The fully autonomous computer control agent can be controlled remotely from a phone.
What if AI worked for you on your computer while you were away? This is the promise of Anthropic which unveiled on March 23 Claude’s Dispatch mode coupled with computer use. After allowing the remote use of Copilot Cowork, via its mobile application, Claude can now take complete control of your computer.
How does Dispatch mode work?
To use Dispatch, you must first link your Claude account between your two devices. The procedure is simple: open the Claude desktop application on your computer, then log in with the same account on the Claude mobile application. Once the two sessions are linked, the phone becomes a remote control capable of assigning tasks to the computer, even remotely. Concretely, the user sends an instruction from their phone, and Claude executes it on the desktop. The agent first favors direct connectors to services like Slack or Google Calendar. When no integration is available, it switches to direct control of the screen: mouse, keyboard, browser. It can open files, browse the web, launch development tools, all autonomously. Claude, however, requests explicit authorization before accessing a new application.
In terms of availability, Dispatch is accessible on iOS and Android via the Claude mobile application. On PC (Windows), the agent is limited to Claude Cowork without computer use: it can interact with Chrome, but does not take complete control of the machine. True computer use (complete control of the computer) remains a macOS exclusive for the moment. In all cases, you will need to have a Claude Max (100 to 200 dollars per month) or Pro account (20 dollars per month).
For the moment, Dispatch and computer use are mainly suitable for very simple everyday tasks. For example, you can ask him to attach a PDF to a Google Calendar invitation, to sort a folder of photos by date or by theme by renaming them in subfolders, to generate a PowerPoint presentation from a text document, or to fill out a spreadsheet from data scattered in several files.
Safeguards, some limits
Anthropic presents this feature as a research preview, and for good reason: computer use remains slower than direct integration, and complex tasks may require several attempts. The computer must also remain on with the desktop application active for Dispatch to work. On the security side, Anthropic has put in place several safeguards. The system automatically scans the model’s activations (the neural network’s internal signals) to detect prompt injection attempts, and the user can interrupt Claude at any time. Furthermore, certain applications are blocked by default to limit the risks, specifies Anthropic without saying which ones. The company nevertheless recommends not letting the agent manipulate sensitive data and starting with trusted applications.
Since the start of the year, Anthropic has been launching launches at an almost daily pace: context with 1 million tokens, interactive visualizations, code review in Claude Code, Dispatch, and now computer use. The philosophy is always the same: release early, even imperfect, and iterate. An approach that allows the company to get closer, release after release, to the promise of a truly autonomous personal assistant. The promise, in short, of a reliable and secure OpenClaw.
