Search YouTube videos and get back ready-to-use results with transcript highlights: the passages where each video actually talks about your topic, with timestamped links.

Best for: finding videos on a topic, finding what creators say about something, collecting examples (hooks, claims, tutorials) across many videos. Returns: one merged, de-duplicated list across all queries. Each video has metadata (channel, views, date, duration, views/day), YouTube's matched snippet and key moments, plus transcript highlights for the top results.

Query tips:

• Send several queries in ONE call to cover the topic: synonyms, angles and languages, e.g. ["AI productivity workflow", "how I use AI to save time at work", {"query": "productividad con IA", "hl": "es", "gl": "ES"}].

• Write queries the way a creator would title the video, not as keyword soup.

• Use filters (uploaded, publishedAfter, minViews, duration, shorts) instead of asking for more results.

• If highlights are not enough, call youtube_fetch on the best videos (it can return only a time range, e.g. the first 30 s for a hook, or only passages matching a query).

Read YouTube videos: transcript (whole, a time range, a chapter, or only passages matching a query), real video frames, details, chapters, comments and thumbnail. Batch many videos in one call.

Best for: reading what was said after youtube_search, pulling the same slice from many videos (e.g. the first 30 s of each video to study hooks), quoting with exact timestamps. Returns: per video a header (title, channel, exact publish date, views, likes, duration) and the requested parts. Transcript lines carry [m:ss] stamps; links with ?t= jump to the moment. Per-video failures are reported inline, never dropped.

Tips:

• Hooks / intros: start=0, end=30 (plus parts ["details","transcript","thumbnail","frames"] if the visual hook matters).

• Frames are real decoded video frames (up to 1280 px): evenly spaced over start–end / the chapter / the whole video (frames=N), or at exact moments (frameAt). The first frames of a video take ~10–20 s (stream lookup); keep frame-heavy calls to a few videos.

• Finding a topic inside long videos: pass query instead of reading whole transcripts.

• Long transcript cut off? Call again with start = the second where it stopped.

List a channel's videos with an outlier score: how many times more views a video got than the channel's median.

Best for: finding a creator's breakout videos (the ones whose hooks/titles/topics worked), reviewing recent uploads, building a list of videos to fetch. Returns: channel stats (subscribers, video count) and videos with views, outlier ×median, date, duration. The median is computed over the videos listed, so list enough (30+) for a fair baseline.

Check that every YouTube path this server uses works from this machine right now (search, video details, transcripts, watch page, channels, yt-dlp and ffmpeg for frames), with timings. Run it when other youtube_* tools fail.