Claude Computer Use
by Anthropic
AI agent that controls your computer like a human coworker using screen, keyboard, and mouse
Claude Computer Use is a beta Claude API tool that gives the model screenshot vision plus mouse and keyboard control of a desktop, so it can drive legacy or API-less software the way a person would. The sandboxed environment runs in your own infrastructure, not Anthropic's.
Claude Computer Use is a beta capability of Anthropic's Claude API that gives the model screenshot vision plus mouse and keyboard control of a desktop environment, letting it operate software that exposes no API. Developers add a computer tool to a Messages API request — currently type computer_20251124 behind the computer-use-2025-11-24 beta header for Claude Opus 5, Sonnet 5, Opus 4.8, 4.7, 4.6 and 4.5 and Sonnet 4.6, or the older computer-use-2025-01-24 header for Sonnet 4.5 and Haiku 4.5 — declaring display width, height and display number. Claude then returns tool_use blocks describing actions; the developer's application executes them against a sandbox and returns screenshots as tool_result blocks, repeating until the task completes. That repetition is the agent loop, and it is deliberately client-side: Anthropic ships no runtime, so all screenshots, keystrokes and files stay in the customer's environment, which makes the feature zero-data-retention eligible. Anthropic's open reference implementation packages a Docker container running an Xvfb virtual X11 display, the Mutter window manager, a Tint2 panel, Firefox and LibreOffice, plus example tool code and a web interface. Computer use is normally combined with the bash and text editor tools for fuller automation. Launched in October 2024, it reached 72.5% on the OSWorld benchmark with Claude Sonnet 4.6, up from under 15% in late 2024, and in February 2026 Anthropic acquired Seattle computer-use startup Vercept, whose founders joined and whose Vy desktop agent was wound down on 25 March 2026.
Automation and platform teams sitting on legacy enterprise software that exposes no API, where every integration today means screen-scraping, brittle RPA scripts or manual data entry.
Automate a workflow by describing it rather than by building an integration, while the desktop sandbox and all its data stay inside your own environment.
At a Glance
- Category
- AI Agents & Orchestration
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CIOs, CTOs, Enterprise Developers, Automation Engineers, QA Leads
- Deployment
- API-based, Hybrid, Multi-cloud
- Founded
- 2024
Key Features
- ✓Screenshot-driven desktop control
Claude sees the screen as images and issues mouse and keyboard actions against whatever application is currently displayed.
- ✓Dated tool versions behind beta headers
Tool definitions carry explicit dates and beta headers, so a model upgrade never silently changes agent behaviour in production.
- ✓Client-side execution model
All screenshots, keystrokes and files stay in your environment; Anthropic processes them in flight only, making the feature zero-data-retention eligible.
- ✓Prompt-injection classifiers
Classifiers scan screenshots for injected instructions and steer the model to request user confirmation before continuing; opting out requires contacting support.
- ✓Open reference implementation
A Docker container with Xvfb, Mutter, Tint2, Firefox and LibreOffice plus a working agent loop shortens time to first prototype.
- ✓Composable with bash and text editor tools
Combining desktop control with shell execution and file editing covers automation that neither pure GUI nor pure CLI tooling can reach.
- ✓Extended thinking tuning for UI tasks
Anthropic recommends medium thinking on Sonnet 4.6 and Opus 4.6 as the best accuracy-to-cost ratio, and warns max adds cost without improving UI accuracy.
- ✓Multi-cloud availability
Runs on the first-party Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, with regional and data-residency endpoint options.
Capabilities
Use Cases
- •Automating API-less enterprise software
Drive an ERP, claims desktop or terminal emulator that exposes no integration surface, without commissioning a custom connector.
- •Adaptive UI regression testing
Run end-to-end tests that survive layout changes, because the model reads the rendered screen rather than brittle CSS selectors.
- •Cross-system data migration
Move records between systems lacking export APIs by having Claude read one interface and type into the other.
- •Background research and information gathering
Collect data across multiple web applications overnight, where slower-than-human latency costs nothing and manual effort is eliminated.
- •Reproducing customer-reported bugs
Replay a described workflow inside a disposable container to confirm a defect before an engineer spends time on it.
Ideal For
Best For
- ✓Driving legacy line-of-business software that has no API and no supported integration path
- ✓Automated UI and end-to-end regression testing where selector-based scripts break on every layout change
- ✓Background research and data gathering across web applications, where throughput matters more than latency
- ✓Teams with strict data-handling requirements, since the environment is client-side and zero-data-retention eligible
- ✓Prototyping agentic desktop workflows quickly using Anthropic's published Docker reference implementation
Not Ideal For
- ✗Latency-sensitive or high-throughput automation — Anthropic states computer use latency 'might be too slow compared to regular human-directed computer actions' and recommends tasks where speed is not critical
- ✗Any workflow touching credentials or sensitive accounts without human oversight: security researcher Johann Rehberger demonstrated a webpage prompt injection that made Claude download and execute a Sliver C2 implant, achieving full command-and-control
- ✗Tasks demanding perfect precision, since the documentation warns Claude 'might make mistakes or hallucinate when outputting specific coordinates' and that reliability drops on niche or multiple simultaneous applications
- ✗Cases where a real API already exists — each screenshot costs roughly 1,000 to 1,800 input tokens, making GUI driving far more expensive than a direct API call
- ✗Account creation or content posting on social and communications platforms, which Anthropic explicitly limits
Integrations
Deployment
Market Analysis
Pros
- ✓Unlocks automation of software with no API, which is most of the enterprise long tail
- ✓Client-side execution keeps screenshots, keystrokes and files in the customer's environment and is zero-data-retention eligible
- ✓Benchmark progress is steep: 72.5% on OSWorld with Sonnet 4.6 against under 15% in late 2024
- ✓Open Docker reference implementation plus availability on Bedrock, Google Cloud and Microsoft Foundry
- ✓Anthropic shipped prompt-injection classifiers as a defence layer rather than leaving the problem entirely to developers
- ✓Backed by Anthropic's SOC 2 Type I and II, ISO 27001:2022 and ISO/IEC 42001:2023 certifications, with a BAA available
Cons
- ✗Still labelled beta roughly two years after its October 2024 launch, with the documentation itself listing eight categories of limitation
- ✗Prompt injection is a demonstrated risk, not a theoretical one: the ZombAIs research achieved full command-and-control compromise via a malicious webpage that Claude obediently downloaded and executed
- ✗Latency is explicitly called out as potentially too slow for human-interactive work, confining it to background and testing tasks
- ✗Screenshots dominate cost at roughly 1,000 to 1,800 input tokens each, so long agent loops get expensive fast
- ✗Scrolling, dragging and spreadsheet cell selection remain unreliable and may need keyboard workarounds or multiple attempts
- ✗Reliability drops noticeably with niche applications or several applications open at once
Pricing
Claude Haiku 4.5 (API)
From $1/MTok input
- ✓$1/MTok input, $5/MTok output
- ✓Uses the computer-use-2025-01-24 beta header
- ✓Lowest-cost option for simple UI loops
Claude Sonnet 5 (API)
From $2/MTok input
- ✓$2/MTok input, $10/MTok output introductory pricing through 31 August 2026
- ✓$3/MTok input, $15/MTok output from 1 September 2026
- ✓Uses the computer-use-2025-11-24 beta header
Claude Opus 5 (API)
From $5/MTok input
- ✓$5/MTok input, $25/MTok output
- ✓Highest-capability tier for complex desktop tasks
- ✓Optional fast mode at $10/MTok input, $50/MTok output
Enterprise / custom
Contact for pricing
- ✓Negotiated volume discounts
- ✓Custom rate limits beyond the Scale tier
- ✓Dedicated support and custom terms
There is no separate licence fee for computer use — you pay standard Claude API tokens, so cost is driven entirely by the agent loop. The tool definition adds 735 input tokens on Claude 4.x models, the beta itself adds 466 to 499 tokens of system prompt, and every screenshot returned to the model costs roughly 1,000 to 1,800 input tokens, which dominates spend on long-running loops. Prompt caching cuts cache hits to 10% of the base input price and the Batch API halves both directions, though batching does not suit interactive desktop work. On Amazon Bedrock, Google Cloud and Microsoft Foundry the cloud provider invoices instead, with regional endpoints carrying a 10% premium; enterprise volume discounts are negotiated rather than published.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 6 on domains other than platform.claude.com.
- 1.platform.claude.com — computer use toolvendor
- 2.platform.claude.com — pricingvendor
- 3.privacy.claude.com — 10015870 what certifications has anthropic obtained
- 4.embracethered.com — claude computer use c2 the zombais are coming
- 5.techcrunch.com — anthropic acquires vercept ai startup agents computer use fo
- 6.hyperight.com — anthropic acquires vercept to enhance claudes computer use
- 7.zylos.ai — 2026 02 08 computer use gui agents
- 8.anthropic.com — claude sonnet 4 6
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Sema4.ai
Enterprise AI agents for knowledge work, running natively in your Snowflake or cloud account
Sapiom
Agent infrastructure that routes, runs and meters AI agents in production
Itential FlowAI
Governed AI agents for network and infrastructure operations, with deterministic execution and full audit trails
OpenAI Presence
Deploy production-grade AI voice and chat agents with enterprise policies, guardrails and evals