Anthropic on Monday launched a research preview of Claude’s computer-use capability, enabling the AI assistant to take direct control of a user’s Mac to complete tasks autonomously. The feature, available now for Claude Pro and Max subscribers, positions Claude as a direct competitor to OpenClaw in the race to build AI agents that operate independently on behalf of users.
How It Works
When assigned a task, Claude first checks whether it has a direct integration with the relevant application, such as Slack or Google Calendar. If no connector exists, Claude falls back to controlling the computer like a human would: pointing, clicking, scrolling, and typing through the screen, according to Anthropic’s blog post. It can open files, navigate web browsers, and run development tools without any additional setup.
The capability is built into both Claude Cowork and Claude Code. Anthropic highlighted developer-specific workflows in particular: making changes in an IDE, running tests, and submitting pull requests, all while the user focuses elsewhere.
Dispatch Integration
The computer-use feature pairs with Dispatch, a mobile tool Anthropic released last week that lets users assign tasks to Claude from their phone. Together, the two features create a remote-agent workflow: a user can message Claude from their phone while away, and the agent executes the task on their desktop.
Anthropic’s demo video showed a user running late for a meeting, asking Claude to export a pitch deck as a PDF and attach it to a meeting invite. Claude completed the task without further input.
Safety and Limitations
Anthropic is running this as a research preview, not a production-ready feature, and the company was explicit about its boundaries. “Computer use is still early compared to Claude’s ability to code or interact with text,” Anthropic wrote. “Claude can make mistakes, and while we continue to improve our safeguards, threats are constantly evolving.”
The system uses a permission-first model: Claude requests access before touching any new application, and users can stop it at any time. Anthropic’s system also scans model activations for prompt injection attempts in real time. Some applications are disabled by default, and the company recommends avoiding sensitive data during the preview period.
For now, computer use is limited to macOS. Windows and Linux users will have to wait, as CNET noted. Complex tasks may require multiple attempts, and screen-based operations are significantly slower than direct API integrations.
The OpenClaw Contest
The launch arrives at a crowded moment in the autonomous agent market. OpenClaw, the open-source framework that went viral earlier this year, established the baseline: users message an AI through WhatsApp or Telegram, and the agent executes tasks on their local device. Nvidia debuted NemoClaw last week as an enterprise-grade OpenClaw framework, and OpenAI hired Peter Steinberger, OpenClaw’s creator, in February.
Anthropic’s approach differs from OpenClaw’s open-source model in a key way: it’s a closed, first-party product built into Claude’s subscription tiers. Where OpenClaw connects to multiple AI providers and runs on any platform, Claude’s computer use is Mac-only, Anthropic-only, and gated behind a paid subscription.
The timing also carries political weight. On the same day Anthropic launched this feature, the company faced Judge Rita Lin in a San Francisco federal courtroom over the Pentagon’s Claude blacklist. Shipping an aggressive new agent capability while simultaneously arguing in court that Claude is safe enough for government use creates a conspicuous contrast that competitors and regulators will likely note.