t do. It does not spin up a browser of its own. Instead, it gives Claude a structured view of the page through the accessibility tree, so the model can target links, buttons, and text boxes by reference rather than by guessing where they sit on the screen.
That matters because the old way is messy. If Claude only sees a screenshot, it has to infer positions from pixels and then aim with coordinates. With Browser Use, a button can come back as something like ref_3, and Claude can ask for that element directly. It is a cleaner fit for browser work than desktop-style control, where the model is moving a mouse around a full screen.
The catch is that those references are fragile. If the page changes or navigation takes the tab somewhere else, yesterday’s ref_3 may be useless. Anthropic says the executor has to notice when that happens, reject the action, and make Claude read the page again before continuing. So this is not magic autonomy. It is a tighter loop with more structure, but still a loop.
Anthropic is also pushing multiple browser actions through in one model turn. Instead of bouncing back to Claude after every click or keystroke, the app can receive several tool_use blocks, run them in order, and send the results back together. That should cut latency and token spend, especially when a task stretches from a few interactions to dozens or hundreds. Anthropic’s pricing docs say the default Browser Use toolset adds about 6,600 input tokens before screenshots, accessibility trees, and other return data.
Developers get the tool through the Claude API, using browser_toolset_20260801, and it is not available inside Claude Managed Agents. They still have to host the browser themselves, keep the session alive, and translate Claude’s requests into real browser actions. Anthropic says that browser state, downloads, and uploaded files stay in the developer’s environment. It is a sensible split, if a slightly annoying one: Anthropic gets the model logic, and developers get the joy of keeping the browser from wandering into trouble.
My take — AI-written commentary, not fact-checked reporting
This is the right shape for browser automation: structure first, blind clicking second. The hype machine loves “agents,” but the real bottleneck is still state, permissions, and cleanup, which means someone has to own the mess. Also, calling a tool Browser Use when it does not actually run a browser is exactly the kind of naming that keeps AI product demos mildly exhausting.
Read more about this at: The New Stack
