channel-chat

Download YouTube channel videos, transcribe, embed, and serve via MCP for semantic search in Claude.

Overview

channel-chat is a CLI tool that:

1. Downloads videos from YouTube channels
2. Extracts or generates transcripts with timestamps
3. Creates vector embeddings for semantic search
4. Serves content via MCP (Model Context Protocol) for use with Claude

Supports two deployment modes:

• Local: SQLite + sqlite-vec for personal use
• Cloudflare: Workers + D1 + Vectorize + R2 for cloud deployment

Prerequisites

• Node.js >= 20.0.0
• yt-dlp - brew install yt-dlp or pip install yt-dlp
• ffmpeg - brew install ffmpeg (required for video clips)

Installation

cd src
npm install
npm run build

To make the CLI globally available:

npm link

Environment Variables

Required (Local Mode)

export GOOGLE_API_KEY=...  # For Gemini embeddings

Optional

export ELEVENLABS_API_KEY=...  # For transcribing videos without subtitles

Cloudflare Mode

export CLOUDFLARE_WORKER_URL=https://your-worker.workers.dev
export CLOUDFLARE_ACCOUNT_ID=...
export CLOUDFLARE_R2_BUCKET=channel-chat-media
export CLOUDFLARE_R2_ACCESS_KEY_ID=...
export CLOUDFLARE_R2_SECRET_ACCESS_KEY=...
export CLOUDFLARE_WORKER_API_KEY=...  # Optional: for authenticated endpoints

CLI Usage

Add a Channel

# Index all videos from a channel (local SQLite)
channel-chat add "https://youtube.com/@channelname"

# Limit to first 50 videos with 3 concurrent workers
channel-chat add "https://youtube.com/@channelname" --limit 50 --concurrency 3

# Index to Cloudflare (downloads video, uploads to R2)
channel-chat add "https://youtube.com/@channelname" --cloudflare

Options:

• --limit <n> - Maximum videos to index
• --concurrency <n> - Parallel processing (default: 1)
• --cloudflare - Index to Cloudflare instead of local SQLite

Search Transcripts

channel-chat search "how to optimize database queries"
channel-chat search "kubernetes deployment" --limit 5

Options:

• -n, --limit <n> - Maximum results (default: 10)

List Channels

channel-chat list          # Show indexed channels
channel-chat list -v       # Verbose: show videos per channel

Index Single Video

channel-chat index-video "VIDEO_ID"

Cloudflare Setup

channel-chat cloudflare-setup  # Display setup instructions

MCP Server

Use channel-chat as an MCP server to search transcripts directly from Claude.

Quick Setup

channel-chat mcp --install

Manual Setup

Add to your MCP config (~/.config/claude/claude_desktop_config.json):

{
  "mcpServers": {
    "channel-chat": {
      "command": "node",
      "args": ["/path/to/channel-chat/src/dist/mcp-server.js"],
      "env": {
        "GOOGLE_API_KEY": "your-key"
      }
    }
  }
}

MCP Tools

Tool	Description
search_transcripts	Semantic search across indexed videos
list_indexed_channels	Show indexed channels with video counts
add_channel	Index a YouTube channel
index_video	Index a specific video
get_stats	Get indexing statistics
set_video_path	Set local video path for clip playback

MCP Resources

Resource	Description
ui://channel-chat/player.html	React-based video player UI
video://clip/{videoId}?start=X&duration=Y	On-the-fly video clip extraction

Cloudflare Deployment

For cloud deployment with video storage in R2:

1. Create Cloudflare Resources

cd cloudflare

# Create D1 database
npm run db:create
npm run db:migrate

# Create Vectorize index
npm run vectorize:create

# Create R2 bucket
npm run r2:create

2. Set API Key

wrangler secret put API_KEY

3. Deploy Worker

npm run deploy

4. Create R2 API Token

In Cloudflare Dashboard:

1. Go to R2 > Manage R2 API Tokens
2. Create token with Object Read/Write permissions
3. Add credentials to your environment

Worker Endpoints

Endpoint	Method	Description
/mcp	POST	MCP JSON-RPC interface
/api/index	POST	Index video content
/api/videos	GET	List indexed video IDs
/api/videos/:id	DELETE	Delete a video
/api/stats	GET	Get statistics
/	GET	Landing page
/ui	GET	Video player UI

Data Flow

Local Indexing

YouTube URL → getChannelVideos → downloadSubtitles → parseSubtitles
           → chunkTranscript (800 tokens, 15% overlap)
           → embedBatch (Google Gemini, 768D)
           → SQLite + sqlite-vec

Cloudflare Indexing

YouTube URL → downloadVideo → uploadToR2
           → chunks → POST /api/index
           → Workers AI embeddings → Vectorize + D1

Tech Stack

CLI

• Commander.js - CLI framework
• youtubei.js - YouTube data extraction
• yt-dlp - Video/subtitle download
• Google Gemini - Embeddings (768D)
• ElevenLabs Scribe - Transcription fallback
• better-sqlite3 + sqlite-vec - Vector storage
• MCP SDK - Claude integration

Cloudflare Worker

• Cloudflare Workers - Serverless runtime
• D1 - SQLite database
• Vectorize - Vector search
• R2 - Object storage
• Workers AI - Embeddings

Project Structure

channel-chat/
├── src/                    # Node.js CLI and MCP server
│   ├── cli.ts             # CLI commands
│   ├── mcp-server.ts      # MCP server (stdio/HTTP)
│   ├── downloader.ts      # YouTube video/subtitle fetching
│   ├── database.ts        # SQLite + vector operations
│   ├── transcriber.ts     # Subtitle parsing, transcription
│   ├── embedder.ts        # Google Gemini embeddings
│   ├── chunker.ts         # Token-based transcript chunking
│   ├── search.ts          # Vector similarity search
│   └── cloudflare-client.ts # R2 upload & Worker API
├── cloudflare/             # Cloudflare Worker
│   ├── src/
│   │   ├── index.ts       # Worker entry point
│   │   ├── api.ts         # API endpoints
│   │   ├── mcp-handler.ts # MCP tool implementations
│   │   ├── db.ts          # D1 operations
│   │   └── vectorize.ts   # Vector storage
│   ├── schema.sql         # D1 schema
│   └── wrangler.toml      # Worker config
└── ui/                     # React MCP App for video player

Database Location

Local database: ~/.channel-chat/channel_chat.db

License

MIT