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

DSH Vision Toolkit allows text-only AI agents in the DeepSeek Harness (DSH) ecosystem to process and understand visual information. It bridges the gap between text-based models and visual tasks by providing a set of specialized tools and a "Skill" (a set of operational playbooks) that guide the agent on how to perceive, analyze, and manipulate images to complete high-value tasks.

How it works

The toolkit combines a vision-capable model (used for understanding) with deterministic local image processing tools (used for action). When a user pastes an image or provides a path, the system can automatically switch the text-only model to a vision-enabled variant.

Instead of generating generic image descriptions, the toolkit uses "focus hints"—passing the user's specific goal to the vision model so it extracts only the evidence relevant to the current task. The agent then uses a bundled vision-skills Skill to decide which tool to call (e.g., OCR, cropping, or pixel comparison) and how to verify the results.

Who it’s for

  • Users of DeepSeek Harness who want a multimodal experience (pasting images directly into chat) with text-only models.
  • Developers and designers who need to automate complex visual workflows, such as converting screenshots or hand-drawn sketches into functional HTML/CSS code, extracting data from long screenshots, or operating a GUI via screenshots.

Highlights

  • Task-Aware Vision: Extracts specific evidence based on the current goal rather than providing generic captions.
  • Comprehensive Toolbox: Includes 10 tools for tasks like vision_glance (general understanding), vision_ground (pixel coordinates), vision_trace (SVG conversion), and vision_pixel_diff (visual regression testing).
  • UI Restoration: Specialized playbooks for rebuilding UIs from screenshots, sketches, or diagrams into structured code.
  • Seamless DSH Integration: Native support for DSH Web and Headless profiles, including automatic model switching upon image paste and native thumbnail previews.

Related

  • Project
    ysr666/dsh-vision-router

    dsh‑vision‑router is a Node‑based plug‑in for DeepSeek Harness that adds a “👁 Vision” toggle, letting a text‑only LLM call pixel‑level vision tools (crop, OCR, grounding, etc.) via configurable cloud or local vision models, with a free OVHcloud fallback and a local‑only privacy mode.

  • Project
    Anionex/agent-vision-toolkit

    A toolkit that gives text-only LLM agents visual capabilities through intent-aware CLI tools and seamless integration proxies, enabling tasks like UI restoration and GUI automation.

  • Project
    Mars-Sea/dsh-commandcode-provider

    dsh‑commandcode‑provider is an unofficial DeepSeek Harness plugin that adds Command Code as an LLM provider. It offers UI integration, multi‑account rotation, plan‑aware model picking, vision support, cost estimates, quota display, web‑search, and optional zero‑data‑retention, working with dsh 0.2.x.

  • Project
    PC2005-cloud/dsh-pet

    dsh-pet is a DeepSeek Harness plugin that adds an animated desktop pet with physics, event‑driven animations, balance display, chat, and an HTTP API, configurable via JSONC and extensible with custom animation packs.

  • Project
    dream-num/dsh-univer-office

    A plugin for DeepSeek Harness that enables an AI agent to create, edit, and manage office documents like spreadsheets, presentations, and reports using natural language.