dsh-computer-use-win

The DeepSeek Harness computer-use ecosystem was macOS-first, leaving Windows users to hand-roll screenshot scripts and key-press utilities. dsh-computer-use-win closes that gap with a single, self-contained bundle: full-screen or per-window capture, top-level window enumeration, Windows UI Automation element discovery, OCR-based text localization, structured visual observation, pointer and keyboard control, and direct Settings launch via the ms-settings URI. Every capability runs inside a bounded PowerShell child process, so there is no native binary to install or keep in sync.

For text-only models the plugin provides a complete see-then-act pipeline: turn a custom-rendered CEF or Chromium window into a structured UI map with computer_observe_window, hit the right control with computer_locate_text or computer_click_text, and finish the job with computer_type_text (clipboard-first for VDI reliability) and computer_press_key. Vision-capable routes such as deepseek-v4-flash-vision-exp skip the OCR intermediate entirely and inspect the raw screenshot pixels, with DSH normalizing each capture to the route budget and falling back to inline upload when the Files API is unavailable.

Built for developers and teams who need to automate Windows desktop workflows, orchestrate RPA-style tasks, or give a DeepSeek conversational model a working set of eyes and hands. Input-injecting tools are approval-gated by default; under dsh Full access preset they execute seamlessly, balancing flexibility with safety.