Video Watch MCP

Let Claude "watch" videos with you. Send a TikTok, YouTube, or any video link - Claude sees the frames and reads the transcript.

Fully cloud-hosted. No local processing. Works on Claude Desktop, Claude mobile, anywhere MCP works.

What it does

Three tools, pick based on content:

Tool	Returns	Best for
video_listen	Transcript only	Talking heads, podcasts, commentary
video_see	Frames only	Dance, visual art, memes, scenery
watch_video	Both	When audio AND visuals both matter

1. You send Claude a video link
2. Claude picks the right tool (or you tell it which)
3. Cloud service downloads, extracts what's needed
4. Claude receives just what it needs - no context bloat
5. You watch it "together"

Quick Start (5 minutes)

1. Create a Modal account

Go to modal.com and sign up. Free tier gives you $30/month in credits - enough for thousands of short videos.

2. Install Modal CLI

pip install modal
modal token set --token-id YOUR_TOKEN_ID --token-secret YOUR_TOKEN_SECRET

(Get your token from Modal's dashboard after signup)

3. Deploy

git clone https://github.com/codependentai/video-watch-mcp.git
cd video-watch-mcp
modal deploy mcp_remote.py

You'll get a URL like: https://yourusername--video-watch-mcp-mcp-server.modal.run

4. Add to Claude Desktop

Edit your claude_desktop_config.json:

Windows: %APPDATA%\Claude\claude_desktop_config.json Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

Add under mcpServers:

{
  "mcpServers": {
    "video-watch": {
      "url": "https://yourusername--video-watch-mcp-mcp-server.modal.run"
    }
  }
}

5. Use it

Restart Claude Desktop. Send any video link and ask Claude to watch it:

"Watch this with me: https://tiktok.com/..."

Claude will see the frames and read the transcript.

Supported Platforms

Anything yt-dlp supports:

• TikTok
• YouTube
• Instagram Reels
• Twitter/X videos
• Reddit videos
• Facebook
• Vimeo
• And 1000+ more

Cost

With Modal's free tier ($30/month credits):

Video Length	Approx. Cost	Videos per Month
30 sec	~$0.002	~15,000
5 min	~$0.01	~3,000
30 min	~$0.05	~600

You'll never hit the limit with normal use.

How it works

You send a link
       ↓
Claude calls watch_video(url)
       ↓
Modal spins up a container with ffmpeg + whisper
       ↓
yt-dlp downloads the video
       ↓
ffmpeg extracts frames (with timestamps burned in)
       ↓
Whisper transcribes the audio
       ↓
Returns frames as images + transcript text
       ↓
Claude sees everything, you discuss it together

Files

• mcp_remote.py - The full MCP server (deploy this)
• video_watch.py - Standalone video processor with web endpoint (if you just want the API)

Configuration

In mcp_remote.py you can adjust:

• fps - Frames per second to extract (default 0.5 = one frame every 2 seconds)
• max_frames - Maximum frames to return (default 10, max 20)
• whisper model - Using "base" for speed, can use "small" or "medium" for accuracy

Updating

Pull the latest code and redeploy — that's it:

cd video-watch-mcp
git pull
modal deploy mcp_remote.py

Your MCP URL stays the same, no client config changes needed.

Coming from the old maryfellowes/video-watch-mcp repo? That repo has been removed. Update your remote:

git remote set-url origin https://github.com/codependentai/video-watch-mcp.git
git pull
modal deploy mcp_remote.py

Limitations

• Very long videos (30+ min) may timeout
• Audio-only content won't have frames (obviously)
• Some DRM-protected content won't download
• Whisper transcription is good but not perfect

Privacy

• Videos are processed in ephemeral containers - nothing stored
• No logs of what you watch
• Your Modal account, your data

License

MIT - do whatever you want with it.

---

Built by Codependent AI.

Support

Built by Codependent AI.