Vision MCP

An STDIO-based MCP Server that provides unified image analysis capabilities for LLMs lacking visual abilities (or with expensive vision models). By switching Providers (via environment variables), you can use multimodal models from different platforms/vendors.

Supported Models / Providers

Select a provider via VISION_MODEL_TYPE:

| type | Provider | Default VISION_API_BASE_URL | Default VISION_MODEL_NAME | Notes | |------|----------|-------------------------------|----------------------------|-------| | glm | Zhipu GLM-4.6V | https://open.bigmodel.cn/api/paas/v4 | glm-4.6v | GLM-4.6V Zhipu vision model | | siliconflow | SiliconFlow (OpenAI compatible) | https://api.siliconflow.cn/v1 | Qwen/Qwen2-VL-72B-Instruct | Rich vision model selection | | modelscope | ModelScope API-Inference (OpenAI compatible) | https://api-inference.modelscope.cn/v1 | ZhipuAI/GLM-4.6V | Requires real-name verification/Aliyun binding, subject to quotas | | openai | OpenAI | https://api.openai.com/v1 | gpt-4o | Chat Completions compatible | | claude | Anthropic Claude (Messages API) | https://api.anthropic.com | claude-3-5-sonnet-20241022 | baseUrl should not include /v1 | | gemini | Google Gemini (generateContent API) | https://generativelanguage.googleapis.com | gemini-2.0-flash-exp | Official entry; proxies/gateways can override via VISION_API_BASE_URL | | coding-plan | Coding Plan (Aliyun) | https://coding.dashscope.aliyuncs.com/v1 | kimi-k2.5 | Kimi K2.5 vision model via Coding Plan API |

Get API Key / Token (from respective platform consoles):

• GLM (Zhipu): https://open.bigmodel.cn/
• SiliconFlow: https://cloud.siliconflow.cn/
• ModelScope: https://modelscope.cn/my/myaccesstoken
• OpenAI: https://platform.openai.com/
• Claude (Anthropic): https://console.anthropic.com/
• Gemini (Google AI): https://ai.google.dev/
• Coding Plan: Aliyun Coding Plan subscription

Features

• One-click provider switching (just change environment variables)
• Image input support: URL / base64 data URL / local file path
• Built-in system prompt templates: UI analysis, OCR, object detection, structured extraction, etc.
• Security: Automatic API key masking in logs, filters model-returned thinking/reasoning content
• MCP compliant: stdout reserved for JSON-RPC, logs go to stderr

Installation & Usage

Requirements: Node.js >= 18

Running as NPM package from MCP client (Recommended)

Configure your MCP client (e.g., Claude Desktop) to use npx:

{
  "mcpServers": {
    "vision-mcp": {
      "command": "npx",
      "args": ["-y", "@lutery/vision-mcp"],
      "env": {
        "VISION_MODEL_TYPE": "siliconflow",
        "VISION_API_KEY": "sk-your-key",
        "VISION_MODEL_NAME": "Qwen/Qwen2-VL-72B-Instruct",
        "VISION_API_BASE_URL": "https://api.siliconflow.cn/v1"
      }
    }
  }
}

Notes:

• VISION_MODEL_NAME / VISION_API_BASE_URL are optional (will use provider defaults)
• For more configuration options, refer to .env.example

You can also install globally and use the executable directly (binary name: vision-mcp):

npm i -g @lutery/vision-mcp
vision-mcp

Local development

cd mcp/vision_mcp
npm install
npm run build
node dist/index.js

On successful startup, you'll see Vision MCP Server is running on stdio in stderr.

Configuration (Environment Variables)

Minimum required:

• VISION_MODEL_TYPE: Select provider
• VISION_API_KEY: Key/token for the selected provider

Common optional:

| Variable | Description | Default | |----------|-------------|---------| | VISION_MODEL_NAME | Model name | Provider built-in defaults | | VISION_API_BASE_URL | API base URL (without specific endpoint) | Provider built-in defaults | | VISION_API_TIMEOUT | Timeout (milliseconds) | 60000 | | VISION_MAX_RETRIES | Maximum retry attempts | 2 | | VISION_STRICT_URL_VALIDATION | Strict validation for image URLs ending in .jpg/.jpeg/.png/.webp | true | | LOG_LEVEL | Log level: debug/info/warn/error | info |

Provider-specific configuration:

• Claude
  • VISION_CLAUDE_API_VERSION: Anthropic API version (default 2023-06-01)

MCP Tools

This server registers 3 tools:

1) analyze_image

Parameters:

{
  "image": "https://example.com/a.png",
  "prompt": "Describe the components in this interface",
  "output_format": "text",
  "template": "ui-analysis"
}

Field descriptions:

• image: Supports URL / base64 data URL / local path
• prompt: Your analysis task description
• output_format: text or json (hint preference; JSON not strictly validated)
• template: Optional system template (see list_templates below)

2) list_templates

Lists built-in system prompt templates (including id, usage description, etc.).

3) get_config

Returns currently active model configuration (API key is masked).

Image Input Specifications

Supports three input types:

1. URL

https://example.com/image.png

Strict validation enabled by default: URL must end with .jpg/.jpeg/.png/.webp, otherwise error. Can be relaxed with VISION_STRICT_URL_VALIDATION=false (warning only).

2. Base64 Data URL

data:image/png;base64,iVBORw0KGgo...

Supported MIME types: image/jpeg / image/jpg / image/png / image/webp.

3. Local file path

./test/image.png
D:\\path\\to\\image.jpg

Requires MCP Server process to have read access; only supports .jpg/.jpeg/.png/.webp.

Note: Gemini provider doesn't support direct URL image input; this project downloads URLs and converts to base64 in the Gemini adapter (subject to size and timeout limits).

About Streaming

All adapters enforce stream: false and parse as "complete JSON response".

SSE / text/event-stream responses are currently not supported (would require additional streaming parser implementation).

Development & Testing

cd mcp/vision_mcp
npm install
npm run build

Testing:

• Unit tests only (no API keys required):

npm run test:unit

• Integration tests (requires VISION_* environment variables configured):

npm test

Troubleshooting

1) Configuration loading failed: Missing VISION_MODEL_TYPE / Unsupported model type

• Ensure VISION_MODEL_TYPE is set
• Valid values: glm / siliconflow / modelscope / openai / claude / gemini / coding-plan

2) Missing VISION_API_KEY

• Ensure VISION_API_KEY is set (in .env or MCP client env)

3) 404 / endpoint errors

• VISION_API_BASE_URL must be the "base" URL, without specific endpoint
  • OpenAI / SiliconFlow / ModelScope: automatically appends /chat/completions
  • Claude: automatically appends /v1/messages (don't write baseUrl as .../v1)
  • Gemini: automatically appends /{apiVersion}/models/{model}:generateContent

4) Image URL validation failed

• Default requires URL to end with .jpg/.jpeg/.png/.webp
• To relax: VISION_STRICT_URL_VALIDATION=false

Security Notes

• Don't log to stdout (stdout reserved for MCP JSON-RPC), all logs go to stderr
• API keys are masked in logs
• Unconditionally filters model-returned thinking/reasoning content to avoid leaking internal inference information