Summarize a YouTube channel with AI and cite the sources

YouTube Video Transcript lets your chat read bulk transcripts to build a source-backed brief on what an expert has said about a topic. When the answer is spread across months of uploads, you need more than a summary of the latest video. Bring the relevant statements together with quotes and links to their original moments, so you can examine the evidence without reopening every transcript by hand.

The server is at https://api.youtubevideotranscript.io/mcp. A channel job with comments off reserves one credit per video, on a paid plan from the chat or within your free 30 credits every 30 days from the web app. Reading saved results adds no service credits. Below is a recorded read of a 30-video job, followed by a brief structure and prompt you can use with your own topic. YouTube Video Transcript supplies the transcripts; the connected assistant writes the analysis.

537 transcripts have been delivered through the YouTube Video Transcript MCP server, across 51 accounts. Figures as of September 2026.

What has this expert said about your topic?

Ask one question about a named person and a defined set of videos, then request evidence for the answer. “What has this expert said about [topic], and can you give me sources?” becomes useful when the assistant knows which channel or job to read and what each citation must contain.

Choose the topic, the videos or period you want covered, and whether you need a current position or a comparison over time. Search within the channel or inspect its playlists to choose a relevant collection. Search results help with selection, but they are not proof that no other video discusses the topic. A job on one channel also does not cover the expert's guest appearances elsewhere.

Collect the transcripts once and let the chat read those saved results. Report the number actually read, any failures, and any parts of the intended scope left out. That gives the brief a defined evidence base rather than implying the assistant has reviewed everything the person has ever said.

How does the chat get every transcript?

Through a job it can create or find, then read page by page. Four tools carry the flow, and only creating the job spends credits. If the collection already exists, start with finding it rather than paying to fetch the videos again.

The free-plan route is the same flow with a different first step. Paste the channel into the web app and check how many videos the button will collect within your remaining balance and the website's 5,000-video cap; then, in the chat, list_jobs finds it and list_job_videos reads it. Use the same Google account for the website and the connector. The channel download post covers that button and the ZIP it produces.

Run: what Auto Focus said about EVs over nine months

Seventeen free calls on 6 September 2026 read a 30-video channel job in full and answered “what has this reviewer said about electric cars, and did it change?” The channel is Auto Focus, a car review channel. The job had been started on 13 August 2026, three weeks before the job tools reached the MCP server, and its 30 videos are a selection made in the web app when it ran, not the channel's whole catalogue; everything below is about those 30. The account is on the free plan, which cannot start a job from the chat but can read one. The balance was 25 credits before the first call and 25 after the transcript reads; the date lookups that followed use a tool that never charges.

list_jobs returned sixteen jobs and the channel job was fourth. get_job reported 30 videos, 29 completed, 1 failed. list_job_videos without transcripts listed all 30 titles in one call and named the failure: one video returned “This content isn't available” and was refunded. Three calls with include_transcript then returned the 29 transcripts, ten, ten, and nine, as timed segments. Joined into prose they came to about 81,000 words, the longest a 4,666-word Cybertruck review and the shortest a 170-word robotaxi ride.

The job rows carried no dates, so three pages of list_channel_videos placed the 30 videos among the channel's uploads at “1 year ago” and “2 years ago”, and get_video_info on the oldest and newest pinned the span exactly: a video about a new daily driver on 11 January 2024 and a Mercedes hybrid review on 10 October 2024. Nine months, two hosts, 29 videos, and mostly automatic captions: of the eight videos checked, one also carried a manual track.

What did reading all 29 show?

Five positions recur across the 29 transcripts. Every claim below links to its video at the quoted moment, and the quotes are the automatic captions' own words, so check them at the link before you repeat them.

• Range is the recurring deal-breaker. Across nine videos the same judgement lands on otherwise-liked cars: the Lexus RZ 450e at 200 miles (“there's no way I can actually recommend this car”), the Toyota BZ4X at 220 (“that's its Achilles”), the Genesis GV70 at 236 (“I would trade a little bit of the performance in a straight line for more range”), the Taycan Turbo GT at 280 (“not going to be livable for a lot of people”). The praise runs the other way for the Silverado EV, which indicated 475 miles, “the most range I have ever tested on any electric vehicle”.
• Some vehicles suit electrification and others do not, stated as a theory and then applied. The Rolls-Royce Spectre: “What more perfect vehicle could there be to go electric?” The Genesis: “there are certain types of cars that are best with today's tech to be going electric”. The Mercedes eSprinter van: “one of the better types of vehicles to go electric”. And the exception, in the video announcing a petrol Porsche as a daily driver: “sports cars though are not yet one of those things”.
• Cheap new EVs lose to used flagships. The $25,000 EV video concludes the price target may never arrive because “batteries are the most expensive part of any electric car”; the VinFast review frames the choice as “a brand new cheap piece of tech or a depreciated older flagship” and picks the flagship; the Fisker Ocean is “the worst car I've ever tested” in its first review and, a month later after a software update, “still the worst car I've ever reviewed” in its second.
• Software separates the field. Tesla's is “the best software in any car that's not CarPlay or Android Auto”; Kia's is “top three most responsive car software I have ever used”; the Fisker is “still a bit of a laggy buggy mess” and the VinFast “just laggy on its face”; two Mercedes interfaces are compared to Windows 95 and Windows Vista. The refreshed Model 3 is “the car I would recommend so easily to anyone just looking for a first EV”.
• The frame that ties them: suitability depends on the vehicle and the use, and the channel applies it to itself. The BZ4X review asks whether Toyota's hybrid strategy “is the way to go right now”; the hybrid AMG sedan that closes the span is judged to have “no real flaws other than the price”; and the January video moves the channel's own daily driver from a Model S Plaid to a petrol 911 while calling EVs the right answer for luxury sedans. These are the same position applied to different cars, not a change of mind; the transcripts show no dated reversal.

None of that is visible from titles. It comes from reading the transcripts against each other, which is the work the job tools make free.

What should a source-backed expert brief include?

Start with a short answer to your question, then put the source beside each finding. A list of videos at the end is not enough if the reader cannot tell which one supports which sentence. Ask the assistant to produce a brief you can verify or share, not only a list of themes.

• State the collection reviewed: channel or job, videos read, failures, and known date range. Identify a selection as a selection rather than the whole channel.
• For each finding, include a short exact excerpt, the video title, and a link to the quoted moment. Keep the assistant's summary separate from the quoted words.
• Identify who is speaking. An interview answer, a host's question, and a quotation from someone else are not all the expert's own position. Mark uncertain attribution for review.
• Include qualifications and counterexamples from the reviewed material. Repetition can show a recurring view, but a position stated once still matters; frequency alone does not establish how strongly someone holds it.
• Close with what the sources do not settle. “Not found in the reviewed transcripts” does not mean the expert has never discussed it. A source shows what was said; it does not independently prove that the claim is true.

The Auto Focus example above shows why context belongs in the brief. Different verdicts on luxury sedans and sports cars should not automatically become a story about the channel changing its mind. Compare the subject and conditions before calling two statements contradictory. When chronology matters, check publication dates with get_video_info and include_dates: true; do not order videos precisely from “2 years ago” labels.

What prompt gets an answer with usable sources?

Name the existing job, the expert, and your question, then specify the evidence the answer must retain. This prompt uses saved transcripts first and asks before fetching anything new.

Use my existing transcript job for [channel or job name].
Expert: [person, or say I want the channel's different speakers compared]
Question: What has this expert said about [topic]?
Scope: [videos or period of interest]

Find the job with list_jobs and check it with get_job.
Read its saved results with list_job_videos and include_transcript: true,
following next_cursor. Report videos actually read, failures, and any
unread pages. Ask before creating a job or fetching new transcripts or
metadata. Do not use get_transcript to fetch these saved videos again.

Give me a short answer followed by findings with sources. For each:
- State the finding as a summary, not an invented quotation.
- Include a short exact excerpt and identify the speaker where supported.
- Give the video title and a timestamped YouTube link built from the
  returned video_id and segment start time. If timing is missing, link
  the video and say the timestamp is unavailable. Do not guess it.
- Explain relevant context, conditions, and any qualifying statements.

Separate the expert's words, other speakers' words, and your interpretation.
If comparing views over time, use checked publication dates; request
get_video_info with include_dates: true only for dates we still need.
Keep unavailable dates unknown. Do not infer a reversal from statements
about different situations.

For a large job, keep a source-linked evidence note per batch and use
those notes for synthesis. Mark any gaps instead of claiming full coverage.
End with unanswered questions and quotes or speaker identities I should
verify in the original videos. Do not fill gaps from general knowledge.

A timestamped link has the form https://www.youtube.com/watch?v=VIDEO_ID&t=SECONDSs. Use the segment's actual start time, rounded down to a whole second, and check that the link reaches the passage. The video ID and transcript segments come from the job; the assistant assembles the citation. Save the brief with those references so a follow-up question starts from evidence you already have.

Where does channel summarizing stop?

The brief can only cover accessible transcripts the assistant actually read. Check these boundaries when choosing a collection and deciding which conclusions it supports.

• Ten transcripts per call. A 29-video job is three calls; a 500-video job is fifty, and whether a chat holds that depends on the client and the videos' length. For large channels, keep an evidence note for each batch with its video links and quotations, then synthesize those notes. Report anything not read instead of treating pagination as unlimited memory.
• Dates are not in the job. The listing gives relative dates and get_video_info with include_dates: true requests exact ones, one video per call. These lookups charge no credits, but can return dates: null. Keep missing dates unknown rather than inventing an order.
• Auto-generated captions. Of the eight videos checked here, seven had only an automatic track, so car names arrived misspelt and sponsor reads sit inside the text. YouTube's automatic-caption guidance recommends reviewing machine-generated captions. Check important quotes and uncertain speakers against the video before publication.
• Starting a job from the chat needs a paid plan. The web app can start a job within its own limits and your free balance, and the connected chat reads it.
• MCP job caps: 500, 2,000, and 10,000 videos on Starter, Pro, and Business, and one, two, or three jobs running at once. Channel collection excludes Shorts. The website has its own 5,000-video cap, within your available balance.
• Hosted download links expire after 100 days. Download the completed files and save your brief; copies you have kept do not expire with the link.

What does it cost?

A transcript-only job reserves one credit per video, with no extra service credits for reading its saved transcripts. Failed or uncollected items are refunded within the billing cycle that funded the job. Listing jobs, checking status, reading transcripts back, listing the channel, and fetching a video's details are free at any balance, which is why the reads here left the balance at 25. The free plan is 30 credits every 30 days with a Google sign-in; Starter is $9 a month for 1,000, Pro $19 for 5,000, Business $49 for 20,000, and unused credits do not roll over.

Leave optional comments off when researching what the expert said. Including them reserves another credit per video for a Top comments sample; only nonempty pages are charged. Use the saved transcript readback for follow-up questions rather than fetching the same video again. Your chat provider's own subscription or usage costs are separate.

Which route fits which reader?

Read the job in the chat when you want a source-backed answer across videos. Choose a different workflow when your end product is a file archive or a searchable application.

• You want an answer you can check: connect the YouTube Video Transcript server, collect or find the relevant job, and ask for a brief with timestamped sources. Saved transcript reads spend no additional service credits.
• You need one quotation from a known video: read its transcript and check the recording. A channel-wide job is unnecessary when the source and the question are already that narrow.
• You want the files: the web app's whole-channel download hands you a ZIP in TXT, SRT, JSON, CSV, or Markdown, and DOCX through the API. That is the channel download post.
• You want a searchable corpus for hundreds of channels or an app: chunk and embed the transcripts, which is the AI training and RAG post. A chat reading pages is the wrong shape for that scale.
• The channel has a dozen videos: read them yourself. The model earns its place when the count and the cross-referencing exceed what you would sit through.

Questions people ask before trying it

You can reuse a completed collection for new questions. The source-backed brief depends on what was read and cited, not on creating another job each time you ask.

Can ChatGPT or Claude summarize a whole YouTube channel?

YouTube Video Transcript lets a connected chat find a bulk job with list_jobs and read its saved transcripts with list_job_videos, up to ten videos per page. The assistant can then answer across the material it has read. Check how much of the channel the job covers, any failed videos, and whether the chat read all pages before calling the answer a whole-channel summary. In this article's recorded run, reading 29 saved transcripts used no additional service credits.

How can I find what an expert said about a topic, with sources?

Use YouTube Video Transcript to collect the relevant videos, then ask your connected chat to read the saved transcripts and build a brief about one topic. Request a short exact excerpt, speaker attribution, video title, and timestamped YouTube link for each finding. Separate the expert's words from the assistant's interpretation, report which sources were read, and check important quotes against the recording. A channel job covers that collection, not every appearance the expert has made elsewhere.

Do I need a paid plan to summarize a channel from a chat?

Only to start the bulk job from the chat. YouTube Video Transcript's create_bulk_job requires Starter or above, with per-job caps of 500 videos on Starter, 2,000 on Pro, and 10,000 on Business. A free account can instead start a website job within its 30 free credits every 30 days and the website's 5,000-video cap, then read it back in the connected chat without extra service credits. Use the same Google account on both surfaces.

How many transcripts can the model read at once?

YouTube Video Transcript returns up to ten video rows per list_job_videos call when transcripts are included, with a cursor for later pages. Reading 500 completed transcripts takes at least fifty calls. Whether the chat can work with the combined text depends on the client and video length. For large collections, keep a source-linked evidence note from each batch, report coverage, and synthesize those notes rather than assuming the whole archive fits in one conversation.

How does the model know when each video was published?

Publication dates are not included in YouTube Video Transcript's job video rows. Listings provide relative dates; get_video_info with include_dates: true requests an exact publish date for one video without charging credits. That lookup can return dates: null, so an unknown date must remain unknown. In the recorded run here, details lookups dated the oldest and newest videos to 11 January and 10 October 2024.

What does a source-backed channel summary cost?

A transcript-only bulk job reserves 1 credit per video, with failed or uncollected items refunded within the billing cycle that funded it. Reading saved transcripts with list_job_videos, finding jobs, and checking progress spend no additional YouTube Video Transcript credits. Fetching a video again is a new charged request. Optional comments cost extra; they are not required for this workflow. Your chat provider may have separate subscription or usage charges.

Ask one question and keep the sources

Start with a relevant collection in the YouTube Video Transcript Workbench, or find a job you already completed. Connect your chat using the connection guide, then ask what the expert has said about your topic, with a quote and timestamped link for each finding. The result is a brief you can inspect, share, and question further without spending more service credits to read those saved transcripts.