YouTube API

Endpoints

Quick start

Coverage & limits

• Four concurrent requests per worker; saturation returns 503 busy. One page or record per request: list operations return one page (30 videos, 48 Shorts, 30 playlists, 30 streams, 10 posts, 20 comments, one search page) and YouTube's continuation token; the caller carries tokens, nothing is stored.
• Base operations (channel, playlists, lives, posts, typeahead, hashtag, comments, replies, community post): 20-second deadline (video, transcript, sponsors and playlist reads: 30 seconds), at most four upstream HTTP attempts (handle resolution, the channel browse, the About panel or sort continuation, and one transient network/5xx retry), 8 MiB decoded bytes (playlist reads follow up to ten continuations under a 12-attempt, 16 MiB budget; sponsors read the page, the Android player and one caption track under five attempts).
• Enrichment operations (channel Shorts, includeExtras on channel-videos and search, trending Shorts): two small public API calls per video (player metadata and watch-next data, issued together; one with extrasDetail=player), 12 videos at a time, at most 48 videos, 128 upstream attempts, 32 MiB decoded bytes and a 60-second deadline. Videos whose details could not be read keep null extras (extras_incomplete) instead of failing the request; an exhausted budget ends the batch early and batches stop starting reads eight seconds before the deadline. Trending Shorts read the feed page plus at most six sequence continuations while enrichment already runs.
• Every parsed document is bounded to depth 96 and 1,000,000 nodes; oversized or malformed responses fail explicitly (response_too_large, payload_too_complex, unsupported_page). A redirect to Google's "sorry" interstitial (an over-used address) is reported as 429 upstream_rate_limited.
• Logged-out reads of the fixed www.youtube.com origin (pages, the public navigation resolver, browse, next, search, player and reel sequence routes and timed text) and the public suggest origin only: no cookies, consent bypass, account session, API key or YouTube Data API quota. Thumbnails, avatars, download and caption URLs are returned as data and never fetched (the transcript operation reads exactly one caption track).
• Relative publish times ("6 days ago", "2w ago") are converted to timestamps at retrieval time; texts with a prefix ("Streamed 2 days ago", "Scheduled for …") keep a null timestamp. Displayed counts reflect the moment of retrieval.
• Channel-videos extras are read with includeExtras=true; transcript text is whitespace-normalised; sponsor inference is a heuristic with evidence and confidence on every suspected sponsor.
• No result cache, billing, shared database, or consuming-product dependency.