- Anthropic launched Computer Use for Claude as a research preview on March 23, 2026, allowing the AI to see, navigate, and control a user’s Mac desktop.
- The feature is available to Claude Pro and Max subscribers on macOS, with Windows and Linux support not yet available.
- A companion mobile tool called Dispatch lets users assign tasks to Claude from their smartphone, which Claude then completes on the user’s computer.
- Anthropic uses a permission-first safety model where Claude requests access before interacting with each new application.
What Happened
Anthropic released Computer Use for Claude as a research preview on March 23, 2026. The feature gives Claude the ability to control a user’s desktop by clicking buttons, opening applications, filling out spreadsheets, scrolling through web pages, and completing multi-step workflows without human intervention.
Computer Use is available to Claude Pro and Max subscribers and currently works only on Mac computers. Anthropic first introduced the underlying capability in October 2024 as a developer API, when Claude 3.5 Sonnet became the first frontier AI model to offer computer use in public beta. The March 2026 release brings that capability directly to consumer subscribers, marking a transition from developer tooling to an end-user product.
Why It Matters
The release shifts computer use from a developer-only API tool to a consumer-facing feature. Rather than building custom integrations for every application, Claude can fall back to screen-based control when it lacks a direct connector. When given a task, Claude first checks whether it has appropriate integrations such as Google Calendar or Slack. If not, it navigates the desktop visually, the same way a human would.
For developers specifically, Claude can modify code in integrated development environments, submit pull requests, and run tests autonomously. The approach trades speed for generality. Screen-based operations are significantly slower than direct API integrations, but they work with any application that has a graphical interface.
This fallback architecture means Claude can theoretically interact with any desktop software, including legacy enterprise applications that lack APIs entirely. That breadth of compatibility is the core value proposition, even as the speed tradeoff limits practical use to tasks where automation saves more time than the slower execution costs.
Technical Details
Computer Use works by capturing screenshots of the user’s display and translating natural language instructions into mouse movements, clicks, and keyboard input. When Anthropic first benchmarked the capability in October 2024, Claude 3.5 Sonnet scored 14.9% on the OSWorld evaluation in the screenshot-only category, compared to 7.8% for the next-best AI system at the time.
The feature launched alongside Dispatch, a companion mobile tool that lets users assign tasks to Claude from their smartphone. Claude then executes those tasks on the user’s computer remotely, enabling workflows where a user can delegate work while away from their desk.
Anthropic has implemented several safety measures. The system uses a permission-first approach where Claude requests access before touching any new application, and users can stop it at any time. The company added safeguards against prompt injection attacks and performs automatic vulnerability scanning. Anthropic acknowledged the feature remains early-stage, noting that “Claude can make mistakes, and while we continue to improve our safeguards, threats are constantly evolving.”
Who’s Affected
The research preview is limited to Claude Pro and Max subscribers using macOS. Windows and Linux users cannot access the feature yet. Early adopter companies that previously integrated computer use through the developer API, including Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company, already had access to the underlying capability.
The consumer launch expands the addressable audience significantly, though Anthropic advises against granting Claude access to sensitive data during the preview phase. Organizations evaluating the feature for business use will need to weigh the convenience of general-purpose computer control against the risk of granting an AI system broad desktop access in environments with confidential information.
What’s Next
Anthropic has not announced a timeline for Windows or Linux support, nor has it indicated when Computer Use will exit research preview status. The company also has not disclosed whether the feature will eventually be available on the free tier or lower-cost subscription plans. Performance benchmarks for the consumer release have not been published, so it is unclear how much the capability has improved since the October 2024 OSWorld scores. Details are available in Anthropic’s announcement and the SiliconANGLE coverage.
