YouTube Transcript Scraper

Extract transcripts from any YouTube video, playlist, channel, or search query — in bulk, in five formats, across 100+ languages.

Built for people who need transcripts reliably: RAG pipelines, content research, competitor monitoring, subtitle workflows, and LLM training data.

What makes it different

• It works when others return nothing. YouTube gates the public web player: caption URLs scraped from the watch page now return HTTP 200 with an empty body. This Actor drives YouTube's mobile app API instead and falls back across three clients, so it keeps returning transcripts where watch-page scrapers quietly fail.
• Bulk input, not one video at a time. Paste a channel URL, a playlist, an @handle, or search:your query and it expands to every video, with paging.
• Translation built in. Pull a transcript in one language and have YouTube translate it to another in the same run.
• Five output formats. Plain text, timed JSON segments, SRT, WebVTT, or timestamped Markdown (ideal for chunking into a vector store).
• You are never charged for a video that fails. Charging happens only after a transcript is actually in hand.

Input

Field	Type	Description
urls	array	Video URLs/IDs, playlist URLs, channel URLs or @handles, and search:<query> terms
languages	array	Preferred caption languages, best first (default ["en"])
fallbackToAnyLanguage	boolean	Use any available track when no preferred language matches (default true)
preferManualCaptions	boolean	Prefer human-written captions over auto-generated (default true)
translateTo	string	Translate captions to this language code via YouTube's translator
outputFormat	string	text, segments, srt, vtt, or markdown
maxVideos	integer	Hard cap on videos processed — protects you from an accidental whole-channel run
aiEnrichments	array	Optional AI summary/key points/chapters — see AI enrichment below
concurrency	integer	Parallel video fetches (default 5)
proxyConfiguration	object	Recommended for large runs — YouTube rate-limits datacenter IPs

Example

{
  "urls": [
    "https://www.youtube.com/watch?v=aircAruvnKk",
    "https://www.youtube.com/@3blue1brown",
    "search:retrieval augmented generation"
  ],
  "languages": ["en"],
  "outputFormat": "markdown",
  "maxVideos": 50
}

Output

One dataset item per video:

{
  "videoId": "aircAruvnKk",
  "title": "But what is a neural network? | Deep learning chapter 1",
  "author": "3Blue1Brown",
  "channelId": "UCYO_jab_esuFRV4b17AJtAw",
  "url": "https://www.youtube.com/watch?v=aircAruvnKk",
  "lengthSeconds": 1120,
  "viewCount": 23914996,
  "thumbnail": "https://i.ytimg.com/vi/aircAruvnKk/sddefault.jpg",
  "language": "en",
  "isAutoGenerated": false,
  "availableLanguages": ["ar", "en", "es", "..."],
  "segmentCount": 286,
  "characterCount": 18145,
  "transcript": "This is a 3. It's sloppily written and rendered at an extremely low resolution...",
  "scrapedAt": "2026-08-16T16:52:20.630Z"
}

Videos that cannot be processed produce an item with videoId, url, and error — so a failed video is visible in your results instead of silently missing.

Notes and limits

• Videos with captions disabled cannot be transcribed. There is no caption track to fetch. The Actor reports this per video rather than failing the run.
• Private, deleted, and age-restricted videos are reported as errors with YouTube's reason.
• Residential proxy is required, and it is the default. YouTube blocks datacenter IP ranges outright with "Sign in to confirm you're not a bot" — measured on Apify's own servers, 4 of 5 videos failed without one and 4 of 4 succeeded with one. Leave proxyConfiguration on its residential default unless you are running from an IP you know YouTube trusts.

AI enrichment (optional, off by default)

The Actor can also generate a summary, key points and timestamped chapters from each transcript. This requires the Actor operator to configure an LLM credential; if none is set, runs return transcripts only and log a warning — they never fail, and you are never charged for an enrichment that did not happen.

Leave aiEnrichments empty unless the listing states AI is enabled.

Configuration (Actor operator)

AI enrichment reaches a model one of two ways. Whichever credential is present wins, preferring OpenRouter because it needs no provider account.

Variable	Purpose
APIFY_TOKEN	Default route. Goes through Apify's OpenRouter Actor — no second account, tokens billed as Apify platform usage. Set automatically in every run, but needs a paid Apify plan (Free returns HTTP 429).
OPENROUTER_API_KEY	Optional override — calls openrouter.ai directly, billed to an OpenRouter balance. Takes precedence if set.
ANTHROPIC_API_KEY	Optional override — calls the Anthropic API with your own key.
LLM_PROVIDER	Force openrouter or anthropic instead of auto-detecting.
LLM_MODEL	Model id. Defaults to deepseek/deepseek-v4-flash (OpenRouter) or claude-haiku-4-5 (Anthropic).
LLM_TIMEOUT_MS	Request timeout, default 180000.

Credential precedence is OPENROUTER_API_KEY → APIFY_TOKEN → ANTHROPIC_API_KEY. With none of them available the Actor logs a warning and returns transcripts only — it never fails a run. See PRICING.md for model costs and the per-character charging model.

Local development

npm install
npm run build
node dist/main.js     # reads storage/key_value_stores/default/INPUT.json