COMPUTER USE / PROFILE

Claude Computer Use: capabilities, evidence & fit

Screenshot, mouse, and keyboard tools for a Claude-driven desktop agent loop.

Computer-use modellive

Claude selects actions through a client-executed computer toolset. The developer operates the sandbox, tool handlers, screenshot capture, and agent loop. Browser-only tasks also have a separate documented browser toolset.

Last verified: Sep 27, 2026
Primary sources: 3
Evidence state: Fresh
Benchmark records: No linked results

Where it fits

CONSIDER IT FOR

Desktop workflows in a controlled sandbox
Developers implementing explicit action approval

LOOK ELSEWHERE WHEN

Low latency human-speed interaction is essential
You cannot isolate credentials and sensitive applications

Fit recommendations are editorial interpretations of the documented architecture.

Documented capabilities

CAPABILITY	DOCUMENTED STATE	SOURCE
Desktop input	Mouse and keyboard actions in a desktop environment	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Desktop input: Mouse and keyboard actions in a desktop environment Verified Sep 27, 2026 · verified platform.claude.com
Screen perception	Screenshot and zoom tool members	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Screen perception: Screenshot and zoom tool members Verified Sep 27, 2026 · verified platform.claude.com
Batch actions	Ordered short action sequences can share a response	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Batch actions: Ordered short action sequences can share a response Verified Sep 27, 2026 · verified platform.claude.com
Execution loop	Client-side tool handlers execute requested actions	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Execution loop: Client-side tool handlers execute requested actions Verified Sep 27, 2026 · verified platform.claude.com
Environment reference	The reference setup uses a sandboxed Linux desktop	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Environment reference: The reference setup uses a sandboxed Linux desktop Verified Sep 27, 2026 · verified platform.claude.com
Toolset version	computer_toolset_20260801; compatibility is model-specific	EVIDENCE RECORD Follow the source. Source verification confirms what the publisher states. It is not our independent performance test. official docs Toolset version: computer_toolset_20260801; compatibility is model-specific Verified Sep 27, 2026 · verified platform.claude.com

Environment, deployment & control

Environment

browser: Full
desktop: Full
mobile: Unknown

Deployment

cloud: Full
local: Unknown
self Hosted: Unknown

Extensibility

api: Full
sdk: Full
mcp: Unknown
bring Your Own Model: Unknown

Control method

screenshots: Full
mouse Keyboard: Full
dom Or Accessibility: Unknown
code Execution: Unknown

Unknown = insufficient public evidence; No evidence = an explicit negative record; N/A = does not apply to this layer.

Trust, oversight & limitations

Documented oversight

human Approval: Unknown
audit Logs: Unknown
sandboxing: Unknown
prompt Injection Defense: Unknown
credential Controls: Unknown

Know the boundary

Anthropic notes latency can exceed ordinary human-directed computer actions.
Toolset support varies by model and platform; earlier tool versions remain beta.
Human confirmation and least privilege must be enforced by the application.

Design human approval →

Pricing snapshot

$2 / 1M input tokens · Claude Sonnet 5

Claude Sonnet 5 · Base API rates

Verified Sep 27, 2026 · USD · usage. Hosted services and software licensing can have different terms.

Cost inputs & calculator →

Compare the tradeoffs

OpenAI Computer Use vs Claude Computer Use

Both provide model-side computer-use capabilities. The decision is which provider integration and action surface fit your runtime; neither comparison entry includes the surrounding production system.

Claude Computer Use vs Gemini Computer Use

Both require an execution environment and application controls. Their documented tool surfaces and safety configuration differ, so choose around the required environment and an explicit approval model.

Relevant use cases

Computer use for legacy software

Evaluate native interface control only after checking supported import, scripting, and integration paths.

RESEARCH NOTES

Sources & verification

Reviewed Sep 27, 2026. Architecture recommendations are editorial analysis; linked vendor documentation supports the underlying capability and safety facts.

How we verify evidence →