Anthropic launches a browser use tool for the Claude API

Anthropic has launched a new browser use tool for the Claude API, giving developers a way to have Claude drive a browser inside their own application rather than controlling an entire desktop.

What's new

"The browser use tool lets Claude navigate, read, and interact with webpages in a browser that your application runs," Anthropic's documentation states. Unlike the existing computer use tool, which works with a whole desktop purely through screenshots and coordinates, the browser use tool works two ways at once: through the page's structure — its accessibility tree, elements, forms, and tabs — and through pixels, via screenshots and viewport coordinates.

It ships as an Anthropic-defined client toolset called browser_toolset_20260801. A single entry for it in a request's tools array gives Claude 27 member tools by default, including navigate, read_page, left_click, and screenshot, with four more — javascript_exec, file_upload, read_console, and read_network — available as opt-in additions. The toolset is available now for Claude Fable 5, Claude Mythos 5, Claude Opus 5, Claude Sonnet 5, and Claude Opus 4.8 on the Claude API.

Anthropic shipped the browser use tool alongside a related change: the existing computer use tool graduated out of beta the same day, as the computer_toolset_20260801 toolset. That toolset no longer requires a beta header, adds batch actions so several actions can run in one turn, enables zoom by default, and adds per-member configuration through a configs field. Earlier beta versions of the computer use tool remain available for existing integrations.

Context

Computer use — letting a model see a screen and issue clicks and keystrokes — has been part of Anthropic's agentic tooling since it first shipped in beta, aimed at letting Claude operate existing software the same way a person would. The browser use tool narrows that idea to the browser specifically, trading whole-desktop generality for tighter, structure-aware control of a single application surface. Reading the accessibility tree and DOM elements directly, rather than inferring everything from pixels, tends to make an agent more reliable at tasks like filling forms, following links, or reading tables, since it isn't solely dependent on the model correctly interpreting a screenshot.

Why it matters

Browser automation is one of the most requested capabilities for agentic coding and research tools, and Anthropic shipping a first-party, structure-aware toolset — with granular member tools rather than a single monolithic "click and screenshot" loop — gives developers finer control over cost, latency, and reliability than screenshot-only automation allows. Pairing the launch with the computer use tool's exit from beta also signals Anthropic now considers both approaches production-ready: teams building agents that need to operate a browser inside their own app have a purpose-built option that doesn't require standing up full desktop control.