magika-mcp

MCP server for Google Magika — AI-powered file type detection.

Magika uses a deep learning model to identify file types from their content, not just extensions. This MCP server makes Magika's capabilities available to any MCP client (Claude Code, Claude Desktop, etc.).

Quick Start

Add to your MCP client config:

{
  "mcpServers": {
    "magika": {
      "command": "npx",
      "args": ["-y", "magika-mcp"]
    }
  }
}

For Claude Code, add it with:

claude mcp add magika -- npx -y magika-mcp

Tools

identify_file

Identify a single file's content type.

Input: path (string) — file path Output: Enriched result with label, MIME type, group, description, extensions, confidence score, is_text flag.

identify_files

Batch-identify multiple files.

Input: paths (string[]) — array of file paths Output: Array of enriched results.

identify_content

Identify content from raw base64-encoded bytes.

Input: content (string) — base64-encoded file content Output: Enriched result.

identify_directory

Recursively scan a directory and identify all files.

Input:

• path (string) — directory path
• recursive (boolean, default: true) — scan recursively
• limit (number, default: 1000) — max files to process

Output: Array of enriched results with file paths.

get_content_type_info

Look up metadata for a known content type label (no file analysis).

Input: label (string) — content type label (e.g., "python", "pdf", "jpeg") Output: MIME type, group, description, extensions, is_text.

list_supported_types

List all content types Magika can detect.

Input: group (string, optional) — filter by group (e.g., "code", "image", "document", "archive") Output: Array of content types with metadata.

Example Output

{
  "path": "/path/to/file.py",
  "label": "python",
  "mime_type": "text/x-python",
  "group": "code",
  "description": "Python source",
  "extensions": ["py", "pyi"],
  "is_text": true,
  "score": 0.997,
  "overwrite_reason": "none"
}

How It Works

• Uses MagikaNode from the magika npm package (TensorFlow.js) for classification
• Enriches results with a bundled content types knowledge base (MIME types, groups, descriptions, extensions)
• The model (~5MB) downloads automatically on first use and is cached by TensorFlow.js
• Lazy initialization — model loads on first tool call, not at server startup

Requirements

• Node.js >= 18

License

MIT