Anthropic Grants Claude AI Direct Control Over macOS for Agentic Tasks

Anthropic has released a research preview enabling its Claude chatbot to directly control macOS devices, positioning the artificial intelligence as a remote operator capable of executing complex desktop tasks. This new computer use functionality is immediately available to paying subscribers through both the Claude Cowork agentic productivity tool and the developer-focused Claude Code environment. The deployment escalates the competition among technology firms, including OpenAI and Google, who are also developing functional, action-oriented AI agents that extend beyond standard conversational interfaces.

The feature allows Claude to interact with the operating system by clicking interface elements, inputting text, and navigating installed software. Currently limited to the macOS platform, Anthropic has stated that a Windows version is forthcoming in the coming weeks. The system employs a tiered hierarchy for task execution, prioritizing direct connectors to established services like Slack or Google Calendar. When direct integration is unavailable, the system defaults to controlling the Chrome browser via an extension; the final recourse involves direct screen interaction, which requires continuous screenshot analysis to interpret the desktop's visual state.

This screen-level interaction, which permits scrolling, clicking, and exploration, represents the most flexible but potentially most error-prone layer of the system. The integration extends to Dispatch, allowing users to issue text instructions via a mobile phone for execution on the connected desktop, establishing a remote workflow pipeline. This immediate release, even as a research preview, addresses a broader industry trend where the capacity to deploy working agents is becoming a decisive factor in the competition for enterprise adoption.

Access to this advanced capability is tiered: Claude Pro subscribers pay $17 per month when billed annually or $20 monthly, while the Claude Max subscription costs $100 or $200 per month. Anthropic has acknowledged that this is an early-stage feature that may fail on complex, multi-step workflows. Early user feedback, such as that from John Voorhees of MacStories, noted that while the feature represents progress, it is not yet reliable for intricate processes when the user is away from the desk.

Granting an AI direct control over a live desktop environment raises inherent security and governance concerns outside of a virtual machine sandbox. Anthropic has implemented safeguards, including requiring explicit user permission before accessing new applications and blocking certain sensitive application categories by default, while also scanning for prompt injection attempts. However, the company has candidly stated that these guardrails are not absolute and the system currently lacks audit logs for Cowork activity, which presents hurdles for adoption in highly regulated enterprise settings.

This development places Anthropic in direct contention with other major technology firms focusing on system-level AI integration. OpenAI is reportedly consolidating ChatGPT, the Atlas browser, and the Codex coding assistant into a single desktop application, while Google is expanding its AI assistant ecosystem utilizing Gemini 2.5 Computer Use and continuous screenshot analysis. The collective movement by these firms signals a clear industry consensus that the next major frontier for AI application lies in deep, system-level automation.