| name | youtube-watcher |
| description | Fetch and summarize YouTube video transcripts. Use when asked to watch, summarize, or answer questions about a YouTube video. |
| agents | ["research_agent","deep_research_agent"] |
YouTube Watcher
When Travis sends a YouTube link or says "summarize this video", "watch this", "what's this video about":
Flow
- Extract the video URL from the message
- Fetch the transcript using yt-dlp (installed at /opt/homebrew/bin/yt-dlp)
- Summarize or answer questions based on the transcript
How to Fetch Transcript
Run this command via the terminal/exec tool:
yt-dlp --write-auto-sub --sub-lang en --skip-download --print-json "VIDEO_URL" 2>/dev/null | python3 -c "
import json, sys
data = json.load(sys.stdin)
subs = data.get('subtitles', {}).get('en') or data.get('automatic_captions', {}).get('en', [])
if subs:
# Get the vtt/json subtitle URL
for s in subs:
if s.get('ext') == 'json3':
print(s['url'])
break
else:
print(subs[0]['url'])
else:
print('NO_SUBS')
"
Simpler approach — just get the subtitle text directly:
yt-dlp --write-auto-sub --sub-lang en --skip-download -o "/tmp/yt_sub" "VIDEO_URL" 2>/dev/null && cat /tmp/yt_sub.en.vtt 2>/dev/null | grep -v "^$" | grep -v "^WEBVTT" | grep -v "^Kind:" | grep -v "^Language:" | grep -v "^[0-9][0-9]:" | head -200
After Getting Transcript
- "Summarize this video" → Read transcript, give 3-5 bullet point summary
- "What does X say about Y?" → Search transcript for relevant sections, quote key parts
- "Key takeaways?" → Extract main points, actionable items
- "Is this worth watching?" → Quick summary + opinion on content quality
Also Get Video Info
yt-dlp --print "%(title)s|%(uploader)s|%(duration_string)s|%(view_count)s" "VIDEO_URL" 2>/dev/null
Returns: title | channel | duration | views
What NOT to Do
- Don't download the actual video file (skip-download always)
- Don't say "I can't watch videos" — you CAN via transcripts
- If no subtitles available, say so and offer to search for a text summary of the video topic instead
