让 Claude Code 中的纯文本 LLM（如 DeepSeek、CodeLlama、Qwen-Coder）通过本地 Ollama 视觉模型获得视觉能力。

支持三种模式：描述剪贴板截图、截取全屏并描述、读取已保存的图片文件。

所有图像处理在本地完成，不上传云端，无需 API key。安装只需两条命令（ollama pull 模型 + pip install），核心脚本 vision 约 150 行。

NoPainNullGain/LocalEyes

Source: https://github.com/NoPainNullGain/LocalEyes

👁️ LocalEyes

Vision for text-only LLMs in Claude Code. Many flagship models like DeepSeek, CodeLlama, Qwen-Coder, and many other models don’t have vision. LocalEyes gives them working eyes via a local Ollama vision model. No cloud, no uploads, no API keys. Private, fast, free.

The problem

DeepSeek is brilliant at reasoning. It’s also completely blind. No screenshots, no UIs, no error dialogs — it can’t see any of it. The usual fix is uploading your images to GPT-4 or Gemini and asking a second model to describe what’s on screen. That breaks the agentic loop. It also means your screenshots are sitting on a cloud provider’s server. Neither is ideal.

LocalEyes gives text-only models local vision. Your screen stays on your machine.

Demo

DeepSeek starts blind — then LocalEyes gives it sight. The model describes the app, understands the UI, and can take its own screenshots during agentic work. All local.

Setup (2 minutes)

Prerequisites

ollama pull qwen2.5vl:7b   # one-time, ~4 GB
pip install Pillow

Install

git clone https://github.com/NoPainNullGain/LocalEyes
cd LocalEyes
python install.py

Verify

# Press Win+Shift+S to screenshot something, then:
python ~/.claude/skills/local-eyes/vision.py   # PowerShell / Git Bash
# Or on cmd.exe:
python %USERPROFILE%\.claude\skills\local-eyes\vision.py

Two modes

You show the model

python vision.py                           # describe the clipboard screenshot
python vision.py "What error is shown?"    # focused question

Model looks itself (agentic mode)

python vision.py screen                              # capture full display
python vision.py screen "Transcribe the terminal"    # focused
python vision.py screenshot.png                      # read a saved file

The model will use screen mode on its own during workflows — after builds, during debugging, when implementing UI from mockups. No prompt needed.

Configuration

Edit ~/.claude/skills/local-eyes/config.json:

{
  "ollama_host": "http://localhost:11434",
  "vision_model": "qwen2.5vl:7b",
  "timeout_seconds": 120,
  "temperature": 0.1
}

Override any setting for a single session:

PowerShell: $env:OLLAMA_VISION_MODEL = "qwen3-vl:latest" Bash/Zsh: export OLLAMA_VISION_MODEL="qwen3-vl:latest"

Environment variables override config.json — useful for trying a new model without editing files.

How it works

┌──────────────────────────┐      ┌─────────────────────────┐
│   Claude Code            │      │  Your machine (Ollama)  │
│                          │      │                         │
│  Your model (text)       │──→──→│  qwen2.5vl:7b (vision)  │
│  DeepSeek / text-only    │←──←──│                         │
│                          │ text │  "The screenshot shows  │
│  "I can see that..."     │      │   a terminal with..."   │
└──────────────────────────┘      └─────────────────────────┘

1. The model runs python vision.py (clipboard, screen, or file)
2. Script captures the image, base64-encodes it
3. Sent to Ollama’s local API at localhost:11434
4. Qwen2.5-VL returns a detailed text description
5. The model reads the description and reasons about it

Nothing leaves your machine.

Platform support

Troubleshooting

Files

LocalEyes/
├── vision.py       # Engine: screen capture, clipboard, Ollama API
├── SKILL.md        # Instructions the model reads to use its eyes
├── config.json     # User settings — edit, don't touch code
├── install.py      # One-command installer
├── LICENSE         # MIT
└── README.md

License

MIT © NoPainNullGain