Anionex/dsh-vision-toolkit

[dsh]为纯文本模型设计更强大的视觉工具箱：一行安装使用、粘贴图片直接识别、多张图片问答、截图到前端UI 还原等｜DeepSeek Harness-native integration for agent-vision-toolkit: image Q&A, long-screenshot OCR, UI restoration, grounding, pixel diff, Artifacts, and Web UI.

What it solves

This toolkit allows text-only AI agents (specifically those running in DeepSeek Harness) to "see" and interact with visual data. It solves the problem of text-only models being unable to process screenshots, missing critical visual details in generic descriptions, and lacking the ability to to perform measurable UI verification or asset extraction.

How it works

It functions as a native plugin for DeepSeek Harness (DSH) that integrates a vision-capable model (like the built-in Gemma 4 service) with a suite of local image processing tools. When a user pastes an image, the plugin switches the agent to a "Vision Toolkit" variant that can request visual evidence, coordinates, or local processing tasks. The agent then chooses between using a remote vision model for understanding and local tools for deterministic tasks like cropping or pixel diffing.

Who it’s for

Developers and AI agent users who use DeepSeek Harness and need their agents to perform visual tasks such as debugging screenshots, recreating UI from sketches, extracting assets from images, or verifying that a rebuilt page matches a reference image exactly.

Highlights

  • Free entry point: Includes a built-in Gemma 4 vision service that requires no API key.
  • Task-focused vision: Agents send the specific reason for inspection to avoid generic, irrelevant captions.
  • Measurable UI restoration: Provides pixel-level difference percentages and heatmaps to verify UI implementation against a reference.
  • Local processing: Operations like SVG tracing, cropping, and color analysis run locally to save API costs.
  • Comprehensive toolbox: Includes 10 specialized tools for grounding, OCR, foreground extraction, and HTML screenshot rendering.

Related

  • Project
    Anionex/agent-vision-toolkit

    A toolkit that gives text-only LLM agents visual capabilities through task-aware vision tools and seamless proxy integrations for image Q&A and GUI automation.

  • Project
    liustack/modlens

    A vision plugin that gives text-only AI models the ability to read images pasted directly into chat by bridging them to external vision engines.

  • Project
    fufankeji/deepseek-harness-studio

    A zero-code desktop application that enhances DeepSeek Harness with a visual plugin marketplace, agent presets, and image understanding capabilities.

  • Project
    ysr666/dsh-vision-router

    A DeepSeek Harness plugin that routes raw images to vision models (with a free OVH fallback) so text‑only agents can call vision tools (grounding, cropping, OCR, pixel‑diff, etc.) without losing pixel fidelity. One‑command npm install, no Python, automatic provider fallback, caching, and eleven ready‑to‑use tools.

  • Project
    Devin-AXIS/deepseek-design

    A visual design system for DeepSeek Harness that enables AI to generate and edit real, editable design files for websites, presentations, and videos.