Source: Anthropic Computer Use Tool Docs

Claude API documentation for the computer use toolset: versions, 17 member actions, coordinate scaling, the batch agent loop, screenshot token costs and security guidance.

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

Summary

Computer use is a client-side toolset: Claude emits mouse/keyboard/screenshot actions, your harness executes them on a (virtual) desktop and returns screenshots. The doc covers tool versions, actions, coordinate handling, the loop contract, context management and safety.

Key claims

  • Versions (as of 2026-11): computer_toolset_20260801 (no beta header; Fable 5.2, Mythos 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5, which return 400 for older versions), computer_20251124 (beta header computer-use-2025-11-24; Opus 4.5-4.7, Sonnet 4.6), computer_20250124 for earlier models.
  • 17 member tools: screenshot, zoom (full-resolution crop of a region), five click types, left_click_drag, mouse_move, left_mouse_down/up, cursor_position, scroll, type, key (with repeat), hold_key, wait. Members can be disabled via configs.
  • Coordinates are screenshot pixels; if you downscale screenshots you must scale clicks back up. On macOS Retina, halve coordinates (2x DPR).
  • Image caps: Opus 4.7+ 2,576 px long edge / 4,784 visual tokens; earlier 1,568 px / ~1.15 MP. Oversized screenshots are rejected, not downscaled.
  • Recommended resolutions 1024x768 or 1280x720 (desktop), 1280x800 or 1366x768 (web apps); avoid above 1920x1080.
  • Claude may return several actions per turn: run in order, stop at first failure, mark later ones with is_error and a fixed “Not executed” message; every tool_use needs a tool_result. Human approval must check each block, because one batch can complete a multistep action.
  • Each screenshot costs about 1,000-1,800 input tokens. Prune old screenshots in batches (e.g. keep last 3, prune every 25 turns) to preserve prompt caching; on Fable/Mythos 5.1-5.2 and Opus/Sonnet/Haiku 5.5, pruning invalidates every later thinking block, so avoid it there (or set prefix_mismatch_behavior: "drop_block").
  • Security: dedicated VM/container with minimal privileges, no credentials or sensitive data, domain allowlist, human confirmation for consequential actions; classifiers scan screenshots for prompt injection; get end-user consent.
  • Tips: explicit steps, screenshot after each step to verify, prefer keyboard shortcuts for dropdowns and scrollbars, put instructions before the screenshot, use zoom for small text.
  • Reference implementation: Docker demo with Xvfb, a window manager, Firefox and LibreOffice.

Feeds

Computer-use agents · Multimodal models

Suggest changes →