Pulling YouTube Shorts transcripts into n8n — the 3 traps that made it hard

I’ve been building a small pipeline around YouTube Shorts transcripts and hit three traps that I think are worth sharing, because most tools out there get them wrong:

  1. The transcript you get is usually a machine translation. In the player response for one English Short there were 21 caption tracks: the video’s own language plus 20 auto-translations. Most tools pick by language preference, so you feed an English video into your workflow and the text comes back in Spanish or German. It’s marked in a field that isn’t the obvious one.

  2. Downloading the whole watch page costs ~1.9 MB per video. There’s a lighter path (mobile player endpoint) that returns the same captions at ~340 KB. If you’re paying for residential proxies per GB, that difference is the difference between profitable and not.

  3. One IP gets rate-limited after roughly 30-40 caption requests, and YouTube then asks you to “sign in to confirm you’re not a bot”. Rotating the proxy on every request and retrying took my success rate from 85% to ~95%.

End result: 38/40 Shorts on an English channel and 12/12 on a Spanish one, with the text in the original language.

I packaged it as an Apify Actor so it runs outside n8n and returns clean JSON I can drop straight into my workflow (one record per Short: transcript with timestamps, title, channel, duration, language, plus a plain-text version). It’s pay per transcript delivered, $5 per 1,000, and Shorts without captions are reported and not charged: YouTube Shorts Transcripts — Captions & Subtitles Scraper · Apify

Happy to answer anything about the rate-limiting part — that was the part that took me longest.

— Alejandro Salamanca, SICTEG LTDA (Colombia)

2 Likes

Fetch the original-language track first. Translations listed in the track list often can’t be fetched at all, a September 2026 test pulled four videos that all listed translations and got zero of four back. Every retry on a translation-only video burns one of your 30 to 40 requests per IP on a track that was never going to arrive, so skip anything that only offers a translation.

The 30 to 40 per IP is the same line the library hits from a home connection, IpBlocked after roughly 30 fast fetches in the same test. Rotating per request keeps every IP under it. If you don’t want to run proxies, a hosted API does the fetch on its side. I use youtubetranscriptapi.com for this, its MCP server works directly in n8n with 7 read-only tools, no HTTP node needed. 100 free credits, no card, errors never charged.

In n8n, put a Wait node between small batches and dedupe on video ID before the loop, so a failed video doesn’t cost two slots out of the 30.