# mcp-vision

MCP server for vision capabilities - screenshot and camera analysis using Ollama vision models.

## Features

- **Screenshot Analysis**: Capture and analyze screenshots with AI
- **Camera Capture**: Take photos from webcam and analyze them
- **Image Analysis**: Analyze existing image files
- **Streaming Output**: Real-time streaming of AI analysis
- **Multiple Models**: Support for various vision models (llava, bakllava, etc.)

## Installation

```bash
cd /Users/bard/Code/mcp-vision
npm install
```

## Prerequisites

- Ollama must be running with a vision model installed:
  ```bash
  ollama pull llava
  ```
- macOS (for screenshot functionality)
- Camera access (for webcam features)

## Tools

### vision_screenshot
Take a screenshot and analyze it with AI.

```javascript
{
  prompt: "What application is open?",  // optional
  model: "llava",                       // optional
  region: {                             // optional
    x: 100,
    y: 100,
    width: 500,
    height: 400
  }
}
```

### vision_camera
Capture from camera and analyze.

```javascript
{
  prompt: "What do you see?",  // optional
  model: "llava",              // optional
  device: "FaceTime HD Camera" // optional
}
```

### vision_analyze_image
Analyze an existing image file.

```javascript
{
  path: "/path/to/image.jpg",
  prompt: "Describe this image",  // optional
  model: "llava"                  // optional
}
```

### vision_list_cameras
List available camera devices.

## Usage with Claude Desktop

Add to your Claude Desktop configuration:

```json
{
  "mcpServers": {
    "vision": {
      "command": "node",
      "args": ["/Users/bard/Code/mcp-vision/src/index.js"]
    }
  }
}
```

## Integration with ELVIS

This tool can be integrated with ELVIS for enhanced visual context:

1. Use `vision_screenshot` to capture current screen state
2. Pass the analysis to `elvis_delegate` for context-aware task processing
3. ELVIS can use visual information to better understand and complete tasks

## Example Workflow

```javascript
// 1. Analyze what's on screen
vision_screenshot({ prompt: "What code is visible?" })

// 2. Use with ELVIS
elvis_delegate({ 
  task: "Fix the syntax error shown",
  context: "Based on the screenshot analysis"
})
```

## Streaming Output

The tool streams AI responses in real-time, providing immediate feedback as the model analyzes images. This is shown in the MCP server logs and can be used for progress tracking.