MCP 服务器

mcp-eyes

A drop-in MCP server that pairs long-context reasoning LLMs with vision models in description-only mode, enabling any reasoning model to 'see' images without the vision model giving advice or solutions.

README

vision-extension

Drop-in vision capability pack for text-only reasoning LLMs. One repo containing an MCP server and a Claude Code skill, both engineered around a single contract: the vision model only describes — your reasoning model does the thinking.

English · 中文

English

What's in this repo

Directory	What it is	Who installs it
`mcp-vision-extension/`	The MCP server (Python package `vision_extension`). Pairs any text-only reasoning model with any vision model over OpenAI or Anthropic protocol.	Required. Install once per machine.
`skills-vision-extension/`	A Claude Code skill that knows the install playbook AND the day-to-day collaboration patterns between the reasoning model and the vision model.	Optional but strongly recommended for Claude Code users. Copy into `~/.claude/skills/`.

These two pieces are designed to work together. The MCP server gives your text model vision; the skill teaches your text model how to use that vision well.

Install everything in 2 commands

# 1. The MCP server
pip install "git+https://github.com/loudMore/vision-extension.git#subdirectory=mcp-vision-extension"

# 2. The Claude Code skill (optional)
git clone https://github.com/loudMore/vision-extension.git /tmp/vx
cp -r /tmp/vx/skills-vision-extension/vision-extension ~/.claude/skills/

Then point your MCP client at the new server. Detailed steps + provider presets in mcp-vision-extension/README.md.

Or just tell your agent

If you have Claude Code (or any MCP-aware agent), copy the skill once:

git clone https://github.com/loudMore/vision-extension.git
cp -r vision-extension/skills-vision-extension/vision-extension ~/.claude/skills/

Then say:

"Install vision-extension. Use the <doubao | openai | qwen | gemini | ollama | …> provider. Here's my key: <KEY>."

The skill handles the rest. You don't write any JSON.

Why this exists

Long-context reasoning models (DeepSeek V4 Pro, GLM 5.2, Kimi K2, Qwen 3 Max, …) are extraordinary at code and analysis but cannot see images. Naively bolting on a vision API has two problems:

No standard pipe — every IDE wires it differently.
Vision models love to "help" — GPT-4o, Gemini, Doubao all reflexively produce advice, debugging hypotheses, and design opinions when you only wanted a description. The reasoning work gets fragmented.

vision-extension solves both:

One MCP server, works with Claude Code, Cursor, Continue, Cline, Roo, or anything else that speaks MCP.
Describe-only contract — the vision model is system-prompted into a pure visual scanner. No advice. No fixes. No opinions. Just verbatim transcription and structured description.
One Claude Code skill that turns the install + daily-use rules into a single trigger phrase.
Provider-agnostic — Anthropic protocol, OpenAI-compatible protocol. Switch with one env var.

License

MIT.

中文

仓库里有什么

目录	是什么	谁要装
`mcp-vision-extension/`	MCP server（Python 包 `vision_extension`），把任意纯文本推理模型和任意视觉模型用 OpenAI/Anthropic 协议接到一起	必装，每台机器装一次
`skills-vision-extension/`	Claude Code skill，把安装流程 + 主模型与视觉模型的日常协作规则打包好	强烈推荐，复制到 `~/.claude/skills/` 即可

两块组件协同设计。MCP server 给文本模型装上视觉；skill 教文本模型怎么用好这套视觉。

两条命令搞定

# 1. 装 MCP server
pip install "git+https://github.com/loudMore/vision-extension.git#subdirectory=mcp-vision-extension"

# 2. 装 Claude Code skill（可选）
git clone https://github.com/loudMore/vision-extension.git /tmp/vx
cp -r /tmp/vx/skills-vision-extension/vision-extension ~/.claude/skills/

然后让你的 MCP 客户端配新 server。详细步骤和 12 个 provider 预设见 mcp-vision-extension/README.md。