All capabilities Agents, tools and MCP

Operate a desktop application

Screenshot, move, click, type. The model works a native application the way a person does, for software that never got an API.

DocsOfficially documentedTestNot tested by benchr Deep · API Docs checked September 1, 2026

Claude APIClaude Fable 5Claude Opus 5Claude Sonnet 5+1

Overview

Provides 17 member tools covering display, mouse, keyboard, scrolling and waiting.
Works against the whole desktop rather than one page - file dialogs, installers, legacy line-of-business software.
Supports a zoom action, so the model inspects a region instead of squinting at a full screenshot.

01Use computer_toolset_20260801 on a supported model. Claude 4.5 through 4.7 reach computer use only through the earlier computer_20251124 version, behind a beta header.
02Keep the virtual display at 1920x1080 or below - the documentation names this as the recommended ceiling to avoid performance problems.
03Scale coordinates yourself if you resize screenshots, because coordinates are in screenshot pixel space.
04Trim screenshot history aggressively. Each screenshot costs roughly 1,000 to 1,800 tokens and a long loop accumulates them fast.

For anything that stays inside a webpage the vendor recommends the browser tool instead.
Screenshots must fit the model's image limits; the API rejects an oversized tool_result image rather than downscaling it.
Security risk is heightened on internet-facing tasks, and the docs require sandboxing plus prompt-injection precautions.
Not available in Claude Managed Agents.

benchr has not run this capability. No test has been published.

benchr test runs →

Source-check history

September 1, 2026DocsOfficially documentedFirst logged. Tool version, the 17 member tools, the 1920x1080 recommendation and the screenshot token range all read from the vendor page.