Anthropic has launched the most ambitious AI agent yet: Claude can now directly control a user’s Mac, clicking buttons, opening apps, and typing as if a person were at the keyboard. This update, available in research preview for paying subscribers, transforms Claude from a chatbot into a remote digital operator, marking a significant escalation in the competition to build AI that does work—not just talks about it.
The Rise of AI Agents: From Chat to Action
The tech industry is locked in a fierce battle to create AI agents that integrate seamlessly into existing workflows. OpenAI, Google, Nvidia, and numerous startups are racing to develop systems that operate within familiar tools, rather than as standalone applications. The stakes are no longer theoretical; Reuters reported Sunday that OpenAI is courting private equity firms in an “enterprise turf war” with Anthropic, where the ability to ship functional agents is becoming the key differentiator.
This push for action-oriented AI is driven by a simple realization: users want solutions that solve problems, not just offer assistance. The current generation of chatbots can provide information, but true productivity requires AI that can execute tasks autonomously.
How Claude Takes Control: Prioritized Access and Security Trade-offs
The new “computer use” feature works through a layered priority system. Claude first checks for direct integrations (Gmail, Slack, etc.). If those aren’t available, it navigates the Chrome browser via an extension. Only as a last resort does it interact directly with the screen, mimicking human actions.
This hierarchy reflects a crucial trade-off: reliability versus reach. Direct integrations are fastest and most secure, while screen-level interaction is flexible but prone to errors. Anthropic acknowledges this, stating that screen interactions are “slower and more error-prone.”
However, direct screen control introduces security risks. Claude takes screenshots to understand the desktop, meaning it can see everything visible—including sensitive data. While Anthropic has implemented guardrails against actions like stock trading or data input, they aren’t foolproof. The company explicitly warns against using the feature for financial, legal, or medical tasks.
Dispatch: From Mobile Command to Desktop Automation
The strategic move isn’t just computer use itself; it’s the integration with Dispatch. This feature allows users to send text commands from their phones to Claude on their desktop, which then executes those instructions. Scheduled tasks can automate repetitive workflows, making Claude a persistent background worker.
Early tests show mixed results: Dispatch accurately retrieves information but struggles with complex multi-step tasks. MacStories’ John Voorhees found it worked “about half the time,” while GitHub users report bugs like payload limits crashing PDF processing. The feature is rough around the edges, but Anthropic is betting on rapid iteration.
The OpenClaw Factor: Anthropic Joins a Crowded Field
Anthropic’s timing isn’t accidental. The company is entering a market reshaped by the viral rise of OpenClaw, the open-source framework that enables AI to autonomously control computers. OpenClaw proved user demand for this functionality, spawning a thriving ecosystem of derivative tools and forcing major players like Nvidia (with NemoClaw) to compete.
Smaller startups, like Coasty, are also pushing into the space, offering polished interfaces and aggressive automation features. Anthropic is betting its tighter integration, consumer-friendly design, and existing subscriber base can compete with free alternatives.
The Unsolved Security Equation
The core challenge remains: how to reconcile AI autonomy with security. Computer use runs outside virtualized environments, exposing the desktop directly to potential misclicks, prompt injections, or malicious instructions. Anthropic’s safeguards—permission requests, blocked apps, prompt scanning—are not absolute.
The lack of centralized audit trails for Cowork activity further complicates matters, especially for regulated industries. As one user pointed out, distinguishing human from AI actions becomes impossible, raising questions about liability and compliance.
User Reactions: Excitement Mixed with Skepticism
Early user reactions are divided. Some are enthusiastic about the productivity gains, while others express concerns about security risks. Many Windows users are frustrated by the macOS-only preview. Several subscribers report rapid usage quota depletion, highlighting the cost implications of aggressive automation.
Anthropic’s bet is that the benefits of AI control outweigh the risks, and that users will tolerate imperfections in exchange for real-world automation. The race to build the ultimate AI agent is just beginning, and the winner will be the one that can deliver on its promise: an AI that doesn’t just talk—it acts.
