meow-vision: A Vision Plugin for DeepSeek Harness, Letting Models See Images and UI Rendering
meimiaoji-creator/meow-vision
meow-vision is a vision plugin for DeepSeek Harness, enabling text-only models to recognize images and multimodal models to preview Vue component rendering via screenshots for UI iteration.
meow-vision is a vision plugin for DeepSeek Harness that gives text-only models the ability to see images and gives multimodal models the ability to see UI rendering. It solves the problem that text-only models cannot directly process images and that Vue page development lacks closed-loop visual verification. It provides two tools: meow_vision lets text models recognize images via a preset vision model and return text descriptions, and meow_preview renders a local Vue demo, URL, or HTML and returns a screenshot, letting multimodal models see real component rendering. The plugin requires zero DSH source changes, automatically injects system prompts, and offers visual configuration in the settings dialog.
How to Install
dsh plugin --profile web add github:meimiaoji-creator/meow-vision- Category
- Tools & Capabilities
- Platform
- DSH Plugin
- Author
- meimiaoji-creator
- Distribution
- Plugin
meow-vision Key Features
meow-vision Repo Summary
What Does It Do?
meow-vision is a DSH plugin for DeepSeek Harness that gives text-only models the ability to "see" images and gives multimodal models the ability to "see" UI rendering. It solves the problem that text-only models cannot directly process images and that Vue page development lacks closed-loop visual verification, letting you have the model "see" screenshots or real component renderings directly inside the harness. Its core capabilities include two tools: meow_vision, which lets a text model recognize images via a designated vision model and return text descriptions, and meow_preview, which renders a page and returns a screenshot to a multimodal model for iterating UI component code.
Core Features
- meow_vision: when the current model cannot directly view images, call it to have a preset vision model recognize the image and return a text description, suitable for screenshots, UI design drafts, error screenshots, etc.
- meow_preview: renders a local Vue demo, URL, or HTML file and returns a screenshot image, letting a multimodal model "see" real component rendering styles (padding, margin, hover, disabled, etc.).
- Automatic system prompt injection: registers two hints (order=150) into DSH's systemPrompt, telling the model when to use meow_vision and how to write minimal demos.
- Zero DSH source changes: coexists with dsh-meow-skill / meow-file-view / meow-dsh-task without interference.
- Visual configuration: adds a meow-vision configuration tab in the DSH settings dialog for specifying the vision model and preview settings.
How to Use This Plugin?
After enabling meow-vision in DSH, restart DSH to see the meow-vision configuration tab in the settings dialog. First, in the "Vision Model" section, select a multimodal model as the vision model (the candidate list automatically filters models whose inputModalities include image), and save; the configuration is written to vision.json. Then the model will automatically call meow_vision to process images. For meow_preview, write a minimal Vue demo (rendering only one component) under .preview/src/demos/, and have the model call meow_preview(target="demos/button-basic") to get a screenshot. Screenshots are saved in .preview/shots/ and can be reused.
This page is an independent rewrite of the plugin's official README — for authoritative documentation and the latest changes, refer to the source: meimiaoji-creator/meow-vision. The plugin is third-party code that runs on your machine once installed; inclusion does not imply endorsement — please review the source before installing.