dsh-computer-use-windows

Getting an AI Agent to actually operate desktop applications on Windows has always been a practical headache. Standard controls expose a UI Automation tree, but self-drawn UIs found in WeChat, Electron apps, and other custom frameworks simply do not ship one. dsh-computer-use-windows fills that gap in the DeepSeek Harness ecosystem, giving the Agent two complementary paths: precise element_index hits for standard controls, and a screenshot-plus-ModLens 3x3 grid workflow that progressively zooms into self-drawn regions until the target coordinate is small enough to click.

Core capabilities span window discovery and foreground activation, UI Automation element probing with focus-state tracking, window screenshots, and observation-bound element or coordinate actions. On the safety side, high-risk operations such as typing, sending, deleting, or paying route through the DeepSeek Harness approval service and fail closed when approval infrastructure is missing. The bundled native helper additionally refuses invisible windows, invalid handles, out-of-bounds points, password fields, terminal windows, authentication UIs, and locked desktops, keeping the Agent strictly within safe boundaries.

This plugin is well-suited for developers building Windows 10/11 x64 workbenches where an AI Agent needs to drive desktop applications. It is fully self-contained: the bundled .NET runtime means no separate dependency installs, and the only external requirement is keeping the @liustack/modlens vision provider enabled. If your workflow involves multi-app orchestration, form filling, or interacting with domestic and custom-built applications that lack accessibility APIs, this plugin meaningfully lowers the integration cost.