Introduction¶

DSH emphasizes “everything is a plugin”. When performing desktop automation on Windows, common challenges include screenshots containing irrelevant windows, unstable coordinate control, OCR text position drift, and the inability to verify whether the expected state is reached after a click. Altairpaca/dsh-computer-use-windows addresses these issues by encapsulating window-bound screenshots, OCR text localization, post-click verification, pure OCR mode, and optional visual models into DSH plugins.

What is this¶

This is a Windows Computer Use for DeepSeek Harness project maintained by Altairpaca, released under the MIT License. It provides a DSH plugin wrapper and exposes a set of computer-use capabilities on the Windows side: computer_screenshot, computer_ocr, computer_click_text, computer_mouse / computer_keyboard, computer_window, computer_use_run, computer_vision, and computer_calibrate, while supporting OCR-only mode, an optional vision provider, and local health checks.

The current status is Experimental alpha. It is suitable for development and controlled testing, but the project does not advocate for production-grade unattended desktop automation. The remaining release hurdles are the clean-install / real-DSH validation matrix, which need to be completed on representative interactive Windows configurations.

Core Features¶

The following lists the available capabilities based on verified facts.

computer_screenshot: Window-bound screenshots.
computer_ocr: Robust OCR.
computer_click_text: Verified clicks.
computer_mouse / computer_keyboard: Mouse, keyboard, and other input operations.
computer_window: Window-related operations.
computer_use_run: Entry point for running computer-use.
computer_vision: Optional visual model access.
computer_calibrate: Calibration-related capabilities.
OCR-only mode: A pure OCR path that does not require an external visual model.
Optional vision provider: A configurable visual provider, with an example using openai-compatible.
DSH plugin wrapper: Accessible as a DSH plugin.
local health checks: Local health checks.

Installation and Enablement¶

Official installation commands are not provided in verified materials; therefore, this article does not include installation commands. If you have obtained the repository source code on your target Windows workstation, you can first run a local check:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts/check-health.ps1

This corresponds to the check script call in package.json. Requirements include: Windows 11 recommended; PowerShell 7.4+; Node.js 20+ for the DSH plugin surface; and OCR-related processes require Windows OCR language packs.

Hosted CI performs only deterministic checks; the full scripts/check-health.ps1 is not used in hosted CI because it depends on Windows Runtime/OCR availability and an interactive desktop session.

Typical Usage¶

The following examples are based on verified usage.

Run a local health check on the target Windows workstation:

./scripts/check-health.ps1

Directly call the helper’s screen example:

$env:CU_ARGS = '{"cmd":"screen"}'
& ./helper/cu.ps1

Enable OCR-only mode:

{
  "vision": {
    "enabled": false
  }
}

Configure an Optional vision provider:

{
  "vision": {
    "enabled": true,
    "provider": "openai-compatible",
    "base_url": "https://your-vlm.example.com/v1",
    "api_key_env": "MY_VLM_KEY",
    "model": "your-model"
  }
}

The base_url, api_key_env, and model mentioned above are documentation examples; API keys should be referenced via environment variable names and should not be written into the repository configuration.

Applicable Scenarios and Notes¶

Suitable for development, debugging, and controlled testing on Windows 11, especially scenarios requiring OCR text localization, post-click verification, and screenshots saved to local paths.
Not suitable for direct use as verified production-grade unattended desktop automation; the current status is Experimental alpha.
This software can inject mouse and keyboard inputs and should only be used on applications and systems for which you have permission to operate.
Target-window checks reduce the probability of accidentally operating on irrelevant windows, but they do not make arbitrary desktop automation risk-free.
If vision is enabled, screenshots may be sent to the configured endpoint; if screenshots must remain local, use the OCR-only mode.
Model/API credentials should be read from environment variables or host credential stores and not committed to configuration.
It executes operations via local scripts and Windows input capabilities; therefore, you should confirm the permissions of the current DSH process and check the source code, MIT license, and dependencies before running.

Conclusion¶

Altairpaca/dsh-computer-use-windows provides a set of Windows desktop operation tools for the DeepSeek Harness: screenshots, OCR, clicking, mouse and keyboard, window management, batch running, optional visual models, and local health checks. It is still in the experimental stage and is suitable for verifying workflows in controlled environments.

GitHub: https://github.com/Altairpaca/dsh-computer-use-windows. The directory page URL is not listed in verified materials; if you need to find it via a community directory, please refer to the actual page of the independent directory site.

DSH Computer Use for Windows: Windows Desktop Operation Tools for DeepSeek Harness