Download YouTube Transcripts — Every Video from a Channel

Get the transcript of a YouTube video, or of every video a channel has ever published, as a spreadsheet. Full text, the language it was written in, and timed segments if you want them. No API key, no quota, nothing to install.

Paste a channel, a playlist, a search term or a single video URL. Transcripts, view counts, publish dates, likes and comments all come back from the same run, so one job gives you the transcript and everything around it.

{ "inputs": ["https://www.youtube.com/@veritasium"],
  "includeTranscript": true, "transcriptLanguages": ["en"] }

Videos with subtitles switched off come back as NO_CAPTIONS, not blank — and you are not charged for them.

Callable from any MCP client. Claude, Cursor, VS Code and anything else that speaks Model Context Protocol can find and run this through Apify's MCP server at mcp.apify.com — search-actors to discover it, call-actor to run it. No setup on your side and no separate integration.

Built for pipelines that run more than once

Most scrapers charge you the same on Monday as they did on Sunday. This one remembers what it already fetched.

Measured on 27 August 2026 — one channel, 500 videos, same input both times:

	first run	run it again
rows delivered	500	500
videos charged	489	11
requests to YouTube	525	30
time	174s	46s

Videos already fetched inside the cache window come back from stored state and are not charged. Eleven were charged the second time because those eleven had failed the first — you are not billed for a row that came back empty, and not billed again for one you already have.

And it is one Actor, not three. Video metadata, transcripts and comments arrive from the same run, in the same dataset, on one bill. No joining three outputs on video_id and no three integrations to keep alive.

What you get back

video_type	title	views	published_at	duration_s
video	Google Pixel 11/Pro/Fold Impressions: It Is What It Is	3138780	2026-08-12T07:00:35-07:00	673
short	This ZOOM is Insane!	3567393	2026-07-20T13:39:08-07:00	66
stream	5 New Phone Updates + Giveaway Update!	940608	2017-11-29T09:15:12-08:00	2501

Real rows from a real run. Exact counts, not 3.1M. Full timestamps, not 7 days ago. Both survive sorting.

---

How do I download the transcript of every video on a channel?

{ "inputs": ["https://www.youtube.com/@veritasium"],
  "includeTranscript": true, "transcriptLanguages": ["en", "es"] }

Each row gains the transcript as text, its language, whether it was auto-generated, and how many timed cues it has.

Timed cues are a separate switch: includeTranscriptSegments. Turn it on and each row also carries transcript_segments — one entry per cue with start_s, duration_s and text — for citing a moment in the video. It costs nothing extra, because the cues arrive in the same request as the text. It is off by default because every export format flattens them: a 1,000-cue video becomes transcript_segments/0/text … /999/text, and a spreadsheet of a few videos turns into thousands of columns. transcript_cue_count ships either way.

transcriptLanguages is a preference order and human-written subtitles beat auto-generated ones. Subtitles off returns NO_CAPTIONS, free.

includeLikes adds a like count from one small extra request, not charged separately.

---

What a row actually looks like

One video, with transcripts, timed segments and comments all switched on. Real output from a real run — long fields trimmed with …, nothing else changed.

{
  "_row_type": "video",
  "video_id": "o4SSoURPODY",
  "video_type": "video",
  "title": "Google Pixel 11/Pro/Fold Impressions: It Is What It Is",
  "channel": "Marques Brownlee",
  "channel_id": "UCBJycsmduvYEL83R_U4JriQ",
  "views": 3148445,
  "likes": 84617,
  "published_at": "2026-08-12T07:00:35-07:00",
  "category": "Science & Technology",
  "duration_s": 673,
  "description": "Every year, a new Pixel, and new hopes and dreams... …",
  "keywords": ["Pixel 11", "Pixel 11 Pro", "Pixel 11 Pro Fold", "…"],
  "thumbnail": "https://i.ytimg.com/vi_webp/o4SSoURPODY/maxresdefault.webp",
  "subscriber_count": "21.1M subscribers",
  "is_live": false,
  "is_private": false,
  "is_crawlable": true,
  "transcript": "[music] >> You know, I've been using Pixel phones for a long time. …",
  "transcript_segments": [
    { "start_s": 2.619, "duration_s": 2.741, "text": "[music]" }, "…"
  ],
  "transcript_cue_count": 331,
  "transcript_language": "en",
  "transcript_is_generated": true,
  "transcript_status": "OK",
  "_source_url": "https://www.youtube.com/watch?v=o4SSoURPODY",
  "_cached": false,
  "_error_kind": null
}

A comment from the same run — its own row, not a nested field:

{
  "_row_type": "comment",
  "video_id": "o4SSoURPODY",
  "comment_id": "Ugz7VRjqhsV-60jGjwR4AaABAg",
  "comment_text": "Adding a physical RGB LED to the back of the phone just to …",
  "comment_author": "@Mosesplusofficial",
  "comment_author_channel_id": "UCr3n8BfGuDIzhO8ALIxMe4w",
  "comment_author_is_verified": false,
  "comment_published_text": "8 days ago",
  "comment_like_count_text": "13K",
  "comment_reply_count_text": "200",
  "comment_is_reply": false
}

---

Three things that make this different

🟢 Failures are typed, never blank

Subtitles off returns NO_CAPTIONS. A broken fetch returns FETCH_FAILED. Different values, so you always know which happened.

🟢 Transcripts and comments in the same run

One actor, one pass. Transcript text, timed segments, comments and replies — alongside the metadata, not in a separate job.

🟢 Failed, filtered and cached rows cost nothing

Charged only when a row has data. A deleted video is free. So is one your filter removed, or one served from cache.

---

How do I scrape all videos from a YouTube channel?

{ "inputs": ["https://www.youtube.com/@mkbhd"], "maxResultsPerSource": 500 }

@handle, a channel URL, or a /channel/UC… ID all work.

You get all three tabs. Long videos, Shorts and live streams sit in separate tabs that do not overlap — one channel measured 30, 48 and 5. video_type says which is which, and the tabs are interleaved so asking for 30 returns a mix. channelTabs narrows it.

Channel rows also carry subscriber_count and channel_description, free — same request that lists the videos.

What happens when something fails?

You get a typed reason in the row, never an empty cell.

_error_kind	means
REJECTED_INPUT	not a recognised channel, playlist, video or search
RESOLVE_FAILED	the listing could not be read
NO_VIDEOS	it was read and contained none
FETCH_FAILED	the video request failed
NO_DATA	it loaded but carried no video details

Blocked requests retry automatically on a fresh IP. In a 36-video run that turned 3 blocks into 0 failures.

How do I export a YouTube playlist to CSV?

{ "inputs": ["https://www.youtube.com/playlist?list=PLbpi6ZahtOH6Blw3RGYpWkSByi_T7Rygb"],
  "maxResultsPerSource": 100 }

Then hit Export — CSV, JSON, Excel, or an API endpoint. Playlists return 100 videos per request, so multiples of 100 go furthest.

How do I get YouTube comments with replies?

{ "inputs": ["https://www.youtube.com/@mkbhd"], "includeComments": true,
  "maxCommentsPerVideo": 100, "maxRepliesPerComment": 5 }

Comments arrive as their own rows, not extra columns — _row_type is video or comment. Folding 100 comments into a video row would export as 600 unusable columns.

Replies are rows too (comment_is_reply: true), capped because each comment with replies costs an extra request.

How do I search YouTube and export the results?

{ "inputs": ["drone review", "sourdough starter"] }

One quirk: a bare 11-character word is read as a video ID, because that is exactly how long a YouTube ID is — helicopters is eleven characters. Write search:helicopters to force a search.

Can I filter, or work in another language?

input	effect
publishedAfter / publishedBefore	YYYY-MM-DD, inclusive
durationMinSeconds	set 61 to exclude Shorts
durationMaxSeconds	upper bound in seconds
language	en, de, ja, es — text YouTube generates
region	US, DE, GB — how results are ranked

Filters run after each video is read, so they are exact — and filtered-out videos are never charged.

subscriber_count follows the language as YouTube wrote it: 21.1M subscribers, 21,1 Mio. Abonnenten, チャンネル登録者数 2110万人. Titles, descriptions and category are unaffected.

What people use this for

📈 Track a competitor's channel every week

1. Run with the competitor's channel and maxResultsPerSource: 50, maxAgeHours: 0 so view counts are fresh.
2. Schedule it weekly in the Apify Console.
3. Join runs on video_id and diff views to get per-video growth, and published_at to get posting cadence.

Set maxAgeHours: 0 here — the cache is a time window, and stale view counts would flatten exactly the trend you are measuring.

🔔 Watch a channel and get only what is new

1. Run with the channel and publishedWithinDays: 7.
2. Schedule it weekly in the Apify Console. That is the whole setup.

publishedWithinDays is a rolling window, counted from the day each run starts — so a saved task keeps returning what is new instead of the same seven days forever. publishedAfter is an absolute date and is the wrong tool for a schedule: set once, it returns the same videos every week until someone edits it. If you set both, the explicit date wins and the run log says which applied.

This is the cheapest thing this Actor does. Everything it has already fetched comes back from stored state and is not charged, and anything outside the window is filtered and free. Measured on a 500-video channel: the first run charged 489 videos, the second charged 11.

Use publishedWithinDays a little wider than your schedule — 8 or 9 days for a weekly run — so a late-publishing channel or a missed run does not leave a hole.

How this differs from a YouTube RSS feed. Every channel has a free RSS feed at youtube.com/feeds/videos.xml?channel_id=…, and if all you need is a title and a link, use it — it is free and instant. It gives you the most recent 15 videos and nothing else: no transcript, no view count, no duration, no description, no comments, and no way to reach further back. This Actor returns new videos with the full record, transcripts included, for any number of channels at once.

🤖 Build an LLM corpus from a channel's transcripts

1. Run with the channel, includeTranscript: true, maxResultsPerSource: 500.
2. Export as JSON and use transcript for plain text. Add includeTranscriptSegments: true when you need timestamps to cite back to a moment in the video — it is free, but it makes the rows much wider.
3. Filter on transcript_status == "OK" — NO_CAPTIONS means that video never had subtitles, and it cost you nothing.

transcript_is_generated tells you whether a human wrote it. Auto-generated captions carry more errors, which matters if you are grounding an answer on them.

💬 Find what an audience actually asks about

1. Run with the channel, includeComments: true, maxCommentsPerVideo: 100, and maxRepliesPerComment: 5 to catch the threads under the popular ones.
2. Filter rows to _row_type == "comment".
3. Group by video_id, sort by comment_like_count_text, read the top of each.

Useful for support-content planning, sponsorship research, and finding the question a whole audience keeps asking that nobody has answered yet.

What fields do I get?

field	example
video_id · video_type	dQw4w9WgXcQ · video, short, stream
title · channel · channel_id	Rick Astley - Never Gonna Give You Up · Rick Astley · UCuAXFkgsw1L7xaCfnd5JJOw
views · likes	1806183998 · 136841
published_at · duration_s	2009-10-24T23:57:33-07:00 · 213
category · keywords · thumbnail · description	Music · array · highest-res URL · full text
subscriber_count · channel_description	from channel inputs
is_live · is_private · is_crawlable	booleans
every row	_row_type _input _source_url _site _cached _cached_at _error_kind _error
with transcripts	transcript transcript_cue_count transcript_language transcript_is_generated transcript_status
with includeTranscriptSegments	transcript_segments — one entry per cue, start_s duration_s text
comment rows	comment_id comment_text comment_author comment_author_channel_id comment_author_is_verified comment_author_is_creator comment_published_text comment_like_count_text comment_reply_count_text comment_is_reply

Three comment fields end in _text because YouTube publishes them rounded and relative — 13K, 7 days ago — and they pass through as given. 13K is not expanded to 13000: it could be 13,499, and inventing digits hands you a wrong number that looks precise.

Not included: no sort order, no video or audio downloads, no total comment count per video.

Does it re-fetch videos I already pulled?

Any video fetched in the last 24 hours comes back from stored state — and is not charged. Marked _cached: true with _cached_at.

This is a time window, not change detection. It checks how long ago a video was read, not whether it changed — so set maxAgeHours to 0 when sorting on views or measuring growth. The window is per actor: yesterday's run counts for today's.

Cached rows still follow this run's settings. Turn transcripts on and a cached video gets its transcript fetched and charged; turn them off and the row carries no transcript fields at all, whatever an earlier run stored. The cache never adds a field you did not ask for, and never withholds one you did.

What does it cost?

you get	you pay
a video row	$0.50 / 1,000
a transcript	$10.00 / 1,000
a comment or reply	$0.20 / 1,000
a failed, filtered or cached row	nothing

Transcripts and comments are billed separately — each is a separate request to YouTube, not an extra field.

Send everything in one run, not one run each: fixed per-run cost is ~86% of a single-video run and ~6% at a hundred. Requests run 8 at a time, one per second each; concurrency adjusts that between 1 and 16.

Will it quietly break?

Every run emits a schema fingerprint — which fields resolved, which came back empty, the row count — compared against a stored baseline. A field that starts returning empty is caught as drift, not discovered months later.

Scrapers rarely break loudly. They thin out.

Frequently asked questions

How do I get a transcript from a YouTube video?

Paste the video URL into inputs and set includeTranscript to true. You get the full transcript as text in the transcript field, along with its language and whether YouTube auto-generated it. No API key and no browser extension.

{ "inputs": ["https://www.youtube.com/watch?v=dQw4w9WgXcQ"],
  "includeTranscript": true }

How do I get the script of a YouTube video?

The same way — what YouTube calls captions or subtitles is what most people mean by the script. transcript holds it as one block of text, ready to paste into a document or feed to a model.

How do I download YouTube transcripts for a whole channel at once?

Give it the channel instead of the video. inputs accepts an @handle, a channel URL or a /channel/UC… ID, and maxResultsPerSource decides how many videos deep to go. Every video comes back as a row with its own transcript.

Can I get the transcript with timestamps?

Yes. Turn on includeTranscriptSegments and each row also carries transcript_segments — one entry per cue with start_s, duration_s and text — so you can cite the exact moment something was said. It costs nothing extra. It is off by default because spreadsheets flatten it into thousands of columns; transcript_cue_count ships either way.

What happens if a video has no subtitles?

transcript_status comes back as NO_CAPTIONS and you are not charged for that video's transcript. It is a different value from FETCH_FAILED, so you can always tell "this video has no subtitles" apart from "we could not reach it" — and re-run only the second kind.

Do I need the YouTube Data API or an API key?

No. There is no key to obtain, no project to create, and no daily quota to run out of. That is the main reason people use this instead of the official API.

Is there a free way to do this?

Not here — this Actor is paid per transcript. What it does give you is that failed, filtered and cached rows cost nothing, so you only pay for transcripts you actually receive, and never twice for the same video inside the cache window.

Can I get transcripts in another language?

transcriptLanguages is a preference order, for example ["en", "es"]. Human-written subtitles are preferred over auto-generated ones when both exist. transcript_language and transcript_is_generated tell you which you got.

Can I use this from Claude, Cursor or another AI assistant?

Yes. It is exposed through Apify's MCP server at mcp.apify.com, so any client that speaks Model Context Protocol can discover it with search-actors and run it with call-actor under the name aquixlabs/youtube-channel-scraper. Nothing to install and nothing to configure — every public Apify Actor is reachable this way, and this one is not in either excluded category (it is LIMITED_PERMISSIONS, and pay-per-event rather than rental).

A useful shape for an agent: ask it for the transcript of a video or a channel, and it can call this directly rather than scraping the page itself.

Can I get the comments as well as the transcript?

Yes, in the same run. Set includeComments and each comment arrives as its own row — not a column — tagged _row_type: comment. The dataset's Comments view is where to read them.

Terms

You are responsible for your use of the data you collect, including under YouTube's Terms of Service and any applicable data protection law.

---

More from aquixlabs

Other actors from this account will be listed here.