Qwen Vision MCP

MCP server that adds image analysis capabilities to AI models that don't natively support vision.

Why?

Models like MiMo V2 Pro, MiniMax M2.1, Nemotron 3, and many others are powerful for text but cannot read or analyze images. This MCP server bridges that gap by routing image tasks to Qwen 3.5 via Ollama — giving any model vision capabilities through tool calls.

How It Works

Your model (MiMo V2 Pro, etc.)  →  calls qwen-vision tool  →  Qwen 3.5 analyzes image  →  returns result

The vision model can be:

• Ollama Cloud (qwen3.5:cloud) — fast, inference on Ollama's servers (requires ollama serve as gateway)
• Local Ollama (qwen3.5:9b, qwen3.5:4b) — private, runs on your machine (requires ollama serve)

Prerequisites

• Ollama installed (brew install ollama or curl -fsSL https://ollama.ai/install.sh | sh)
• Ollama authentication — sign in with ollama and authenticate your account
• ollama serve running — required for both cloud and local models (Ollama acts as the gateway to cloud inference)

Tools

Input Formats

• File paths: /path/to/image.png
• URLs: https://example.com/image.jpg
• Base64 data URIs

Setup

1. Pull the model

# Cloud (recommended — no local resources needed)
ollama pull qwen3.5:cloud

# Or local (requires ~6GB RAM for 9B)
ollama pull qwen3.5:9b

2. Add to opencode config (~/.config/opencode/opencode.json)

Option A: Ollama Cloud (recommended)

ollama serve must be running — it acts as the gateway to Ollama's cloud inference.

{
  "qwen-vision": {
    "type": "local",
    "command": ["npx", "-y", "qwen-vision-mcp"],
    "environment": {
      "QWEN_VISION_MODEL": "qwen3.5:cloud",
      "OLLAMA_BASE_URL": "http://localhost:11434",
      "QWEN_VISION_THINK": "false"
    },
    "enabled": true,
    "timeout": 30000
  }
}

Option B: Local Ollama

Requires ollama serve running and ~6GB RAM for qwen3.5:9b.

{
  "qwen-vision": {
    "type": "local",
    "command": ["npx", "-y", "qwen-vision-mcp"],
    "environment": {
      "QWEN_VISION_MODEL": "qwen3.5:9b",
      "OLLAMA_BASE_URL": "http://localhost:11434",
      "QWEN_VISION_THINK": "false"
    },
    "enabled": true,
    "timeout": 30000
  }
}

Other local options: qwen3.5:4b (3GB RAM), qwen3.5:2b (1.5GB RAM).

3. Restart opencode

opencode

The qwen-vision_* tools will be available to any model — even ones without native vision.

Configuration

Development

git clone https://github.com/sahilchouksey/qwen-vision-mcp.git
cd qwen-vision-mcp
bun install
bun run dev

Plugin Mode (Non-Multimodal Models)

For models that cannot call tools or don't support MCP, use the OpenCode Plugin in .opencode/plugin.ts. This plugin intercepts chat messages containing images, analyzes them automatically via Qwen 3.5, and injects the analysis into the conversation — no tool calls needed.

How Plugin Mode Works

User pastes image → Plugin detects image → Qwen 3.5 analyzes it → Analysis injected as text → Model responds using analysis

When to Use Plugin vs MCP

Plugin Setup

1. Copy the plugin file to your opencode config:

cp .opencode/plugin.ts ~/.config/opencode/plugins/qwen-vision.ts

2. Add to ~/.config/opencode/opencode.json:

{
  "plugin": [
    "file:///Users/YOUR_USERNAME/.config/opencode/plugins/qwen-vision.ts"
  ]
}

3. Ensure Ollama is running:

ollama serve

4. Restart opencode. The plugin will:
   • Detect when you paste an image
   • Automatically analyze it with Qwen 3.5
   • Inject the analysis into your message text
   • The model will respond as if it can see the image

Plugin Features

• Automatic image detection — works on paste/upload
• No tool calls required — models don't need tool support
• Transparent injection — analysis appended to user message
• Fallback mode — injects as system message if text part not found
• Built-in tools — vision and vision_text tools available for explicit calls

Plugin Configuration

Same environment variables as MCP mode:

License

MIT