AI Vision MCP Server

A powerful Model Context Protocol (MCP) server that provides AI-powered image and video analysis using Google Gemini and Vertex AI models.

Features

• Dual Provider Support: Choose between Google Gemini API and Vertex AI
• Multimodal Analysis: Support for both image and video content analysis
• Flexible File Handling: Upload via multiple methods (URLs, local files, base64)
• Storage Integration: Built-in Google Cloud Storage support
• Comprehensive Validation: Zod-based data validation throughout
• Error Handling: Robust error handling with retry logic and circuit breakers
• TypeScript: Full TypeScript support with strict type checking

Quick Start

Pre-requisites

You could choose either to use google provider or vertex_ai provider. For simplicity, google provider is recommended.

Below are the environment variables you need to set based on your selected provider. (Note: It’s recommended to set the timeout configuration to more than 5 minutes for your MCP client).

(i) Using Google AI Studio Provider

export IMAGE_PROVIDER="google" # or vertex_ai
export VIDEO_PROVIDER="google" # or vertex_ai
export GEMINI_API_KEY="your-gemini-api-key"

Get your Google AI Studio's api key here

(ii) Using Vertex AI Provider

export IMAGE_PROVIDER="vertex_ai"
export VIDEO_PROVIDER="vertex_ai"
export VERTEX_CREDENTIALS="/path/to/service-account.json"
export GCS_BUCKET_NAME="your-gcs-bucket"

Refer to the guideline here on how to set this up.

Installation

Below are the installation guide for this MCP on different MCP clients, such as Claude Desktop, Claude Code, Cursor, Cline, etc.

{
  "mcpServers": {
    "ai-vision-mcp": {
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "google",
        "VIDEO_PROVIDER": "google",
        "GEMINI_API_KEY": "your-gemini-api-key"
      }
    }
  }
}

{
  "mcpServers": {
    "ai-vision-mcp": {
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "vertex_ai",
        "VIDEO_PROVIDER": "vertex_ai",
        "VERTEX_CREDENTIALS": "/path/to/service-account.json",
        "GCS_BUCKET_NAME": "ai-vision-mcp-{VERTEX_PROJECT_ID}"
      }
    }
  }
}

claude mcp add ai-vision-mcp \
  -e IMAGE_PROVIDER=google \
  -e VIDEO_PROVIDER=google \
  -e GEMINI_API_KEY=your-gemini-api-key \
  -- npx ai-vision-mcp

claude mcp add ai-vision-mcp \
  -e IMAGE_PROVIDER=vertex_ai \
  -e VIDEO_PROVIDER=vertex_ai \
  -e VERTEX_CREDENTIALS=/path/to/service-account.json \
  -e GCS_BUCKET_NAME=ai-vision-mcp-{VERTEX_PROJECT_ID} \
  -- npx ai-vision-mcp

{
  "env": {
    "MCP_TIMEOUT": "20000", // Give the MCP server 20s to start.
    "MCP_TOOL_TIMEOUT": "300000" // Allow each tool calls before timeout.
  }
}

{
  "mcpServers": {
    "ai-vision-mcp": {
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "google",
        "VIDEO_PROVIDER": "google",
        "GEMINI_API_KEY": "your-gemini-api-key"
      }
    }
  }
}

{
  "mcpServers": {
    "ai-vision-mcp": {
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "vertex_ai",
        "VIDEO_PROVIDER": "vertex_ai",
        "VERTEX_CREDENTIALS": "/path/to/service-account.json",
        "GCS_BUCKET_NAME": "ai-vision-mcp-{VERTEX_PROJECT_ID}"
      }
    }
  }
}

{
  "mcpServers": {
    "timeout": 300, 
    "type": "stdio",
    "ai-vision-mcp": {
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "google",
        "VIDEO_PROVIDER": "google",
        "GEMINI_API_KEY": "your-gemini-api-key"
      }
    }
  }
}

{
  "mcpServers": {
    "ai-vision-mcp": {
      "timeout": 300,
      "type": "stdio",
      "command": "npx",
      "args": ["ai-vision-mcp"],
      "env": {
        "IMAGE_PROVIDER": "vertex_ai",
        "VIDEO_PROVIDER": "vertex_ai",
        "VERTEX_CREDENTIALS": "/path/to/service-account.json",
        "GCS_BUCKET_NAME": "ai-vision-mcp-{VERTEX_PROJECT_ID}"
      }
    }
  }
}

npx ai-vision-mcp

MCP Tools

The server provides four main MCP tools:

1) analyze_image

Analyzes an image using AI and returns a detailed description.

Parameters:

• imageSource (string): URL, base64 data, or file path to the image
• prompt (string): Question or instruction for the AI
• options (object, optional): Analysis options including temperature and max tokens

Examples:

1. Analyze image from URL:

{
  "imageSource": "https://plus.unsplash.com/premium_photo-1710965560034-778eedc929ff",
  "prompt": "What is this image about? Describe what you see in detail."
}

2. Analyze local image file:

{
  "imageSource": "C:\\Users\\username\\Downloads\\image.jpg",
  "prompt": "What is this image about? Describe what you see in detail."
}

2) compare_images

Compares multiple images using AI and returns a detailed comparison analysis.

Parameters:

• imageSources (array): Array of image sources (URLs, base64 data, or file paths) - minimum 2, maximum 4 images
• prompt (string): Question or instruction for comparing the images
• options (object, optional): Analysis options including temperature and max tokens

Examples:

1. Compare images from URLs:

{
  "imageSources": [
    "https://example.com/image1.jpg",
    "https://example.com/image2.jpg"
  ],
  "prompt": "Compare these two images and tell me the differences"
}

2. Compare mixed sources:

{
  "imageSources": [
    "https://example.com/image1.jpg",
    "C:\\\\Users\\\\username\\\\Downloads\\\\image2.jpg",
    "data:image/jpeg;base64,/9j/4AAQSkZJRgAB..."
  ],
  "prompt": "Which image has the best lighting quality?"
}

3) detect_objects_in_image

Detects objects in an image using AI vision models and generates annotated images with bounding boxes. Returns detected objects with coordinates and either saves the annotated image to a file or temporary directory.

Parameters:

• imageSource (string): URL, base64 data, or file path to the image
• prompt (string): Custom detection prompt describing what to detect or recognize in the image
• outputFilePath (string, optional): Explicit output path for the annotated image

Configuration: This function uses optimized default parameters for object detection and does not accept runtime options parameter. To customize the AI parameters (temperature, topP, topK, maxTokens), use environment variables:

# Recommended settings for object detection (these are now the defaults)
export TEMPERATURE_FOR_DETECT_OBJECTS_IN_IMAGE=0.0     # Deterministic responses
export TOP_P_FOR_DETECT_OBJECTS_IN_IMAGE=0.95          # Nucleus sampling
export TOP_K_FOR_DETECT_OBJECTS_IN_IMAGE=30            # Vocabulary selection
export MAX_TOKENS_FOR_DETECT_OBJECTS_IN_IMAGE=8192     # High token limit for JSON

File Handling Logic:

1. Explicit outputFilePath provided → Saves to the exact path specified
2. If not explicit outputFilePath → Automatically saves to temporary directory

Response Types:

• Returns file object when explicit outputFilePath is provided
• Returns tempFile object when explicit outputFilePath is not provided so the image file output is auto-saved to temporary folder
• Always includes detections array with detected objects and coordinates
• Includes summary with percentage-based coordinates for browser automation

Examples:

1. Basic object detection:

{
  "imageSource": "https://example.com/image.jpg",
  "prompt": "Detect all objects in this image"
}

2. Save annotated image to specific path:

{
  "imageSource": "C:\\Users\\username\\Downloads\\image.jpg",
  "outputFilePath": "C:\\Users\\username\\Documents\\annotated_image.png"
}

3. Custom detection prompt:

{
  "imageSource": "data:image/jpeg;base64,/9j/4AAQSkZJRgAB...",
  "prompt": "Detect and label all electronic devices in this image"
}

4) analyze_video

Analyzes a video using AI and returns a detailed description.

Parameters:

• videoSource (string): YouTube URL, GCS URI, or local file path to the video
• prompt (string): Question or instruction for the AI
• options (object, optional): Analysis options including temperature and max tokens

Supported video sources:

• YouTube URLs (e.g., https://www.youtube.com/watch?v=...)
• Local file paths (e.g., C:\Users\username\Downloads\video.mp4)

Examples:

1. Analyze video from YouTube URL:

{
  "videoSource": "https://www.youtube.com/watch?v=9hE5-98ZeCg",
  "prompt": "What is this video about? Describe what you see in detail."
}

2. Analyze local video file:

{
  "videoSource": "C:\\Users\\username\\Downloads\\video.mp4",
  "prompt": "What is this video about? Describe what you see in detail."
}

Note: Only YouTube URLs are supported for public video URLs. Other public video URLs are not currently supported.

Environment Variables

# AI Parameters configuration
TEMPERATURE=0.3
MAX_TOKENS=600

# Task-specific overrides
TEMPERATURE_FOR_IMAGE=0.1  # More precise for image analysis
MAX_TOKENS_FOR_VIDEO=1200   # Longer responses for video content

# Function-specific overrides
TEMPERATURE_FOR_ANALYZE_IMAGE=0.05     # Very precise for single image analysis
TEMPERATURE_FOR_COMPARE_IMAGES=0.2     # More creative for comparisons
MAX_TOKENS_FOR_COMPARE_IMAGES=1500      # Longer responses for image comparisons
TEMPERATURE_FOR_DETECT_OBJECTS_IN_IMAGE=0.0   # Deterministic for object detection
MAX_TOKENS_FOR_DETECT_OBJECTS_IN_IMAGE=8192   # Higher token limit for structured output
TEMPERATURE_FOR_ANALYZE_VIDEO=0.1      # Precise video analysis

# Model selection configuration
ANALYZE_IMAGE_MODEL="gemini-2.5-flash-lite"          # Fast, cost-effective for single image analysis
COMPARE_IMAGES_MODEL="gemini-2.5-flash-lite"
DETECT_OBJECTS_IN_IMAGE_MODEL="gemini-2.5-flash-lite"  # Structured output capable model
ANALYZE_VIDEO_MODEL="gemini-2.5-flash-pro"           # Most capable for video analysis

# Task-specific models (existing pattern still works)
IMAGE_MODEL="gemini-2.5-flash"
VIDEO_MODEL="gemini-2.5-flash-pro"

# Resolution order for analyze_image:
# 1. ANALYZE_IMAGE_MODEL
# 2. IMAGE_MODEL
# 3. System default ("gemini-2.5-flash-lite")

# Resolution order for detect_objects_in_image:
# 1. DETECT_OBJECTS_IN_IMAGE_MODEL
# 2. IMAGE_MODEL
# 3. System default ("gemini-2.5-flash-lite")

# LLM can still override at runtime via tool parameters

Development

Prerequisites

• Node.js 18+
• npm or yarn

Setup

# Clone the repository
git clone https://github.com/tan-yong-sheng/ai-vision-mcp.git
cd ai-vision-mcp

# Install dependencies
npm install

# Build the project
npm run build

# Start development server
npm run dev

Scripts

• npm run build - Build the TypeScript project
• npm run dev - Start development server with watch mode
• npm run lint - Run ESLint
• npm run format - Format code with Prettier
• npm start - Start the built server

Architecture

The project follows a modular architecture:

src/
├── providers/          # AI provider implementations
│   ├── gemini/        # Google Gemini provider
│   ├── vertexai/      # Vertex AI provider
│   └── factory/       # Provider factory
├── services/          # Core services
│   ├── ConfigService.ts
│   └── FileService.ts
├── storage/           # Storage implementations
├── file-upload/       # File upload strategies
├── types/            # TypeScript type definitions
├── utils/            # Utility functions
└── server.ts         # Main MCP server

Error Handling

The server includes comprehensive error handling:

• Validation Errors: Input validation using Zod schemas
• Network Errors: Automatic retries with exponential backoff
• Authentication Errors: Clear error messages for API key issues
• File Errors: Handling for file size limits and format restrictions

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Google for the Gemini and Vertex AI APIs
• The Model Context Protocol team for the MCP framework
• All contributors and users of this project