Youtube MCP

Fast, minimal, and reliable YouTube MCP for AI agents.

Install

Claude Code

claude mcp add youtube -- uvx youtube-watch-mcp

Before publish (local dev): point uvx at the checkout instead: claude mcp add youtube -- uvx --from /path/to/youtube-mcp youtube-watch-mcp

Claude Desktop / Codex / other MCP clients

Add to the client's MCP config:

{
  "mcpServers": {
    "youtube": {
      "command": "uvx",
      "args": ["youtube-watch-mcp"]
    }
  }
}

CLI only

uvx --from youtube-watch-mcp youtube-watch-mcp-cli info "https://youtu.be/VIDEO_ID"

That's it. uvx pulls youtube-watch-mcp, yt-dlp, and dependencies into an isolated environment automatically. Nothing to install globally.

Optional: ffmpeg on PATH is required only for --asr (speech-to-text on caption-less videos). Core transcript extraction needs nothing.

Related MCP server: VidLens

Optional API key

A YouTube Data API key is not needed to read videos. Add one only to enable cross-YouTube search:

claude mcp add youtube -e YOUTUBE_API_KEY=your_key -- uvx youtube-watch-mcp

Transcript extraction never uses the key (YouTube only allows caption download for video owners).

Tools

Design principle: pull, don't dump. Transcripts write to a local cache file; tools return a path and a short preview. The agent reads or searches on demand — long videos never flood the context.

/get_info $url
/get_transcript $url
/search_transcript $url
/get_segment $url

Architecture

Adapters (thin):   cli.py   mcp_server.py   skill
                        │  call
Core (all logic):  fetch → clean → chunk → cache
                        │  uses
Backends:          youtube-transcript-api · yt-dlp · faster-whisper

Fetch fallback chain:

1. youtube-transcript-api — fastest, no download

2. yt-dlp auto-captions

3. yt-dlp manual captions

4. --asr: audio → local faster-whisper

On yt-dlp failure the engine self-updates yt-dlp and retries once — most breakage is a stale yt-dlp.

Caching: results are keyed by video ID under ~/.cache/youtube-mcp/<id>/. Repeat calls are instant.

Cleaning: auto-captions are de-duplicated (rolling-caption overlap removed), stripped of timestamps and [Music] noise, and whitespace-collapsed before the agent ever sees them.

Requirements

• Python 3.11+ (managed automatically by uvx)

• ffmpeg — optional, only for --asr

Docs

• FAQ — keyless? no-caption videos? --from gotcha? long-video handling?

• Architecture — core/adapter split, fetch chain, the cleaning moat, cache.

License

MIT