Skip to content
  • Home
  • Plugin manual
  • image-vision

wangyang10/image-vision

80Last commit Aug 14, 2026

image-vision DSH plugin

This plugin adds vision capabilities to DeepSeek Harness (DSH), enabling pure-text models like DeepSeek to process images. It automatically calls OpenAI-compatible vision APIs (OpenRouter, SiliconFlow, Zhipu, Kimi, Tongyi, local Ollama, etc.) for image description, question answering, and OCR. It also supports multi-image comparison and structured JSON output, with built-in preprocessing (HEIC/AVIF conversion, resizing, compression) on macOS.

visionskills-workflows

How to install the image-vision DSH plugin

dsh plugin --profile web add dsh-image-vision

Copying does not run this command. Review the repository and version before installing the image-vision DSH plugin.

image-vision DSH plugin data source

image-vision DSH plugin snapshot date: Aug 16, 2026

discovered

What the image-vision DSH plugin can do

  • Automatically invokes a vision API when the model lacks visual ability
  • Supports local images, http(s) URLs, and data URLs
  • Preprocesses images on macOS: HEIC/HEIF/AVIF to JPEG, TIFF/BMP to PNG, auto-scaling long edge to 2048px, auto-compression over 10MB
  • Multi-image comparison and Q&A
  • OCR text extraction with optional --json structured output
  • Configurable via environment variables or a .env file; DSH plugin also supports credential storage

Where the image-vision DSH plugin fits

  • Ask a pure-text model to describe the content of a photo
  • Extract and read text from screenshots or documents (OCR)
  • Compare two images and highlight differences
  • Include images in a conversation for context-aware Q&A
  • Automatically process images in a workflow without leaving the chat

Who the image-vision DSH plugin is for

  • DSH users who want to add vision capabilities to their text-only models
  • Developers building AI assistants that need image understanding without a multimodal model

image-vision DSH plugin limitations

  • Requires a third-party vision API key (VISION_API_KEY) and network access at runtime
  • Image preprocessing (HEIC/AVIF conversion, resizing) currently only works on macOS (via sips)
  • The DSH plugin version does not depend on Python, but the standard skill version requires Python 3
  • Image size limits: maxEdge=2048px, maxBytes=10485760 (10MB) by default, configurable

image-vision DSH plugin: from the repository README

Quoted from the wangyang10/image-vision README, the upstream source of the image-vision DSH plugin. Copyright remains with the original authors.

让没有视觉能力的模型（如 DeepSeek 等纯文本模型）也能"看图"：自动调用 OpenAI 兼容的识图模型 API（OpenRouter、SiliconFlow、智谱、Kimi、通义千问、 本地 Ollama 等），完成图片描述、问答、OCR 等任务。 **一个仓库，多个 Agent 宿主**：同一套技能逻辑分别以标准 skill 和宿主插件的形式 提供给 Codex、Claude Code、DeepSeek Harness（DSH）等使用。 ## 支持的 Agent | Agent | 接入物 | 推荐安装方式 | 调用方式 | 运行时依赖 | |---|---|---|---|---| | **Codex** | 标准 skill（`skills/image-vision/`） | 自动安装 / 一键脚本 `TARGET=codex` | 发图提问自动触发；或手动跑 python 脚本 | python3 + sips（macOS） | | **Claude Code** | 标准 skill | 自动安装 / 一键脚本 `TARGET=claude` | 同上 | python3 + sips（macOS） | | **DeepSeek Harness (DSH)** | 插件（`dsh-image-vision/`，推荐） | 一键脚本 / 手动 / 把仓库地址发给 DSH 自动装 | 模型直接调 `vision_query` 工具；输入框 `/image-vision` | 无（纯 Node 内置模块） | | **DeepSeek Harness (DSH)** | 标准 skill（轻量方案） | 一键脚本 `TARGET=dsh` | 发图提问触发技能，模型跑 python 脚本 | python3 + sips（macOS） | | 其他支持 SKILL.md 的宿主 | 标准 skill | 复制到对应技能目录 | 依宿主而定 | python3 | ### 仓库结构 ``` image-vision/ ├── skills/image-vision/ # 标准 skill：Codex / Claude Code / 其他 SKILL.md 宿主 / DSH 轻量接入 │ ├── SKILL.md # 技能说明（各宿主通用） │ ├── scripts/vision_query.py │ ├── agents/openai.yaml │ └── references/providers.md ├── dsh-image-vision/ # DSH 插件：深度接入（Cordis 插件，独立可 publish） │ ├── lib/*.js #

Read the full READMERepository license: MIT

image-vision DSH plugin questions

How do I configure the API key for the Image Vision plugin?

Create a .env file at ~/.dsh/image-vision.env (or ~/.codex/image-vision.env) with VISION_API_KEY=your_key. You can also set the environment variable directly. The plugin also supports storing the key in DSH's credential system (setting page or ~/.dsh/.credentials.yaml).

Which vision API providers are supported?

Any OpenAI-compatible vision API: OpenRouter, SiliconFlow, Zhipu, Kimi, Tongyi (Qwen), and local Ollama. You specify the base URL and model via VISION_API_BASE and VISION_MODEL. See the references/providers.md file for examples.

Can I use the plugin without Python installed?

Yes, the DSH plugin version (dsh-image-vision) is a pure Node.js ESM module with zero Python dependencies. Only the standard skill version (for other hosts) requires Python 3.

What image formats are supported?

Local images: HEIC/HEIF/AVIF/TIFF/BMP/JPEG/PNG (macOS preprocesses the first four). URLs and data URLs: any format supported by the remote vision API. The plugin also handles large images by resizing and compressing them.

Does the plugin work on Windows or Linux?

The plugin itself runs on any OS that supports Node.js (DSH). However, the automatic image preprocessing (HEIC/AVIF conversion, resizing) currently only works on macOS because it uses the built-in sips tool. On other platforms, raw images are sent as-is; you may need to manually convert them first.