Best MCP Server for YouTube (2026)

YouTube Video Transcript brings YouTube search, caption checks, viewer comments, and background bulk transcript exports into one MCP connection. It suits researchers who need to find relevant videos, compare what creators and viewers say, and collect transcripts across a channel. Pasting a URL into a chat and getting a summary back solves the last step of that work; finding the video, checking it has captions, and reading the reactions are where the hours go.

The server lives at https://api.youtubevideotranscript.io/mcp and ChatGPT, Claude, and Grok all connect to it. It exposes fourteen tools. Three spend credits: transcripts, comments pages, and bulk jobs. The other eleven cover search, listings, video details, job results, and usage without spending credits. A Google sign-in gives 30 credits every 30 days with no card, and paid plans start at $9 a month for 1,000 credits. Below: the six things to ask of any YouTube MCP server, how this one and two other hosted servers with YouTube tools answer them according to their own documentation, two runs you can repeat, and where this server stops.

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

What should a YouTube MCP server be able to do?

Six questions separate a server that reads transcripts from one that works YouTube for you. Ask them of any server before you connect it, ours included.

1. Can the model find a video without being handed a URL? Search across YouTube, search inside one channel, and list a channel or playlist page by page are three different jobs, and a server that only accepts a video ID leaves all three to you.
2. Can it tell you a video has captions before it charges you? No channel listing on YouTube carries that signal, so the server has to look at the video itself.
3. Does the transcript come in a shape the model can use? Plain prose for summaries, timestamped lines for citing a moment, structured data for anything programmatic. One fixed format means the model pays tokens it does not need or lacks the timings it does.
4. Can it read what viewers said, not only what the creator said? Comments, sorted top or newest, with reply threads, are the other half of a video.
5. Can it take a whole channel as one job rather than one video per call? A chat client will not hold a tool call open for the minutes a large channel takes, so the job has to run in the background and report back.
6. What does looking around cost before you fetch anything? Search and listing pages that each spend a credit change how freely the model can explore.

How do the hosted servers compare on those questions?

TranscriptAPI's MCP documents YouTube search and channel and playlist listing at one credit per page; Supadata's documents transcripts across several platforms, media metadata, and web extraction; neither lists comments, a caption-availability check, or a channel-wide transcript job. The table reads each row from what the vendor's MCP documentation stated in September 2026, and “not listed” means exactly that: the MCP page does not mention it. Their REST APIs may differ and are not compared here. Sources: TranscriptAPI's MCP documentation, and Supadata's MCP documentation and tool list.

Two things the table should not hide. TranscriptAPI's listing pages carry about 100 videos where ours carry 30, so a model paging a long channel there makes fewer calls, and its latest-15 call costs nothing. Supadata's server is the one to look at when the material is not only YouTube: it documents TikTok, Instagram, and Twitter transcripts and full web scraping, which we do not do at all. Our case is narrower and it is the research case: find, check, read, and collect across a YouTube channel, with the exploring free and the extraction charged.

Run one: from a topic to a transcript and its comments

Three content calls, one credit, on 5 September 2026, before comments became paid, from a free account with 26 credits remaining. The same workflow now costs 2 credits when both the transcript and comments succeed. The question was “find an MKBHD iPhone review and tell me what viewers disagreed with”; here is what each call returned.

search_channel_videos on @mkbhd for “iphone review” returned 28 videos, each with its duration, relative date, and view count, and charged nothing. The ranking is YouTube's own relevance order, and it is loose: a 51-second Short came first and the 22-minute “iPhone 14 Pro Review: This Will Be Copied!” sat at position 21. The model picks by reading titles and durations, which is why the results carry both.

get_video_info on that video, with dates, returned the facts below and charged nothing. The video listed four caption tracks, including English, Hindi, and Spanish.

title        iPhone 14 Pro Review: This Will Be Copied!
channel      Marques Brownlee
duration     22:21
views        13,547,779
published    2022-09-14
likes        356,252
captions     en (auto-generated), en-US, hi, es

get_transcript with include_comments returned the full transcript, opening with Title and Author lines, followed by the first page of comments under a header reading “Top comments (18,606 total)”. The top comment, at 32K likes and 498 replies, argued that the Dynamic Island is a marketing device that made an unchanged notch feel new; the second, at 17K likes, credited LG with the ideas Apple is praised for. That is the disagreement the question asked for, and it took one credit under the old pricing: the recorded balance went from 26 to 25.

Run two: a whole channel as one job, read back in the chat

A 90-video channel job on the same account, read through the job tools on 5 September 2026 without spending anything. The job was started from the web app on 12 August 2026 by pasting the channel URL; the account used for these runs is on the free plan, which cannot start bulk jobs from the chat, and the MCP job tools read a job the same way whichever surface created it.

list_jobs returned the account's sixteen jobs newest first with their status, counts, and a link to each job page. get_job on the channel job returned:

status            completed
source            https://www.youtube.com/@bootuseua
total_videos      90
videos_completed  87
videos_failed     3
url               https://youtubevideotranscript.io/app/history/job/9c02...

list_job_videos filtered to failed rows showed why three of ninety did not complete. One video had no captions, and two failed on upstream fetch errors during the run. Failed videos are refunded when a job settles, so the job cost 87 credits, not 90. The messages are as the tool returned them:

nHoUDAWwpME   No captions available for video nHoUDAWwpME
p5wpYmbnL2o   Captions player request timed out after 10000ms
GT1-mn_uzCw   InnerTube fetch failed: fetch failed

list_job_videos with include_transcript returned a completed video's transcript as timed segments, language uk, straight from the job and without a credit. That is the step that makes a channel job useful in a chat: the model reads what the job already paid for, ten videos per page, and answers from it. The ZIP of all 87 files downloads from the job page in the web app.

{ "text": "Сьогодні в має на тестуванні ось такий", "start": 0.24,  "duration": 3.28 }
{ "text": "корпус для самостійного збирання",       "start": 1.839, "duration": 4.841 }

Where does this server stop?

Five limits are worth knowing before you connect, because each one will show up as a refusal in the chat rather than a surprise on a bill.

• Bulk jobs from the chat need a paid plan. A free account gets the same jobs in the web app, running on its 30 free credits every 30 days, but create_bulk_job returns an error until you are on Starter or above. Per-job caps apply as well: 500 videos on Starter, 2,000 on Pro, 10,000 on Business.
• Captions only. The server fetches the captions YouTube already has, manual or auto-generated, and does not transcribe audio. A video with no caption track has no transcript here, which is why get_video_info exists.
• Channel listing is 30 videos per page. A long channel takes several pages to read in full, and the model pages them one at a time. The exact count and the bulk job cover a whole channel without that.
• Job files expire after 100 days. The transcripts you read in the chat are yours, but the ZIP link on the job page stops working after that, so pull it somewhere durable.
• Six export formats, and WebVTT is not one of them. Jobs deliver TXT, SRT, JSON, CSV, DOCX, or Markdown. If a pipeline needs .vtt files, yt-dlp writes them natively and costs nothing.

One more thing that is a property of the clients rather than the server: ChatGPT and Claude cache the tool list when a connector is added. When new tools ship, refresh the connector in the client settings to see them. The steps are in the Claude setup guide.

What does it cost inside a chat?

Three tools spend credits. get_transcript costs 1 credit per successful transcript. get_video_comments costs 1 credit per nonempty page, including replies. Adding comments to get_transcript costs 2 credits when both succeed. Failed fetches and empty comments pages are not charged. create_bulk_job charges 1 credit per video up front and refunds failed videos when the job settles. The other eleven tools, including search, listings, video details, job results, and usage, are free at any balance.

The free plan is 30 credits every 30 days with a Google sign-in and no card. Starter is $9 a month for 1,000, Pro is $19 for 5,000, Business is $49 for 20,000. Unused credits do not roll over. The balance is one number shared by the MCP server, the web app, and the REST API, so there is no separate subscription per surface. Every transcript fetch still costs a credit wherever it runs, so fetching a video in Claude and again in the browser charges twice; downloading a finished bulk job is free from any of them.

Which server fits which reader?

Connect the YouTube Video Transcript server for research across YouTube channels; the other three cases below each have a better fit, and the free web app covers bulk until you pay.

• You ask an AI chat about YouTube videos and want it to find, check, read, and compare them, and to collect transcripts across a channel, without you pasting anything: connect the YouTube Video Transcript server. Fourteen tools, one endpoint for ChatGPT, Claude, and Grok, 30 free credits to try it.
• You mostly page through long channels and playlists and want fewer, larger pages: TranscriptAPI's MCP lists about 100 videos per page against our 30, at one credit per page, and its latest-15 call is free.
• Your material is not only YouTube: Supadata's MCP documents transcripts from TikTok, Instagram, and Twitter and full web scraping and crawling, none of which this server does.
• Your transcripts are mostly meetings, and YouTube is occasional: Tactiq's MCP is built for its meeting transcripts and ships Slack, Notion, and HubSpot integrations beside it. Its YouTube tool exports to TXT and the clipboard only.
• You want hundreds of transcripts and have no budget yet: run the bulk job in the web app on the free 30 a month, and connect the MCP server for the reading. Bulk from the chat is the one thing that waits for a paid plan.

Questions people ask before connecting

What does a YouTube MCP server actually let a chat do?

It lets the model call YouTube itself instead of waiting for you to paste text. With the YouTube Video Transcript server connected, ChatGPT, Claude, or Grok can search YouTube, list a channel or playlist, check whether a video has captions, read its transcript and its public comments, and start a job that fetches every transcript in a channel. You ask in plain language and the model picks the tool. Fourteen tools are exposed; transcripts, comments pages, and bulk jobs spend credits; the other eleven tools are free.

Which tools cost credits and which are free?

Three tools spend credits. get_transcript costs 1 credit per successful transcript. get_video_comments costs 1 credit per nonempty page, including replies. Adding comments to get_transcript costs 2 credits when both succeed. Failed fetches and empty comments pages are not charged. create_bulk_job charges 1 credit per video up front and refunds failed videos when the job settles. The other eleven tools, including search, listings, video details, job results, and usage, are free at any balance.

Can I fetch a whole YouTube channel from inside Claude or ChatGPT?

Yes, on a paid plan. Ask for every transcript from a channel URL and the model calls create_bulk_job, which lists the channel for you, charges one credit per video, and runs in the background. The job is capped at the smaller of the channel size, your plan's per-job cap (500 videos on Starter, 2,000 on Pro, 10,000 on Business), and your remaining credits. The model can report progress with get_job and read finished transcripts back with list_job_videos without paying again. The ZIP downloads from the job page in your account. Free accounts get the same bulk jobs in the web app, where they run on the 30 free credits every 30 days.

How do I know a video has a transcript before I spend a credit?

Ask the model whether the video has captions. It calls get_video_info, which returns the caption tracks and their languages along with the title, exact view count, duration, and description, and charges nothing. An empty caption list means get_transcript would fail for that video. The server fetches YouTube's existing captions, manual or auto-generated; it does not transcribe audio, so a video with no captions has no transcript to fetch.

Do I need a different server for ChatGPT, Claude, and Grok?

No. One endpoint, https://api.youtubevideotranscript.io/mcp, serves every client. Claude.ai, Claude Desktop, and ChatGPT connect with OAuth and a consent screen; Claude Code and Grok take an API key. The same account and the same credit balance also work in the web app and the REST API, so a transcript fetched in a chat and one fetched from a script draw from one allowance. Click-by-click setup for each client is in the companion post on connecting the server.

Connect it once

Sign in with Google at youtubevideotranscript.io, add https://api.youtubevideotranscript.io/mcp to ChatGPT, Claude, or Grok, and ask your chat a question about a YouTube video you have not opened. The 30 free credits every 30 days cover single-video transcripts or comments pages. Eleven discovery and job reading tools need no credits; bulk jobs from the chat wait for a paid plan. Per-client setup is in the connection guide and the full tool reference is at /docs/mcp.

Related reading: how to download every transcript from a channel from the web app or the API, and the comparison of YouTube transcript APIs for anything scripted rather than conversational.