perrylink/dsh-click
Windows-first native desktop control: screenshots, accessible tree structured reading, click/input/scroll/key press, application startup, approval access control for change operations, rejection of screen changes, process identity verification before and after operations
Gives a DeepSeek Harness agent native desktop control: it looks at the screen and then acts, with every click gated and every action audited. On Windows it drives the UI through UIAutomation and Win32 input, so the agent works with real application controls rather than pixel coordinates. screen_read returns structured text — which means a text-only model can read the screen without vision — while vision-capable models can additionally request screen_shot images. Platforms without a backend fail closed with an explicit reason, so an unsupported machine cannot silently half-work.
Install
dsh plugin --profile web add dsh-clickREADME install: the npm channel (dsh plugin --profile web add dsh-click). The git channel tracks latest main (dsh plugin --profile web add "github:PerryLink/dsh-click#main", prepare builds with production dependencies only) and a tarball channel is documented (pnpm pack, then dsh plugin --profile web add ./dsh-click-<version>.tgz). Uninstall with dsh plugin --profile web remove dsh-click.
Compatibility
Windows first: UIAutomation plus Win32 input through a bundled PowerShell helper. macOS and Linux backends are reserved and fail closed with a clear reason rather than degrading silently. Node ^22.19.0 || >=24.0.0; verified against the dsh-v0.1.5-rc.1 checkout on 2026-09-10. Text-only models are fully supported (screen_read returns structured text); vision models additionally get screen_shot images.
Details
- Repo: perrylink/dsh-click
- Category: Agent Capabilities
- Stars: 0
- Version: 0.3.9 (npm dsh-click; registry repository field → github.com/PerryLink/dsh-click, verified 2026-09-11)
- Last push: 2026-08-18
- First seen: 2026-08-16
Recent updates
README documents the platform policy: Windows first with UIAutomation and Win32 input via a bundled PowerShell helper, while macOS/Linux backends are reserved and fail closed with a clear reason.
FAQ
- How do I install dsh-click?
- Run: dsh plugin --profile web add dsh-click. A git channel (github:PerryLink/dsh-click#main) and a tarball channel are also documented.
- Which platforms are supported?
- Windows first, using UIAutomation and Win32 input via a bundled PowerShell helper. macOS and Linux backends are reserved and fail closed with a clear reason.
- Can a text-only model use it?
- Yes — screen_read returns structured text. Vision models additionally get screen_shot images.