AgenticVision-MCP

MCP server for AgenticVision — universal LLM access to persistent visual memory.

What it does

AgenticVision-MCP exposes the AgenticVision engine over the Model Context Protocol (JSON-RPC 2.0 over stdio). Any MCP-compatible LLM gains persistent visual memory — capture screenshots, embed with CLIP ViT-B/32, compare, recall.

Install

cargo install agentic-vision-mcp

Configure Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "vision": {
      "command": "agentic-vision-mcp",
      "args": ["serve", "--vision", "~/vision.avis"]
    }
  }
}

Configure VS Code / Cursor

Add to .vscode/settings.json:

{
  "mcp.servers": {
    "vision": {
      "command": "agentic-vision-mcp",
      "args": ["serve", "--vision", "${workspaceFolder}/.vision/project.avis"]
    }
  }
}

MCP Surface Area

Category	Count	Examples
Tools	10	`vision_capture`, `vision_similar`, `vision_diff`, `vision_compare`, `vision_query`, `vision_ocr`, `vision_track`, `vision_link`, `session_start`, `session_end`
Resources	6	`avis://capture/{id}`, `avis://session/{id}`, `avis://timeline/{start}/{end}`, `avis://similar/{id}`, `avis://stats`, `avis://recent`
Prompts	4	`observe`, `compare`, `track`, `describe`

How it works

Capture — vision_capture accepts images from files, base64, or URLs. Embeds with CLIP ViT-B/32, stores in .avis binary format.
Query — vision_query retrieves by time, description, or recency. vision_similar finds visually similar captures by cosine similarity.
Compare — vision_compare for side-by-side LLM analysis. vision_diff for pixel-level differencing with 8×8 grid region detection.
Link — vision_link connects captures to AgenticMemory cognitive graph nodes.

Performance

Operation	Time
MCP tool round-trip	7.2 ms
Image capture	47 ms
Similarity search (top-5)	1-2 ms
Visual diff	<1 ms

License

MIT

agentic-vision-mcp 0.1.1