Text models in DeepSeek Harness usually cannot directly understand image content. To give these models vision capabilities and complete tasks such as image Q&A, long screenshot parsing, or UI recovery, an external tool is needed. mengruoa/dsh-vision-toolkit is a plugin designed for this purpose. It provides vision capabilities to text models in DeepSeek Harness through a set of tools and Skills.

Plugin Positioning

This is a toolkit that provides vision capabilities to text models. The plugin includes vision tools and Skills, and supports image Q&A, long screenshot OCR, UI recovery, and GUI visual tasks. The maintainer is mengruoa, and the license is MIT.

Core Features

The plugin provides the following capabilities:

  • Image Q&A: Ask questions directly about image content.
  • Long Screenshot OCR: Process long screenshots containing large amounts of text.
  • UI Recovery: Rebuild screenshots into editable code or pages.
  • GUI Visual Tasks: Operate graphical user interfaces through screenshots.
  • Vision Skills and Tools: Built-in Skills guide Agents to perform visual tasks and provide multiple vision tools.
  • Multi-Provider Support: Supports configuring multiple vision providers and has an automatic fallback mechanism.

Installation and Enabling

Install the plugin via the command line. The installation command is as follows:

dsh plugin --profile web add @mengruo/dsh-vision-toolkit

You can also install it into the Headless Profile:

dsh plugin --profile headless add @mengruo/dsh-vision-toolkit

If using DSH Desktop, because there are known issues with the built-in installation, it is recommended to open DSH Terminal and run the command:

dsh plugin --profile desktop add @mengruo/dsh-vision-toolkit

After installation, restart the Web Profile, go to Settings → Vision Toolkit to configure the vision provider, and run Test vision model to confirm the connection.

Typical Usage

After installation and configuration, you can perform the following operations:

  • Paste an image directly into the conversation and ask questions.
  • Rebuild a screenshot into HTML code.
  • Convert sketches or design mockups into usable frontend pages.
  • Restore UI layouts from screenshots.
  • Operate a GUI through screenshots.

Applicable Scenarios and Notes

This plugin is suitable for scenarios where text models need vision capabilities. Note that the plugin runs with the permissions of the current dsh process, and the source code and license should be reviewed before installation. The built-in plugin marketplace of DSH Desktop has known installation issues; terminal commands are a more reliable installation path.

Conclusion

This plugin adds vision processing capabilities to DeepSeek Harness, enabling it to handle more complex visual tasks. Project address: GitHub.