List YouTube comment replies

Where to start. Three values work: a top-level comment id (e.g. the comment.id of any row from /v1/youtube/video/comments), that row's comment.ext.replies_token, or the pagination.next_cursor from a previous replies response. A token that is none of these is rejected.

Text format for reply bodies: 'html' (default) or 'plainText'.

Optional CSV of SocialCrawl labels to add to every comment: sentiment, spam, question, purchase_intent, complaint, toxic, low_quality, injection. Without this param every page already carries sentiment, question, purchase_intent and complaint, free; label= adds the labels you name to them (the defaults keep running). Each comment gains computed.labels with probabilities (0 to 1), so you choose the cut-off; a comment that could not be judged carries labels: null. sentiment, question, purchase_intent and complaint are free when asked for too; spam, toxic, low_quality and injection add 1 credit per started 25 newly judged comments (comments already labelled for anyone are free, and so is a cached page). injection flags text that addresses an AI system and tries to direct it (flagged, p); it never drops or rewrites a row. judgments=off (or label=none) turns the default labels off. data.labels reports what was judged and billed.

Optional CSV (spam, low_quality), only with the same preset in label. Drops comments whose probability is 0.8 or higher and lists their ids in data.labels.dropped_ids. A comment that could not be judged is never dropped.

Optional, on (the default) or off. By default every row gains free SocialCrawl judgments (computed.labels, and computed.relevance on search endpoints), reported in data.labels (mode default) and data.relevance (origin default), each with a status (complete, partial or skipped) and pending: the rows still being judged when the page was sent, which carry null now and are filled on your next call or cached read. Default judgments never add credits, never change an existing field, and never drop or reorder a row. off returns the page exactly as before, with none of those keys. label=none does the same.

Optional. When 1, return a cost preview for this labelled or relevance-filtered request without fetching the page or judging any row. data.estimate reports rows_expected, rows_cached, label_credits_min, label_credits_max and base_credits. 0 credits charged.

Optional, only with label=. When 1, every labelled row also carries computed.labels_evidence.<preset> = { quote, sentence_index }: the sentence in the row that most clearly shows the label, copied verbatim. Absent or null when no single sentence shows it.

Optional. When goal, keep the rows and fields needed for the goal you pass in goal= (plus any that are uncertain, and the first and last), and replace the rest with a stub. data.held_back lists the held ids and a recall id that re-reads the full page from cache at no extra charge. Without this param the page is unchanged.

Required by fit=goal. What you are trying to do, in your own words, up to 300 characters.

Optional, only with fit=goal. Soft cap on how much of the page to keep, in tokens. Uncertain blocks and the first and last block are kept even if they exceed it.

Only rows published on or after this date: YYYY-MM-DD (midnight UTC) or an ISO 8601 timestamp. Rows older than it are left off each page. This list is not in date order, so the walk does not stop at the date: keep following next_cursor, and each page costs what a page costs even when every row on it is filtered out. A row with no date is kept. published_after, start_date, from_date, publish_time (yesterday, this-week, this-month, last-3-months, last-6-months) and max_age_days are accepted and applied as since.

Only rows published on or before this date: YYYY-MM-DD (that whole UTC day is included) or an ISO 8601 timestamp. Newer rows are left off each page and the walk does not stop, so older pages follow. A row with no date is kept. published_before, end_date and to_date are accepted and applied as until.

Universal pagination cursor. Send pagination.next_cursor from the previous response back verbatim: the API maps it to this endpoint's native continuationToken (cursor style). You never construct, decode, or look up a cursor. Omit it for page 1.

Send no-cache to bypass the response cache and force a live fetch. Billed at the normal endpoint cost; the fresh result is written back to cache for the next caller. Only the no-cache directive triggers this. See the Response Schema guide for details.

Optional UUID that makes the request safely retriable. The key is reserved atomically per account before billing. A concurrent retry receives 409 IDEMPOTENCY_IN_PROGRESS with a Retry-After of 1 second while the original is young, growing to a quarter of the time it has been running (at most 30); a crashed reservation expires after 5 minutes. A replay keeps the cached payload immutable except for billing metadata: credits_used becomes 0, idempotent_replay becomes true, and credits_remaining is refreshed to the current balance. A known current balance appears in both the body and X-Credits-Remaining header; no balance row resolves to 0. On a transient lookup failure, body credits_remaining is null and X-Credits-Remaining is omitted. Completed outcomes expire after 24 hours. If the claim store is unavailable, the request fails closed before billing. GET endpoints do not combine idempotent replay with Server-Sent Events: use Accept: application/json when sending this header. The unsupported key-plus-SSE combination is rejected before billing.