dsh-computer-use-macos

In DeepSeek Harness, models can perceive the interface but cannot act on it. dsh-computer-use-macos closes that gap on macOS: it registers a single computer tool that lets a vision model capture the real screen inline in the conversation, then move the mouse, perform left/right/double clicks at pixel coordinates, type text, send key combinations, scroll in any direction, and read the current cursor position. Every OS interaction goes through three macOS-native binaries (screencapture, sips, osascript), meaning zero installed dependencies, zero network calls, and zero compiled helper binaries.

Security and UX engineering go a long way: every command is invoked via Node execFile with an argv array to eliminate shell injection; screenshots are automatically downscaled to stay within common vision-model pixel limits; a machine-wide lockfile prevents two DSH processes from driving the mouse and keyboard concurrently; and optional features include hiding other apps while acting, excluding the host terminal from self-screenshots, posting macOS notifications on start/stop, and adding risk warnings when the frontmost app is a terminal, IDE, Finder, or system-settings pane.

Ideal for macOS-based DSH users and power developers who want a vision model to complete multi-step GUI tasks hands-on. A one-time grant of Accessibility and Screen Recording permissions is all that is needed.