| name | yt-fetch |
| description | Fetch a YouTube video as readable markdown (title, channel, description, clean transcript). WebFetch can't read YouTube — this wraps a tested yt-dlp pipeline. |
| argument-hint | <youtube-url> [output-dir] |
| allowed-tools | Bash, Read |
yt-fetch — make a YouTube video readable
Reach for this whenever you need to read / study / summarize / quote a YouTube video and WebFetch returns nothing — YouTube blocks AI fetch. Produces a clean transcript + metadata as markdown.
WebFetch can't read YouTube. This runs a tested yt-dlp wrapper that pulls the video's captions + metadata (no video download) and cleans the auto-caption "rolling window" duplication into plain text.
Use
scripts/yt-fetch.sh <youtube-url> [output-dir]
It prints the path to a single <video-id>.md (default dir: cwd) holding title · channel · duration · url · description · clean transcript. Then Read that file to study/quote/summarize. Capture the path:
md=$(scripts/yt-fetch.sh "https://youtu.be/XXXX" /tmp) # then Read "$md"
- No captions → the Transcript section says so (nothing to read — pick another source).
- Unavailable / private / blocked → exits non-zero with a message.
- Single video only (not playlists). English captions by default.
Why a wrapper, not a raw command (the traps it handles)
Verified against the yt-dlp docs — the naive one-liner fails in ways the script fixes:
--sub-langs "en.*" is a regex that matches every English variant (en, en-orig, en-de-DE, en-ja…) → HTTP 429. The script uses a bounded exact list en-orig,en,en-US,en-GB.--print implies --simulate → it silently skips writing files. The script uses --write-info-json + jq.- Auto-caption
.vtt is full of rolling-window duplicate lines; scripts/vtt2text.py strips them to clean text (no ffmpeg).
To fetch non-English or more langs, edit --sub-langs in scripts/yt-fetch.sh. If this stays heavily used, that's the single place to harden it.
