Takuma-AI/youtube-transcribe-mcp

YouTube Transcribe MCP Server

Enables Claude to download and transcribe YouTube videos directly into Takuma OS's knowledge base.

Features

• Download & Transcribe: Automatically downloads YouTube videos and transcribes them using OpenAI's Whisper
• Knowledge Integration: Saves transcripts directly to knowledge/_inbox/ for processing
• Configurable Models: Choose from tiny to large Whisper models based on accuracy needs
• Metadata Preservation: Captures channel, upload date, and URL alongside transcripts
• Timestamp Support: Includes VTT files for precise timestamp reference

Prerequisites

The following tools must be installed on your system:

1. yt-dlp: pip install yt-dlp
2. whisper: pip install openai-whisper
3. ffmpeg: brew install ffmpeg (macOS) or apt-get install ffmpeg (Linux)

Installation

Already installed! The server is connected to Claude Code at:

/Users/kate/Documents/Manual Library/Projects/takuma-os/tools/mcp/youtube-transcribe

Usage

Basic Transcription

Ask Claude to transcribe any YouTube video:

"Transcribe this video: https://youtube.com/watch?v=..."

With Options

Specify model or keep audio:

"Transcribe this video using the large model and keep the audio file"

Extract Playlist Videos

"Get all video URLs from this playlist: [playlist URL]"
"Get playlist videos with titles and metadata"

Batch Transcribe

"Transcribe all videos from this playlist"
"Transcribe these 5 videos: [URLs]"

Check Dependencies

"Check if YouTube transcription dependencies are installed"

Configure Settings

"Configure YouTube transcribe to use the turbo model"

Available Tools

1. transcribe_video - Main transcription tool

   • url: YouTube video URL (required)
   • whisper_model: Model size (tiny/base/small/medium/large/turbo)
   • keep_audio: Whether to save the audio file
   • save_path: Custom save location

2. get_playlist_videos - Extract video URLs from a playlist

   • playlist_url: YouTube playlist URL (required)
   • include_metadata: Include titles and durations (optional)

3. batch_transcribe_videos - Transcribe multiple videos in sequence

   • video_urls: List of YouTube video URLs (required)
   • whisper_model: Model to use for all videos
   • save_path: Base path for all transcripts

4. check_youtube_dependencies - Verify all dependencies are installed

5. configure_youtube_transcribe - Update default settings

6. get_youtube_config - View current configuration

Whisper Models

• tiny: 39M params, fastest, basic accuracy
• base: 74M params, good balance (default)
• small: 244M params, better accuracy
• medium: 769M params, high accuracy
• large: 1550M params, best accuracy
• turbo: 809M params, fast and accurate

Output Structure

Transcripts are saved to:

knowledge/_inbox/[video-title]/
├── transcript.md      # Formatted transcript
├── timestamps.vtt     # Timestamp file
└── audio.mp3         # (optional) Audio file

Configuration

Default settings in config.json:

{
  "whisper_model": "base",
  "keep_audio": false,
  "audio_format": "mp3",
  "default_save_path": "knowledge/_inbox"
}

Integration with Takuma OS

This MCP server extends Takuma OS's knowledge capture capabilities by:

• Automatically organizing video content in the inbox for processing
• Preserving metadata for context and attribution
• Creating markdown-formatted transcripts ready for knowledge synthesis
• Supporting the OS's philosophy of capturing authentic content from external sources

Troubleshooting

1. Missing dependencies: Run check_youtube_dependencies() to see what's missing
2. Transcription fails: Check ffmpeg is installed and accessible
3. Download fails: Verify the YouTube URL is valid and accessible
4. Out of memory: Use a smaller Whisper model (tiny or base)

Privacy Note

• Audio files are processed locally using Whisper
• No data is sent to external transcription services
• Downloaded content respects YouTube's terms of service via yt-dlp