Introduction¶
In the DeepSeek Harness (DSH) plugin ecosystem, implementing desktop automation is a common requirement. Existing MCP protocol approaches or third-party drivers may increase system overhead or depend on additional middleware. The dsh-computer-use plugin aims to provide a lighter, more native Windows desktop control method, invoking it directly through DSH’s Cordis tool registry without an MCP proxy layer.
What Is This¶
dsh-computer-use is a plugin for DeepSeek Harness that provides a native Windows computer-use toolset. It directly uses DSH’s Cordis tool registry and does not depend on the MCP protocol. The plugin supports only the Windows platform, implements input and screenshot functionality by calling user32.dll and System.Drawing, and does not install global hooks or run background processes.
Core Features¶
The plugin provides a standard set of computer control tools, supporting the following operations:
computer_screenshot: Captures the virtual desktop and returns a DSH image attachment.computer_click,computer_drag,computer_scroll: Control the pointer using virtual desktop coordinates.computer_type,computer_keypress: Send Unicode text or keyboard shortcuts to the current control.computer_wait: Allows the application to stabilize state before the next observation.computer_observe: Captures the desktop and sends it to a configured vision model for inference when the current DSH model cannot directly view images.
All computer_* tools request approval by default through DSH’s native approval service, ensuring confirmation before sending input.
Installation and Activation¶
Installing the plugin requires specifying a DSH configuration profile, such as web, tui, or headless.
dsh plugin --profile web add @crazy_th/dsh-computer-use
After installation, you must restart DSH to load the new plugin. You can verify whether the configuration has taken effect with the following command:
dsh --profile web --dump-config
In the output configuration, you should find the two nodes computer-use-settings and computer-use. If they are not found, rebuild or reinstall the plugin.
Vision Model Configuration¶
The plugin supports sending screenshots to an external vision model for analysis.
- Open Settings > Plugins > Plugin configuration.
- Expand the Computer Use Vision configuration card.
- Fill in the following fields:
| Field | Description |
|---|---|
| Enabled | Enabled only after the endpoint, API key, and model have been filled in. |
| Base URL | An OpenAI-compatible API URL in HTTPS format, for example https://api.example.com/v1. |
| API Key | Stored in DSH’s settings as a ciphertext field, used for local storage only. Leave it blank when updating to retain the existing key. |
| Model | The identifier of the vision model. |
| Max Tokens | Integer, range 64 to 8192, default 1000. |
Note: The computer_observe tool only sends the current screenshot to the endpoint configured above. Make sure you understand where the data is processed.
Use Cases and Notes¶
- Platform Limitations: The plugin is a Windows-only automation driver and depends on
user32.dllandSystem.Drawing. - Security: DSH’s native approval service is enabled by default. If you encounter a
user rejected toolerror, it may be becauseapproval: neverwas used (for example, thedanger-full-accesspreset), in which case the tool is automatically rejected and misleading text is displayed. It is recommended to enter/permission workspace-writein the conversation to restore prompting. - Configuration Activation: The vision configuration registered by the plugin is read through DSH’s browser configuration API to ensure security.
- License: The plugin uses the MIT license.
Brief Ending¶
dsh-computer-use provides native support for desktop automation in Windows environments with DSH. By integrating directly into the Cordis tool registry and utilizing DSH’s native approval mechanism, it simplifies environment setup and is suitable for developers who need fine-grained control over desktop operations.
- GitHub repository: https://github.com/ThreeBody6666/dsh-computer-use