Urgent.News

What's breaking now, across thousands of outlets.

Editions

AI

Anthropic’s new browser tool doesn’t actually run a browser

Anthropic launched a new Browser Use tool that gives Claude a structured view of a web page in addition to The post Anthropic’s new browser tool doesn’t actually run a browser appeared first on The New Stack .

Anthropic’s new browser tool doesn’t actually run a browser

Anthropic has introduced a new Browser Use tool that enhances Claude's ability to interact with web pages. This tool provides Claude with a structured view of a webpage, allowing it to interact with elements directly, rather than relying on visual rendering or calculating positions based on pixel data. This is part of a broader release by Anthropic that also includes Computer Use, the Skills API and Files API.

Unlike traditional methods that involve pinpointing elements by coordinates, Browser Use utilizes the page's accessibility tree. When Claude requests to interact with a specific element, it can use a reference (such as "ref_3") associated with that element, simplifying the interaction process.

However, the tool has limitations. If the page changes or the tab navigates to a new page, references may become invalid. Developers need to monitor for such changes and instruct Claude to read the page again to maintain functionality.

Anthropic's Browser Use operates through a specific client-toolset protocol, distinct from other browser automation tools like Playwright MCP or Puppeteer. While these tools may offer similar functionality, they use different communication protocols, necessitating an adapter for developers to integrate them with Claude's system.

A key improvement in this release is Anthropic's ability to batch multiple actions into a single model turn. This allows Claude to request several operations simultaneously, reducing the number of back-and-forth communications needed. For instance, a sequence of actions like clicking a button, filling in a field, and submitting a form can be sent as a single batch, cutting down on model calls and associated costs.

Despite these advancements, developers must host the Browser Use tool themselves and integrate it into their Claude processes. The tool is currently only accessible via the Claude API and not directly within Claude Managed Agents. Additionally, default Browser Use operations come with a significant token cost, so developers can optimize performance by selectively disabling unnecessary operations.

Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at thenewstack.io →

More in AI

AI Is Learning to Write Genetic Code

This sort of research is both exciting and terrifying: The two models in question were told to generate complete genomes for a viable bacteriophage—a type of virus able to infect and replicate itself…

More from Friday 21 August →