Anthropic has officially launched a revolutionary “Computer Use” feature for Claude, marking a major leap from a chatbot to an autonomous agent. The capability, released in research preview on March 23, 2026, allows Claude to see, navigate, and control your computer exactly like a human would—moving the cursor, clicking buttons, and typing text across any application.
This release follows the debut of Claude Sonnet 4.6 in February, which significantly improved the model’s “pixel-perfect” reasoning and visual understanding required for these complex tasks.
1. How It Works: The “Dispatch” System
Anthropic has introduced a unique mobile-to-desktop workflow called Dispatch, allowing you to delegate tasks from your phone to your computer.
- Visual Perception: Claude takes high-frequency screenshots of your desktop to “understand” the UI, identifying buttons, text fields, and menu hierarchies.
- Universal Literacy: Unlike traditional integrations that need specific APIs, Claude uses “universal computer literacy.” It can interact with any software, including legacy internal dashboards or specialized dev tools that don’t have official “connectors.”
- Decision Loop: For every task, Claude first checks for Connectors (like Google Drive or Slack). If none exist, it moves to the Browser. If the task still requires desktop action, it initiates Screen Interaction to manually click and type.
2. Key Use Cases
During the research preview, early users have deployed Claude for multi-step workflows that were previously impossible for AI:
| Capability | Example Workflow |
| Data Consolidation | Opening a PDF, a spreadsheet, and a CRM to cross-reference data and update records. |
| UX Testing | Opening a phone simulator, interacting with a new app build, and identifying navigation bugs. |
| Competitive Analysis | Browsing multiple competitor sites, taking screenshots of pricing, and compiling a structured report in Excel. |
| DevOps | Launching a terminal, running tests, and fixing code directly in the IDE based on error logs. |
3. Availability and Limits
The feature is currently rolling out with a “safety-first” staggered approach:
- Platform: Currently limited to macOS via the Claude Desktop app. Windows support is reportedly “coming soon.”
- Subscription: Available for Claude Pro, Claude Max, and Team plan subscribers.
- Research Preview: Anthropic warns that the feature is still “early” and more error-prone than Claude’s text or coding skills. It can consume a high volume of tokens due to constant screenshot processing.
4. Safety & “Red Lines”
Given the power of an AI that can control a mouse, Anthropic has implemented strict guardrails:
- Explicit Permission: Claude must ask for permission before accessing a new application. Users can approve a “session” but are advised to monitor the screen.
- Sensitive Blocks: By default, Claude is restricted from accessing certain high-risk categories, such as financial trading platforms or government ID portals.
- Privacy Monitoring: The system does not “train” on your screenshots by default, but it does log them for safety moderation (retained for 30 days unless otherwise specified).
5. The Competitive Landscape
The launch is widely seen as a counter-move to the rising popularity of OpenClaw, an open-source framework for autonomous agents. While OpenClaw is highly customizable, Anthropic is betting that its integrated, “zero-setup” approach within the Claude ecosystem will win over professional users who want “AI that actually does things.”
Pro-Tip: Anthropic recommends keeping your screen resolution at 1024×768 (XGA) for the best accuracy during the preview, as higher resolutions can sometimes cause “click-drift” in the current model.
