multimodal-reader-mcp

MCP server for reading local audio and video files with Google Gen AI and returning structured observations, timelines, and transcripts.

It analyzes a local media file and returns:

• a short summary
• a timeline of key moments
• transcript snippets for spoken or visible text
• key observations and notable signals
• relevant clues tailored to the user's question
• open questions plus a confidence level

Requirements

• uv
• Python 3.14
• GOOGLE_API_KEY

Model configuration

The default model is gemini-2.5-flash.

You can override the default model for all requests by setting:

• MULTIMODAL_READER_MODEL

Users can also still pass model directly to the read_media tool call.

MCP client configuration

Example Cursor MCP config:

{
  "mcpServers": {
    "multimodal-reader": {
      "command": "uvx",
      "args": ["multimodal-reader-mcp"],
      "env": {
        "GOOGLE_API_KEY": "${env:GOOGLE_API_KEY}",
        "MULTIMODAL_READER_MODEL": "gemini-2.5-flash"
      }
    }
  }
}

Tool

The package exposes one MCP tool:

• read_media(file_path, question=None, model="gemini-2.5-flash")

file_path must be an absolute path to a local media file.