TL;DR

• Anthropic has enabled Claude to directly control a user’s computer — clicking, scrolling, typing, and navigating applications — when it lacks purpose-built tool integrations
• A companion feature called Dispatch allows users to assign tasks from their phone and have Claude complete them on a desktop machine, even when the user is away
• The features are available as a research preview on macOS only, limited to Claude Pro and Max subscribers

From chatbot to desktop operator

Anthropic has released what amounts to a fundamental expansion of what its AI assistant can do. Claude can now operate a user’s computer directly — controlling the mouse, keyboard, browser, and screen to complete tasks when no dedicated integration exists. The feature works within both Claude Cowork and Claude Code, Anthropic’s collaborative and developer-focused interfaces.

The system follows a hierarchy of precision. Claude first attempts to use purpose-built connectors to services like Slack or Google Calendar. When no connector is available, it falls back to direct screen interaction, navigating applications the same way a human would. Anthropic states that explicit user permission is required before Claude accesses any new application, and users can halt the process at any point.

Dispatch ties it together

The computer use capability is paired with Dispatch, a feature Anthropic released last week for Claude Cowork and has now extended to Claude Code. Dispatch creates a persistent conversation thread across devices, allowing users to assign a task from a mobile phone and find the completed work waiting on their desktop.

The combination creates a workflow where Claude can operate autonomously on a user’s machine while they are elsewhere. Anthropic’s own examples include preparing a morning briefing while the user commutes, running tests and creating pull requests in a development environment, and managing file organisation across applications.

Safeguards and limitations

Anthropic has built in several restrictions. The system includes automatic scanning for prompt injection attacks — a recognised risk when AI models interact with arbitrary screen content. Certain applications are restricted by default, and the company acknowledges that computer use remains an early-stage capability prone to errors.

The research preview is currently limited to macOS, requiring the Claude desktop application with the feature manually enabled in settings.

Where this fits in the AI agent race

Computer use positions Anthropic alongside other firms building AI systems that go beyond conversation into direct action. Google’s Project Mariner and OpenAI’s Operator have explored similar territory, though with different approaches to scope and user control. Microsoft’s Copilot has pursued a more constrained model, operating within its own application ecosystem rather than across the full desktop.

For UK businesses evaluating AI assistants for productivity, the release raises practical questions about security, data governance, and acceptable use. An AI system with the ability to read screens, access browsers, and operate applications has obvious utility — but also introduces risks that most enterprise IT policies were not written to address.