How to Research a Topic Across YouTube from Your Chat (2026)

YouTube Video Transcript lets ChatGPT, Claude, and Grok search YouTube, narrow to one channel, check whether a video has captions, and pay a single credit only for the transcript it finally reads. The finding is free; only the reading is paid. That order matters because most of the time in a research session goes on locating the right video, not on reading it, and the chat can do the locating with tools that never touch your balance.

The server is at https://api.youtubevideotranscript.io/mcp. A Google sign-in gives 30 credits every 30 days with no card, and every search and lookup described here is free at any balance. This post covers the tools that do the finding, a real run from a topic to a transcript on 5 September 2026, the ranking behaviour that will mislead you if you do not know about it, and where free discovery stops.

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

What does researching a topic on YouTube involve?

Four steps, and only the last one has ever cost anything. You find candidate videos for a topic, you narrow to the creator or channel you trust, you check that the video you picked can be read at all, and then you read it. Done in a browser, the first three steps are tab-switching and the fourth is copy-paste. Done from a chat with the server connected, each step is a tool call the model makes itself, and the first three are free.

Which tools do the finding, and what do they cost?

Six tools cover discovery and one covers reading. The costs come from the server's own tool descriptions, which the model sees too.

Every row feeds the next: a search row's video_id goes to get_video_info or get_transcript, a channel_id to the channel tools, a playlist_id to list_playlist_videos. The tool descriptions say so, which is why the model chains them without being told the tool names.

Run: from a topic to a transcript and comments

On 5 September 2026, on a free account, the question was “what did a well-known reviewer think of the iPhone 14 Pro, and how did viewers react?” Three content calls, plus a balance check before and after, answered it for one credit: the balance went from 26 to 25. That was before comments became paid. The same workflow now costs 2 credits when the transcript and comments succeed; without comments it still costs 1. Everything below records the original run.

search_channel_videos on @mkbhd for “iphone review” returned 28 videos with a token for more. The first result was a 51-second Short, “Samsung broke this”, at 7,093,985 views. The full review, “iPhone 14 Pro Review: This Will Be Copied!”, sat at position 21 with 13,547,791 views. That is YouTube's relevance ranking, and it is the first lesson of the run: the results are ranked, not filtered, and a model reading them needs the duration column as much as the title.

get_video_info on that review with include_dates returned 13,547,779 views, 22 minutes 21 seconds, published 14 September 2022, 356,252 likes, category Science and Technology, a description carrying the creator's own chapter markers, and four caption tracks: English auto-generated, English (United States) manual, Hindi, and Spanish. The listed tracks are what made the next step safe to pay for.

get_transcript with include_comments returned the transcript with the Title and Author lines at the top, followed by the first page of top comments from the 18,606 the video carried that day. One credit under the old pricing. The model had the review's own words and the audience's first reaction in one response, and nothing before that call had spent anything.

How does the ranking mislead, and how do you steer it?

Three examples from the same day show what the sort and type options change, all through search_youtube and all free.

• Relevance and views disagree. “mkbhd iphone review” with type set to video returned 20 results against an estimated 362,674, led by “iPhone 16e Review: Who Are You?” at 8,759,453 views. The same query sorted by views led with the channel's Apple Vision Pro video at 28,650,908 views, which matches the name and the word review but not the topic. Whichever sort you use, read the titles and snippets against the question before spending a credit on any of them.
• The default mixes Shorts in. “minecraft” with type left at all returned 31 results of which 25 were Shorts and 6 were videos. Shorts rarely carry captions worth reading, so for transcript work set type to video, and set duration to medium or long when the topic needs a full treatment.
• Channel search resolves names. Searching with type set to channel for the reviewer's name returned Marques Brownlee, @mkbhd, 21.2M subscribers, and the channel ID that every channel tool accepts. That is the step that turns “that tech reviewer” into a channel the model can search inside.

Page two of the relevance search returned 20 further results with none repeated from page one. Estimated totals in the hundreds of thousands describe YouTube's index, not the number worth reading.

What should you ask the model to do?

Name the topic, the constraint, and the check, and let the tools do the rest. These prompts each produce a chain that ends in one paid call or none.

• “Find three videos from the last month explaining [topic], each over ten minutes, and tell me which have English captions.” Search with upload_date and duration filters, then a details check on each; no credit spent.
• “Which of [creator]'s videos cover [product]? List them with views and dates.” Channel search, one free call, with the Shorts visible by their duration so you can rule them out.
• “Pick the most-viewed full review of [product] on YouTube, confirm it has captions, then summarize it.” Search sorted by views with type video, details check, one transcript, one credit.
• “Before you read anything, tell me how many credits I have and how many videos you plan to read.” get_usage is free, and the answer keeps a research session inside a budget.
• “Find the channel behind this topic and how many long-form videos it has.” Channel search then the exact count, both free, which tells you whether a whole-channel read is a small job or a large one.

Where does free discovery stop?

The tools find and check; they do not judge, and a few limits shape what a search can tell you.

• Ranking is YouTube's. Relevance, views, date, and rating are YouTube's orderings applied server-side, and an in-channel search has no sort at all. The model can re-rank what it fetched, but it cannot ask for a different index.
• Results carry relative dates. A search row says “1 year ago”; the exact publish date is one get_video_info call with include_dates per video, still free but a request each.
• A listed caption track is a strong signal, not a guarantee. The fetch is a separate step and can fail on a restricted video or a transient error, in which case nothing is charged. The missing-transcript post covers what an empty list means.
• No comments or transcript in search results. The row tells you a video exists and how it performed; what it says costs the credit, and how viewers reacted costs 1 credit per nonempty page from get_video_comments.
• Twenty results a page is a YouTube page, not a cap you can lift. A model that needs more pages through tokens, and each page is a free call, but the relevance of page five is what it is.

Which route fits which reader?

Free discovery inside the chat is the right fit for most research sessions, and there are two cases where something else is.

• You research topics and want the model to find, check, and read in one conversation, paying only for transcripts it reads: connect YouTube Video Transcript. Six free discovery tools and one paid reading tool, 30 credits every 30 days free to start.
• You want to find a video and watch it, not read it: YouTube's own search page is better at that. It has autocomplete, personalisation, and every filter, and none of it needs a connector. What it cannot do is check captions or hand the words to your chat.
• You already use TranscriptAPI's server: its MCP documentation lists YouTube search at about 20 results a page and in-channel search at about 30, each at 1 credit per page, and channel listing at 1 credit per page. The same discovery steps run there on a per-page charge. A caption-availability check and comments are not listed in its MCP documentation. Supadata's documents multi-platform transcripts, metadata, and web scraping; YouTube search is not listed in its MCP documentation.
• You run research as a script rather than a conversation: yt-dlp can search with a ytsearch prefix and list results for a pipeline, with the cloud-host and caption caveats the dataset tools post covers.

Questions people ask before searching from a chat

Does searching YouTube through the MCP server cost credits?

No. search_youtube, search_channel_videos, get_video_info, list_channel_videos, list_channel_playlists, and get_channel_video_count are free at any balance, including an account that has used all of its credits. get_transcript costs 1 credit per successful transcript. Comments are optional and cost 1 additional credit per nonempty page, including replies. Finding the right video and checking its captions remain free.

How many results does one search return?

About 20 per page from search_youtube and about 30 per page from search_channel_videos, each with a continuation token for the next page. The first page of a YouTube-wide search also carries YouTube's estimated result count, which was 362,674 for a reviewer's name plus a topic. There is no total for an in-channel search. Paging is free, so the model can keep going, but relevance falls quickly and the useful results are usually on the first page or two.

Why did a Short come first when I searched inside a channel?

Because in-channel search returns YouTube's relevance ranking, not a filter, and Shorts that mention the topic rank alongside full reviews. In the run in this post a 51-second Short led the results for a phone review query and the full review sat at position 21. Ask the model to skip results under a few minutes, or use search_youtube with type set to video and duration set to medium or long, which are filters YouTube applies before ranking.

How do I know a video has a transcript before paying for it?

Call get_video_info first. It returns the video's caption tracks with language codes and whether each is auto-generated, along with the exact view count, duration, and description, and it charges nothing. An empty caption_tracks list means get_transcript would fail. A listed track means a fetch is likely to succeed, though fetching is a separate step that can still fail, in which case nothing is charged.

What does the connector add that a chat's own browsing does not?

The connector gives the model dedicated YouTube search, channel search, caption checks, and transcript retrieval in one workflow, with the tool it called visible at each step. A client's built-in browsing may find a video; the connector is what turns a found video into a caption check and a transcript in the same conversation, and it is the only part of that chain that spends a credit.

Connect it once

Setup is the same for every tool on this page and takes a few minutes: the install guide covers ChatGPT, Claude, and Grok step by step, and the MCP docs list every tool with its cost. For how this server compares with others on the same questions, see the 2026 comparison.