video-watcher MCP server

An MCP server that lets Claude Desktop "watch" videos from YouTube, Instagram, Facebook, or a direct video URL (mp4/webm/etc), and answer questions about them.

What it does

For any supported URL, the server:

1. Downloads the video with yt-dlp (works for YouTube, Instagram, Facebook, and hundreds of other sites, plus falls back to a raw HTTP download for plain file URLs).
2. Extracts a transcript — YouTube captions via youtube-transcript-api if available, otherwise audio transcription via faster-whisper.
3. Extracts 5 key frames (sampled across the video's timeline) as base64 JPEGs using OpenCV.

Results are cached in memory per URL for the life of the process, so calling multiple tools on the same video only downloads/transcribes it once.

Tools

• watch_video(url) — full analysis: title, duration, transcript, frames.
• summarize_video(url) — ~200-word summary + bullet list of main points.
• answer_about_video(url, question) — transcript + frames so Claude can answer a specific question about the video.

Setup

cd D:\Projects\video-mcp-server
pip install -r requirements.txt

You'll also need ffmpeg on your PATH (required by both yt-dlp for muxing and faster-whisper for audio decoding). On Windows, install via winget install ffmpeg or download from https://ffmpeg.org and add it to PATH.

Test the server standalone first:

python video_mcp_server.py

It should sit waiting on stdio (no output = good; it logs to stderr).

Connect to Claude Desktop

Edit %APPDATA%\Claude\claude_desktop_config.json (create it if it doesn't exist) and merge in:

{
  "mcpServers": {
    "video-watcher": {
      "command": "python",
      "args": ["D:\\Projects\\video-mcp-server\\video_mcp_server.py"]
    }
  }
}

Adjust the path if you cloned this elsewhere. If python isn't on PATH inside Claude Desktop's environment, use the full interpreter path instead, e.g. "C:\\Users\\<you>\\AppData\\Local\\Programs\\Python\\Python311\\python.exe".

Fully quit and reopen Claude Desktop after editing the config (closing the window is not enough on Windows either — use the system tray icon to Quit).

Testing it

Try these once connected:

• watch_video("https://www.youtube.com/watch?v=dQw4w9WgXcQ")
• summarize_video("https://www.youtube.com/shorts/<some-short-id>")
• answer_about_video("https://www.instagram.com/reel/<reel-id>/", "What product is being promoted?")

Ask Claude directly, e.g.: "Watch this video and summarize it: <url>"

Notes & limitations

• Instagram: public posts/reels work via yt-dlp. Private accounts or posts requiring login will fail with a clear error — this server does not store or use Instagram credentials. instagrapi is listed as an optional dependency for future extension but is not required by the current code path.
• Facebook: public videos only; the same login-wall limitation applies.
• Whisper fallback: the first call that needs transcription will download the base Whisper model (~150MB) on first use.
• Rate limits: platforms may throttle repeated requests; the server surfaces a clear "rate limited, try again later" error rather than hanging.
• Storage: downloaded videos are written to a temp cache directory (%TEMP%\video_mcp_cache) and are not automatically cleaned up — clear it periodically if disk space matters.