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

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:

1. No standard pipe — every IDE wires it differently.

2. 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.

Related MCP server: image-tiler-mcp-server

中文

仓库里有什么

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。

或者直接让你的 agent 装

装完 skill 之后，对你的 Claude Code（或任何支持 MCP 的 agent）说：

"装个 vision-extension。视觉模型用 <豆包 | openai | 通义 | 智谱 | ollama | …>，key 是 <KEY>。"

skill 会按 7 步确定流程把剩下的全做完。你不用写任何 JSON。

为什么做这个

DeepSeek V4 Pro / GLM 5.2 / Kimi K2 / Qwen 3 Max 这类长上下文推理模型推理超强，但看不见图。直接接个视觉 API 拼起来有两个老问题：

1. 没有统一通道 —— 每个 IDE 接法都不一样

2. 视觉模型爱"帮忙" —— GPT-4o / Gemini / 豆包都会条件反射地给方案、提假设、写评价，把推理工作抢走一半，你只想要个描述

vision-extension 一并解决：

• 一个 MCP server，Claude Code / Cursor / Continue / Cline / Roo 通用

• describe-only 契约 —— 视觉模型被系统提示锁成纯扫描器，不给建议、不给方案、不给评价，只做逐字转录和结构化描述

• 一个 Claude Code skill 把安装流程和日常使用规则压成一句话触发

• 协议解耦 —— Anthropic 协议、OpenAI 协议都支持，一个环境变量切换

License

MIT。