GitHub

Agent.exe lets Claude 3.5 Sonnet control a local computer using new computer-use APIs. Intended for testing Claude’s capabilities, the tool executes actions directly on the computer, including booking tasks or simple commands. Users can observe the app’s performance and use a "stop" button to intervene if the model faces problems. A semi-automatic mode requiring user confirmation for each action was initially planned, but its slower execution speed rendered it unnecessary, rendering Agent.exe effective for hands-off control testing.