Claude Computer Use and Dispatch: Desktop Control by Phone

Two features, one release

Anthropic's announcement bundles two things under one launch. Computer use lets Claude Cowork and Claude Code fall back to pointing, clicking, typing and reading the screen directly when a task has no purpose-built tool or connector, rather than stopping and asking a person to do that step by hand. [1]

Dispatch keeps a single conversation running across phone and desktop, so a person can assign Claude a task from a phone, close the laptop, and come back later to review or continue the same thread rather than starting a new one on each device. [1][2]

Atlas interpretation: The pairing is deliberate rather than incidental: computer use is what lets a dispatched task actually finish a step that has no API or connector behind it, and Dispatch is what lets a person start or check that work without being at the desk where it runs. [1]

Tools first, the screen as a last resort

Anthropic describes an order of preference: Claude tries an integrated connector, such as one for Slack or Google Calendar, before it tries direct computer control. Screen-based interaction is the fallback for the gap those connectors leave, not the default path. [1]

Anthropic's own framing is qualified rather than promotional: the model can make mistakes, screen-based control is slower than a direct integration, and a complex task can need more than one attempt. Certain sensitive applications are restricted by default, and Anthropic recommends starting with trusted apps and keeping sensitive data out of the loop. The system asks for explicit permission before it acts and includes safeguards Anthropic attributes to prompt-injection defense. [1][2]

Who can use it, and on what

The release is a research preview open to Claude Pro and Max subscribers on macOS and Windows, and it requires the Claude desktop app to be running for computer use to reach the machine. [1]

Engadget's report on the same day describes the launch as reaching macOS, with Windows coverage not part of its account; Anthropic's own announcement states both platforms. The summary above follows the primary source on platform availability. [2][1]

The second attempt at giving Claude hands

Atlas interpretation: Anthropic first shipped a screen-control capability as a raw API tool in October 2024, aimed at developers building their own agents rather than at people using Claude directly. This release is the same underlying idea moved into the products Anthropic ships itself, Cowork and Claude Code, wrapped in permissioning and a tool-first fallback order that the 2024 API left to whoever built on top of it. [1]