Anthropic’s Claude Gets Computer Use Capabilities in Preview

Why It Matters

By giving an AI direct control of a computer, Anthropic moves closer to autonomous agents that can execute complex workflows, potentially reshaping enterprise automation. The preview also intensifies the competitive race with OpenAI and Google to deliver safe, actionable AI assistants.

Key Takeaways

• Claude can click, scroll, navigate on Mac computers
• Feature available as research preview for Claude Pro and Max
• Claude requests permission before accessing new applications
• Supports developers: edit code, submit pull requests, run tests
• Limited to macOS; Windows and Linux users excluded

Pulse Analysis

Anthropic’s latest "computer use" preview marks a significant step toward AI‑driven automation beyond text generation. By integrating Claude with the Dispatch mobile interface, users can assign tasks from a smartphone and watch the model manipulate a Mac’s graphical interface as if a human were at the keyboard. The system first checks for native integrations—such as Google Calendar or Slack—and only resorts to screen‑based control when necessary, adhering to a permission‑first safety model that prompts users before any new application is accessed. This approach balances convenience with a layer of oversight, though it currently confines the feature to macOS, leaving Windows and Linux users on the sidelines.

For developers, the capability promises tangible productivity gains. Claude can open integrated development environments, modify code, submit pull requests, and run test suites without manual intervention, effectively acting as an autonomous coding assistant. This mirrors OpenAI’s recent plugin ecosystem, where AI models invoke external APIs, but Claude’s visual interaction sidesteps the need for bespoke API endpoints, potentially accelerating adoption in legacy environments. As enterprises seek to streamline repetitive development tasks, such AI‑powered agents could reduce bottlenecks, free up engineering talent for higher‑value work, and accelerate release cycles.

Security remains the chief concern as AI gains direct system access. Anthropic warns that granting Claude "keys" to a computer could expose Macs to prompt‑injection attacks or vulnerabilities in the underlying OpenClaw "Claws" ecosystem. While the company implements automated scans and safeguards, the research preview advises users to avoid sensitive data exposure. The broader AI community watches closely, as the race to safe, autonomous agents intensifies, with OpenAI and Google also pursuing similar capabilities. Success will hinge on robust risk mitigation, user trust, and the ability to expand beyond macOS while maintaining stringent security controls.