Agent.exe

Agent.exe is a proof-of-concept Electron application that enables Anthropic's Claude 3.5 Sonnet model to directly control your local computer. It leverages Claude's computer use API to perform mouse movements, clicks, keyboard input, and screen reading, treating the model as an autonomous agent that can operate your desktop. Built by Kyle Corbitt, the project is explicitly unmaintained — the creator recommends forking and extending it for production use. The tool is designed for developers and researchers who want to experiment with AI-driven desktop automation. Unlike cloud-based automation tools, Agent.exe runs entirely on your machine, giving Claude direct access to your local environment. This allows it to interact with any desktop application — browsers, IDEs, file managers, or custom software — as a human would. Setup is minimal: clone the repo, add your Anthropic API key, and run `npm start`. Key features include direct mouse and keyboard simulation, screen capture for visual context, and a simple stop button to halt the agent. It currently only supports macOS (theoretically cross-platform) and works only on the primary display. The creator notes that Claude performs best with Firefox and may struggle with other browsers. Agent.exe is strictly for experimentation. It has no safety sandboxing — Claude can perform any action a user could, including destructive operations. The project is offered under the MIT license, and the roadmap is essentially nonexistent. For production use, consider forking the code to add guardrails, multi-monitor support, and ongoing maintenance. Alternatives like the official Anthropic demo provide more structure but are heavier to set up.