Analyze YouTube comments with ChatGPT or Claude: do viewers agree?

YouTube Video Transcript lets your chat compare a review's transcript with viewer comments and reply threads to find agreement and important objections. A confident review and a popular comment can point in different directions, but the disagreement may be about different claims. Read what the reviewer actually said, match the reactions to it, and open the relevant replies before deciding what the discussion tells you.

The server is at https://api.youtubevideotranscript.io/mcp. Reading comments costs 1 shared credit per nonempty page; the transcript costs 1 credit separately. Google sign-in gives 30 credits every 30 days with no card. Below is a recorded discussion analysis from a video with 18,604 comments, followed by a sampling plan and prompt for your own review. The tools supply the words and source references; your assistant compares them.

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

Do viewers agree with the review, or are they arguing about something else?

Compare reactions with specific statements in the transcript, not with an overall positive or negative label. A viewer can agree with a test result but reject the recommendation because their budget, use case, or priorities differ. That qualification may be more useful than another general thumbs-up.

Start by identifying the claims relevant to your decision and keeping their timestamps. Then separate comments that support those claims, challenge them, add a condition, or ask something the review did not settle. Keep jokes, general brand opinions, and arguments aimed at another commenter separate from direct responses to the review.

An important objection is one that could change the conclusion for your situation, even if it has few likes. A report of a different result is a lead to examine, not a verified test. Look for the conditions described, the reviewer's original qualification, and any reply that addresses the objection. The output should explain what needs checking, not declare a winner from the comment ranking.

What can a chat read from a video's comments?

The comments tool returns text, engagement indicators, source links, and continuation tokens for further reading. A first call to get_video_comments usually returns about 20 top-level comments and a total count when YouTube supplies it. Use these fields to inspect the discussion, not to count votes for the review.

Top uses YouTube's ordering, not a strict sort by likes; in the run below a 17K-like comment came before a 32K one. Newest gives a view of recent responses. YouTube documents both comment orderings and links to individual threads. Keep the returned comment URLs with your quotes so a reader can inspect the discussion behind the summary.

How does a real analysis go?

Four calls on 5 September 2026, before comments became paid: three comments calls and one transcript-with-comments call cost 1 credit in that run. The same four calls now cost 5 credits when all pages are nonempty. They were made against Marques Brownlee's iPhone 14 Pro Review: This Will Be Copied!, which had 18,604 comments in the recorded response. The question was “what did viewers actually think, and has that changed?” Everything below describes that sample, not a new collection. The responses reported total counts two apart a few minutes apart.

The first page, sorted Top, held 20 comments with likes from 5 to 32K. Three themes were visible without reading further. The two most-liked comments were about the phone's new screen cutout: at 32K likes, “just another one of the small marketing schemes that Apple does that works so well”, and at 17K, credit to LG for having done the same ideas years earlier, echoed by three more comments at 3.4K, 883, and 807 likes. The third theme was crash detection, told as personal stories: a fiancé's phone that called 911 after an accident, and a first responder who wrote that most crash-detection callouts are phones left on car roofs.

The 32K comment had 498 replies, so the second call opened that thread with its reply_continuation. The first ten replies complicated the headline. One agreed, at 1.7K likes. The rest pushed back: “the main idea is the software-hardware integration, not the pill shaped camera” at 1.4K, “It is new though” at 808, and a 226-like reply asking why it counted as a scheme at all. Reading only the top-level page would have missed those responses. In that sample of ten, the most-liked take from the sampled Top page drew more pushback than support. Reading the replies exposed disagreement the top-level page missed.

The third call, sorted Newest, showed a different set of responses to the older review. The twenty returned comments, from two days to five months old, were almost all from people buying the phone now at a discount: “Finally bought the 14 pro yesterday”, “Is £290 a good price for this phone?”, “3 months away from buying mine”. Those questions suggest a buying context different from the novelty debate in the Top sample. They do not establish that the whole audience changed.

A fourth call, get_transcript with include_comments, now costs 2 credits when both succeed, and it is worth it when the question is whether viewers argued with the video rather than with each other. On this review it returned the full transcript and the same first page of top comments in one reply, so the model could set the creator's line that the cutout “will be one of the most copied features” beside the 32K-like comment calling it marketing.

What does this example say about agreement with the review?

It shows a contested interpretation of the feature, not a demonstrated error in the review. Calling the cutout marketing does not directly contradict a prediction that other companies will copy it. The replies add a different criterion: whether the software and hardware work together in a useful way.

A useful brief would retain the reviewer's prediction, the skeptical comment, and the replies that qualify that skepticism. It would describe the newer buying questions separately. Flattening all of this into “viewers disagree” loses both the target of the disagreement and the question a buyer still needs answered.

How should you collect a review, comments, and replies?

Start with a timestamped transcript and one Top comments page, then choose whether a reply thread or a Newest page would help answer your question. The separate comments tool supplies the comment links and reply tokens needed for this workflow.

1. Request get_transcript with format: "timestamps" and comments off. Identify the claims relevant to your question and keep the moments where the reviewer states or qualifies them. A successful transcript costs 1 credit.
2. Request get_video_comments with sort: "top". Keep the returned comment IDs, URLs, and reply tokens. A nonempty page costs 1 credit and usually contains about 20 top-level comments.
3. Open one relevant objection's replies when a reply_continuation is available. Pass it as continuation to the same comments tool and retain the parent comment alongside the replies. A nonempty reply page costs another credit; it is not necessarily the whole thread.
4. If you need recent reactions, make a fresh comments call with sort: "newest" and no continuation token. That nonempty page costs another credit. A continuation keeps its original ordering; changing the sort beside it does not start a Newest sample.

All four steps cost 4 credits when the transcript succeeds and each comments page is nonempty. This is a suggested sampling plan, not another recorded test. Unlike the historical sequence above, it does not fetch the first Top page twice. Reuse anything already loaded and ask the assistant to stop before additional pages unless you approve them.

For a quick transcript-plus-reactions read, you can instead use get_transcript with include_comments: true. It costs 2 credits when both return data, but the bundled sample does not include reply continuation tokens. Use separate calls from the start when you intend to investigate threads.

What prompt finds important objections without inventing consensus?

Ask the assistant for a claim-by-claim comparison with sources and an explicit account of what it read. This prompt keeps replies attached to their parent comments and limits new fetching.

Review: [YouTube URL]
My question or buying situation: [what I need to decide]
Maximum new collection spend: 4 service credits. Reuse data already here.

Get the transcript with format: "timestamps" and include_comments: false.
Read one Top page through get_video_comments. If useful, open one page
of replies under a relevant objection using its returned reply_continuation,
and one Newest page with no continuation. Ask before any additional fetch.
Do not invent tokens or keep trying failed calls without asking me.

List the review claims relevant to my question, with short exact excerpts
and timestamped video links. Match sampled reactions to those claims.
Separate agreement, direct objections, qualifications, unanswered questions,
and comments that address another commenter rather than the review.

For each important reaction, quote the words and retain the returned
comment URL. Explain which claim or parent comment it addresses and
what the replies add. Mark self-reported experience as unverified.
Do not treat different preferences as a factual contradiction.

Report how many unique top-level comments and replies you read, by ordering
or thread. Deduplicate repeated comment IDs across pages. Like counts show
engagement, not audience votes; do not turn them into agreement percentages.

Answer: where did the sampled viewers agree, what objections could change
the decision, and what remains unresolved? Include context and counterexamples.
If the transcript or a thread is unavailable, state the missing evidence.
Do not generalize the sample to all viewers or fill gaps from general knowledge.

The finished answer should let you move from an objection to its comment, its replies, and the review passage it concerns. Keep a short exact quote beside each interpretation. Check the recording before relying on a captioned number or a claim that depends on a visual test the transcript cannot show.

Where does comment analysis stop?

The answer describes the accessible comments you read, not every viewer's opinion or an independent product test. Keep these boundaries visible in the brief.

• Top and Newest are different samples and may overlap. Count unique comment IDs, keep replies separate from top-level comments, and state how much of each thread was read.
• Top is YouTube's order, not a likes sort. A prominent comment can contain a joke, a question, or criticism of another viewer rather than an objection to the review.
• Like counts are display strings. 32K is not 32,000 exactly, and adding likes across comments does not count distinct people or measure agreement with the reviewer.
• Comments turned off means an error, not an empty page. The tool says so. A transcript may still be available, but caption availability is a separate check. Without the transcript, limit the answer to the comments rather than claiming a comparison with the review's exact words.
• Public comments only. YouTube's comment settings allow creators to hold comments before they become public. The tool does not expose those held comments, hidden material, or anything requiring a YouTube sign-in.
• The tool returns comments; the model does the judging. Sentiment, themes, and summaries are the model's reading of the text, so ask for quotes and links behind every claim.

Of the two other hosted MCP servers with YouTube tools that we compared in the MCP server roundup, neither TranscriptAPI's nor Supadata's MCP documentation listed a comments tool in the documentation read on 5 September 2026.

What does it cost?

One credit per nonempty comments page, one per transcript, from the same balance on the website, API, and MCP, including reply pages. Errors and empty pages are not charged. A successful transcript with a nonempty comments page costs 2 credits. You can read comments without buying a transcript first. Search and video details remain free.

The free plan gives 30 credits every 30 days with a Google sign-in and no card. Starter is $9 a month for 1,000 credits, Pro $19 for 5,000, Business $49 for 20,000. Unused credits do not roll over. Fetching the same comments page again costs another credit; asking follow-up questions about comments already in the chat does not require another fetch.

Your chat provider may have separate subscription or usage costs. The four-credit plan above limits the new data you ask it to collect; further analysis of the same supplied text uses no additional YouTube Video Transcript credits.

Which route fits which reader?

Connect the MCP server when you want to compare a review and its discussion inside your chat. Use the website collection tool when the result you need is a larger set of files.

• You want an answer with evidence from both sides: connect YouTube Video Transcript and compare the timestamped review with comments and selected replies. Keep the sources so you can inspect an objection before relying on it.
• You moderate your own channel's comments: YouTube Studio is where you reply, filter, and manage them. That is a moderation task; this article's workflow is for comparing accessible public discussion with a review's claims, including videos on someone else's channel.
• You need a downloadable dataset of comments: start a standalone collection in the website's Comments tool. It exports CSV and JSON for a video, playlist, or channel, with optional replies and a credit limit. The comments download guide covers that workflow separately from analysis in chat.
• You need one quick answer from a short discussion: read it directly. Use the connected workflow when tracing claims through comments and reply threads is the work you want help organizing.

Questions people ask before trying it

Start with the review claim you want to examine and collect only enough discussion to investigate it. These answers clarify the costs, reply access, and coverage of that workflow.

How can I check whether viewers agree with a YouTube review?

Use YouTube Video Transcript to read the review's timestamped transcript alongside a sample of Top or Newest comments and selected reply threads. Ask your chat to match each reaction to the specific claim it addresses, distinguish objections to the reviewer from arguments between commenters, and include video and comment links. Report agreement or disagreement within the comments read, not as a verdict from the whole audience.

Does reading YouTube comments through the MCP server cost credits?

Yes. YouTube Video Transcript charges 1 shared credit per nonempty comments page, including Top, Newest, continuation pages, and reply pages. Top-level pages usually contain about 20 comments; page sizes vary. Empty pages and failed requests are not charged, but at least 1 credit is required before fetching. A transcript costs 1 credit separately. Adding include_comments costs another credit only when a nonempty comments page is returned; a comments failure leaves a successful transcript available.

Can the chat read the replies under a comment?

Yes. When get_video_comments returns a reply_continuation for a comment, pass it back as continuation to read that thread. Retain the parent comment so the assistant knows what the replies address. Reply pages cost 1 credit when nonempty, and their size varies. In the recorded run here, the first ten replies under a comment with 498 replies mostly pushed back on it; that finding describes those ten, not the entire thread.

Can it read all of a video's comments?

The MCP tool can follow available continuation tokens, but a chat analysis usually works with a stated sample and a credit budget. Top comments, Newest comments, and replies are different views of the discussion, not a representative audience poll. For a larger downloadable collection, use YouTube Video Transcript's standalone Comments tool with All available, optional replies, and a job credit limit. Its actual coverage depends on what is accessible and the budget.

Does it work on every video?

No. YouTube Video Transcript fetches accessible public comments, not held or hidden comments or content requiring a YouTube sign-in. Disabled comments return an error rather than an empty sample. Caption availability is separate: if a transcript cannot be retrieved, the chat can still analyze accessible comments, but should not claim to have compared them with the review's exact words.

Can I export the comments to a file?

Yes. YouTube Video Transcript's standalone Comments tool collects public comments and optional replies from a video, playlist, or channel into a ZIP with CSV and JSON files. You set an amount per video or choose All available, then set a credit limit for the job. Start it on the website and download it from History. The MCP comments tool returns pages in chat; it does not start these collection jobs.

Check the objection before making the decision

Sign in to YouTube Video Transcript and connect your chat using the connection guide. Pick one review and a question that matters to you. A successful transcript and a nonempty Top comments page use 2 of your 30 free credits; follow a relevant thread if the initial comparison leaves something unresolved. Keep the quotes and links so the answer explains the objection, not only the mood of the comments.