This MCP (Model-Client-Protocol) server provides a powerful audio indexing and search capability built on top of the Pixeltable SDK. It allows you to:

Create searchable indexes of audio content
Transcribe audio files using OpenAI's Whisper model
Perform semantic search on audio content using embeddings
Retrieve relevant audio segments based on natural language queries

Features

Audio Transcription: Automatically transcribes audio files using OpenAI's Whisper model
Chunking: Splits audio into manageable segments for better search precision
Sentence Splitting: Further divides transcriptions into sentences for fine-grained retrieval
Semantic Search: Uses sentence embeddings to find content based on meaning, not just keywords
Multiple Indexes: Create and manage multiple audio indexes for different collections

Setup

Clone the repo

gh repo clone pixeltable/mcp-server-pixeltable

Docker Setup

Build the Docker image

cd servers/audio-index
docker build -t audio-index-mcp-server .

Run the Docker container

docker run -p 8080:8080 audio-index-mcp-server

This will start the MCP server on port 8080, making it accessible at http://localhost:8080.

Add the tool to Cursor

Go to Cursor MCP settings
Add MCP > Add Name > Type = 'SSE'
For the URL, enter: http://localhost:8080/sse

Available Tools

The server provides the following tools:

setup_audio_index: Create a new audio index
- Parameters: table_name (name for your index), openai_api_key (for Whisper transcription)
insert_audio: Add an audio file to an index
- Parameters: table_name (index to use), audio_location (URL or path to audio file)
query_audio: Search for content in an audio index
- Parameters: table_name (index to search), query_text (your search query), top_n (number of results, default=5)
list_tables: Show all available audio indexes

Requirements

The server requires the following dependencies:

pixeltable
openai-whisper
sentence-transformers
mcp
uvicorn
starlette
and other dependencies listed in requirements.txt