YouTube Comments Scraper — Innertube API, No 10K-Quota, No OAuth

Get the full comment stream from any YouTube video into CSV / JSON without an API key. The actor calls YouTube's internal Innertube endpoint (the same one youtube.com uses on every page load) instead of the YouTube Data API v3 — so there is no 10,000-unit/day quota to budget, no OAuth console to register, and no key to rotate.

19 production runs to date by sentiment-analysis, creator-economy, and NLP-dataset teams.

---

Why Innertube over the Data API

Factor	YouTube Data API v3	Innertube (this actor)
API key	Required	Not needed
Daily quota	10,000 units (≈ 5–8 fully-paginated videos)	No published quota
Setup	Google Cloud project + OAuth	Paste video URLs
Comment payload	commentRenderer (older)	commentEntityPayload (current)
Reply count + verified flag	Available, costs quota	Available, no quota

Honest disclosure on "no quota": YouTube can — and does — rate-limit Innertube traffic when it looks abusive. The actor includes random 0.8–2 s delays between paginated requests and 1.5–3 s between videos to stay under that radar. At default settings it has run cleanly across 19 jobs; bursting harder is your call.

---

Output schema (per comment, verified against src/main.js)

{
  "videoId": "dQw4w9WgXcQ",
  "commentId": "Ugzge340dBgB75hWBm54AaABAg",
  "author": "@YouTube",
  "authorChannelId": "UCBR8-60-B28hp2BmDPdntcQ",
  "text": "can confirm: he never gave us up",
  "likes": 200,
  "publishedAt": "10 months ago",
  "isReply": false,
  "replyCount": 961,
  "isHearted": false,
  "isVerified": true,
  "scrapedAt": "2026-04-29T12:00:00.000Z"
}

12 fields per comment. publishedAt is the relative string YouTube exposes (e.g. "10 months ago") — there is no absolute timestamp on Innertube's public-page response, so we don't pretend to derive one.

isHearted reflects creator-heart state. isVerified is the channel-verified flag on the comment author. authorChannelId is the YouTube channel ID of the commenter (useful for joining against youtube-channel-scraper).

---

Input

Parameter	Type	Default	Range	Description
videoUrls	string[]	required	≥1	YouTube video URLs or 11-character video IDs
maxCommentsPerVideo	integer	100	1–100000	Top-level cap per video (replies flagged, NOT recursed)
sortBy	string	"top"	"top" | "newest"	"newest" is best-effort (see Honest limitations)

videoUrls accepts:

• https://www.youtube.com/watch?v=ID
• https://youtu.be/ID
• https://www.youtube.com/shorts/ID
• Bare 11-char video IDs (regex ^[a-zA-Z0-9_-]{11}$)

---

How it works

1. Fetch the watch page — https://www.youtube.com/watch?v=<videoId> with a hardcoded Chrome 120 User-Agent. Extract ytInitialData and INNERTUBE_API_KEY from inline <script>.
2. Find the comments continuation token inside twoColumnWatchNextResults.
3. POST to youtubei/v1/next with INNERTUBE_CONTEXT (clientName=WEB) and the continuation token.
4. Optionally re-fetch with the "Newest first" sort token if sortBy="newest" and the sort menu surfaces in the response.
5. Walk frameworkUpdates.entityBatchUpdate.mutations — each commentEntityPayload becomes one record. Continue paginating via onResponseReceivedEndpoints until the cap is hit or the continuation chain ends.

Pure fetch(), no Cheerio, no Playwright, no browser automation.

---

Honest limitations (read before bulk runs)

• maxCommentsPerVideo=0 returns ZERO comments, NOT unlimited. The pagination loop guard is allComments.length < maxComments, so 0 < 0 is false → loop never enters → empty output. The schema now enforces minimum: 1. For "as much as possible," set a high explicit number (e.g. 100000); the run will end naturally when the continuation chain ends or the cap is hit.
• Replies are flagged but NOT recursed. Each record exposes isReply (bool) and replyCount (int). The actor never POSTs into reply-panel continuation tokens. If you need full reply threads, see Custom builds below.
• sortBy="newest" is best-effort. The actor re-fetches with the "Newest first" token only if findSortToken() finds it in the response. If the comments-header sort menu is missing for a given video (rare but possible), the actor silently falls back to default "top" sort. Check the run log to confirm; the line Switching to "Newest first" sort... indicates success.
• Hardcoded User-Agent (Chrome 120). Both the watch-page GET and the Innertube POST send the same UA string. If YouTube fingerprints UA + IP and that combination triggers a soft-block, this actor cannot rotate. Custom build can wire in UA rotation and proxy.
• No proxy. Direct fetch from the Apify worker IP. Bursting many videos in parallel from the same IP raises rate-limit risk.
• Hardcoded fallback Innertube API key. If the regex "INNERTUBE_API_KEY":"([^"]+)" fails to match the watch page (YouTube response shape change), the actor falls back to the well-known public web-embed key (AIzaSyAO_FJ2SlqU8Q...). This usually works but is not guaranteed long-term.
• Comments disabled / video unavailable / age-gated — actor logs No comments section found. or Video not available and skips that video; no record pushed.
• No absolute timestamps. publishedAt is the relative string YouTube returns ("10 months ago"). We do NOT infer absolute timestamps from a relative string — that would be a fabrication. If you need ISO 8601 timestamps, use the Data API v3 (at the cost of quota).
• Continuation chain ends naturally — for popular videos with 100K+ comments, YouTube eventually stops returning a next-token. Practical hard cap is whatever YouTube serves you for that specific video.
• Inter-video delay 1.5–3 s, intra-pagination delay 0.8–2 s. Adjust expectations: 100 videos × 100 comments ≈ 5–10 min wall-clock, even though each fetch is sub-second.

---

Python integration

from apify_client import ApifyClient

client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("knotless_cadence/youtube-comments-scraper").call(
    run_input={
        "videoUrls": [
            "https://www.youtube.com/watch?v=dQw4w9WgXcQ",
        ],
        "maxCommentsPerVideo": 500,
        "sortBy": "newest",
    }
)
for c in client.dataset(run["defaultDatasetId"]).iterate_items():
    flag = " ✓" if c["isVerified"] else ""
    print(f"{c['author']}{flag}: {c['text'][:120]}  ({c['likes']} likes)")

---

Common questions

Q: Will this stop working if YouTube changes their internal API? A: Innertube is YouTube's core internal API — they can't remove it without breaking youtube.com itself. The actor parses the commentEntityPayload schema, which has been stable on the WEB client. If a field is renamed, the actor returns partial records (the ||-fallback chains across likeCountLiked / likeCountNotliked, replyCount, etc. mean a rename on one path doesn't crash extraction — it degrades gracefully).

Q: Can I get full reply threads? A: Currently the actor records isReply (bool) and replyCount (int). It does not recurse into reply panels. Full-reply-thread extraction is a custom build — see Custom scraping below.

Q: Why does publishedAt look like "10 months ago" instead of an ISO date? A: That's the literal value YouTube returns in the WEB-client Innertube response. We don't infer absolute timestamps from a relative string — that would be a fabrication. If you need exact timestamps, you'll need the Data API v3 (which exposes publishedAt as ISO 8601, at the cost of quota).

Q: What if a video has comments disabled? A: The actor logs "No comments section found. Comments may be disabled." and continues to the next video. No record is pushed for that video.

Q: Cost on Apify? A: Apify charges by compute units, not per-comment. Each fetch is sub-second over plain fetch(); pagination delays dominate runtime. Free tier covers small batches. For sustained large pulls, email and we can quote a fixed-price custom build.

---

Use cases

• Sentiment analysis on creator audience.
• AI / NLP training-data collection (real conversational text).
• Competitor monitoring on rival channels.
• Influencer engagement-quality audit (verified-vs-unverified ratio, reply density).
• Brand-mention surveillance across creator videos.
• Academic research into platform discourse.

---

YouTube tool family (related actors)

• YouTube Channel Scraper
• YouTube Search Scraper
• YouTube Transcript Scraper

Cross-platform pairing:

Source	Actor	Use
Trustpilot	Customer reviews	External brand sentiment (949 production runs)
Reddit Discussion	Community discussions	Unfiltered audience opinion
Bluesky Scraper	AT Protocol	Open social-graph mentions
Hacker News	Tech sentiment	Developer-audience reactions

All 31 published actors free to inspect on Apify Store.

---

Custom scraping — pilot tiers

Need full reply threads, multi-video pipelines, UA rotation + proxy, or a different schema (translation, profanity flags, sentiment scoring)? Three tiers:

• Pilot — $97 · 1 actor, basic config, 7-day support. Good entry point — useful for a single creator-comment dump or a one-off competitor channel audit.
• Standard — $297 · custom actor + Slack/email alerts on results, 30-day support. Most NLP-dataset and creator-economy projects fit here.
• Premium — $797 · custom actor + dashboard + 90-day support + 1 modification round. For ongoing pipelines (daily creator-channel sentiment, multi-platform fan-out with TikTok / Instagram / Threads), full reply-thread recursion, proxy-routed UA rotation.

Email: spinov001@gmail.com — drop video URLs and the schema you need; quote within 48h.

Proof of work: 31 published Apify scrapers (78 total in portfolio) — Trustpilot 949 runs, Reddit 80+, Google News 43, Glassdoor 37, Email Extractor 36+. Recently delivered a paid 3-article series for a client in the proxy industry ($150).

More tips: t.me/scraping_ai · blog.spinov.online

---

Disclaimer

Designed for sentiment analysis, audience research, and academic use. Respect YouTube's Terms of Service, applicable data-protection law (GDPR, CCPA), and scrape publicly visible content only. Not affiliated with YouTube or Google LLC.

Honest disclosure: extracts 12 fields per comment (videoId, commentId, author, authorChannelId, text, likes, publishedAt, isReply, replyCount, isHearted, isVerified, scrapedAt). publishedAt is the relative string YouTube returns ("10 months ago") — no inferred absolute timestamps. Reply threads are flagged but NOT recursed. maxCommentsPerVideo=0 returns 0 comments (use a high integer for "all"). Hardcoded Chrome 120 UA + fallback Innertube API key on watch-page extraction failure. Innertube has no published quota but YouTube can rate-limit aggressive callers; the actor's default 0.8–2 s pagination delay keeps runs polite.