jeffi369
SuperMate-Harness-System
Give DeepSeek Eyes · 给 DeepSeek 装眼睛 — a DeepSeek Harness (DSH) Skill: local vision models or vision APIs let DeepSeek read images and graphic files
- Stars
- 1
- Language
- TypeScript
- Created
- Jan 22, 2026
- Updated
- Aug 17, 2026
Introduction
🧩 SuperMate Harness System
"Everything is a plugin" — a true agent assembled from the DeepSeek Harness (DSH) plugin architecture.
Cloud inspiration · Local controllability · Unified local/cloud tools — 天人合一（heaven–human unity: cloud inspiration + local control + unified tools）.
中文版：README-cn.md · 中文说明
DSH Ecosystem · DeepSeek Harness · topic dsh-plugin · Agent Skills
License · Apache-2.0（含专利授权条款）· 双许可边界见 双许可说明（核心闭源商业层：调度/记忆治理/沙箱/评测）
🧩 What is SuperMate Harness System
A self-contained agent system built on DSH's "everything is a plugin" philosophy: Skills (the how-to layer) + Plugins (the can-do layer) freely composed and replaced. Downloading this repository gives you both official deepseek-ai skills/plugins and SuperMate's original ones — no extra fetching required.
SuperMate Harness System
├── skills/ → Skill layer (task-level: how-to)
│ ├── Deepseek-eyes/ 👁️ vision skill (Ollama)
│ ├── MiniMax h3-video-producer/
│ ├── Supermate/
│ └── DSH Official/ official skills (deepseek-ai · 13 · Apache-2.0 fork copy)
└── plugins/ → Plugin layer (system-level: can-do)
├── Supermate/ in-house plugins (planned; core enhancements = closed-source commercial)
├── DSH Official/ official plugins (deepseek-ai · 49 families · fork copy)
│ ├── packages/ official plugin source (ships with the repo)
│ ├── README.md · NOTICE.md
└── README.md plugin catalog (model / memory / schedule / tools / sandbox / eval / UI)
✨ Original Skills
| Skill | What it does |
|---|---|
| Deepseek-eyes | 👁️ Give text models eyes — local vision (Ollama) / OpenAI-compatible vision API → structured text |
Official content ships with this repo — 13 official skills & 49 official plugin families (deepseek-ai, Apache-2.0, NOTICE-attributed): skills/DSH Official · plugins/DSH Official.
👁️ Featured Skill · Give DeepSeek Eyes
Image → vision model converts it to text → the text model "sees"
deepseek-eyes gives DeepSeek (or any text model) a pair of eyes inside DeepSeek Harness — screenshots, photos, charts, design drafts, illustrations, character sheets, AI-generated images all become structured text the model can reason about.
⚙️ How it works
The Skill calls a local vision model (Ollama) or an OpenAI-compatible vision API to convert images and graphic files — screenshots, photos, charts, design drafts, illustrations, character sheets, AI-generated images — into structured text. DeepSeek then reads that text, which is exactly how it gains the ability to "read" images and graphic files.
Image / graphic file → Skill → local vision model or vision API → structured text → DeepSeek reads & reasons
✨ The Skill
| Skill | What it does | Manuals |
|---|---|---|
| deepseek-eyes | Image analysis for text AI models: image → structured description + retrieval tags. Local Ollama or OpenAI-compatible vision API. | English · 中文 |
Architecture
Measured performance (RTX 5080 16GB)
🔗 DeepSeek Harness Ecosystem
This repository is part of the DeepSeek Harness plugin ecosystem — find it under the dsh-plugin topic, alongside DeepSeek Harness itself. Drop a skill folder into ~/.dsh/skills/, and the DSH agent automatically "opens its eyes" whenever it meets an image, injecting the image description into the text model's context.
🚀 Install into DSH
Put the skill folder into DSH's skill directory:
~/.dsh/skills/deepseek-eyes/
Restart the DSH session — the agent automatically "opens its eyes" when it meets an image, and injects the image description into the text model's context.
✅ Prerequisites (for deepseek-eyes)
- Local vision model: Ollama +
ollama pull llava:13b(or any vision model) - or OpenAI-compatible vision API + API key
SuperMate Harness System · Built on the DeepSeek Harness (DSH) plugin ecosystem — topic dsh-plugin