dsh-computer-use

In DeepSeek Harness, letting a model directly interact with the Windows desktop has lacked a lightweight, native option. dsh-computer-use fills that gap: it plugs straight into the DSH Cordis tool registry with no MCP middle layer, giving the model a ready-to-use set of desktop-control tools.

The capability set covers the full loop of everyday desktop interaction—capturing the virtual desktop as an image attachment, performing bounded-coord clicks, drags, and scrolls, sending Unicode text and supported keyboard shortcuts to the active control, waiting for an application to settle, and, when the current model cannot inspect images, forwarding the screenshot to a user-configured vision endpoint for analysis. Under the hood it relies on user32.dll for input and System.Drawing for capture; it installs no global hook and runs no background process.

Security is a first-class design concern. Every computer_* call passes through DSH's native approval service, and illegal inputs such as empty text, unsupported keys, excessive durations, or out-of-bounds coordinates are rejected before anything is sent. Two independent gates reinforce the safeguard: a master On/Off pill on the composer side that denies all tools with an explicit reason when off, and an approval gate that adopts a fail-closed, auto-reject strategy when approval is set to never—eliminating the risk of unintended actions.

It is built for Windows users of DSH who want their model to genuinely perceive and operate the desktop. No screen-recording software or virtualisation layer is required; a single DSH restart after installation is all it takes.