Source: Anthropic Browser Use Tool Docs

Claude API documentation for the browser use toolset: 31 member tools mixing accessibility-tree refs and coordinates, tab state reporting, and a hardening checklist.

Contents
  • What: official API documentation (Browser use tool)
  • Author/publisher: Anthropic
  • Date: living document; read 2026-11-14
  • Original: Browser use tool

Summary

browser_toolset_20260801 is a client toolset for driving a browser your application runs. Unlike computer use, it sees page structure (accessibility tree, element refs, forms, tabs) as well as screenshots.

Key claims

  • 31 members (27 on by default): navigation and capture, pointer actions, keyboard/timing, page reading (read_page, find, get_page_text), forms, tabs; file_upload, read_console, read_network, javascript_exec are off by default.
  • Targets are either viewport coordinates or ref handles from read_page/find. Refs go stale when the DOM changes and the API cannot detect it; your executor must return errors. Prefer refs; fall back to coordinates for canvas, video, virtualized lists, cross-origin iframes.
  • A tree read often costs fewer tokens than a screenshot. read_page output is capped at 50,000 characters.
  • Executor reports tab inventory in a browser_state block (max 100 tabs).
  • Same batch contract as computer use: in order, stop at first failure, answer every call.
  • Use computer use for whole-desktop tasks; browser use for web-confined tasks; web fetch/search when no interaction is needed. Both toolsets can be declared together.
  • Hardening: dedicated container with fresh profile and no credentials; network-layer domain allowlist re-checked after redirects; block private IP ranges; treat page content as untrusted; only http(s) in navigate; keep javascript_exec and file_upload off; human confirmation for purchases, account changes, messages, accepting terms.
  • javascript_exec runs with the page’s full privileges including cookies.

Feeds

Computer-use agents

Suggest changes →