Everyone who works with an agent has run this exchange at least once.
You: Summarize this. https://www.youtube.com/watch?v=kCc8FmEb1nY
Agent: I'm not able to watch videos. If you paste the transcript here, I'll gladly help!
The transcript is precisely the thing the agent cannot get on its own. With this server
connected, the same message simply resolves.
YouTube transcripts, search, channels, playlists and bulk transcript jobs for AI agents. 14 tools.
2026-11-08
Sao
—
Công cụ
—
Thời gian hoạt động
—
You: Summarize this. https://www.youtube.com/watch?v=kCc8FmEb1nY
Agent: → get_transcript(video="kCc8FmEb1nY", video_metadata=true) 1 credit
That's "Let's build GPT: from scratch, in code, spelled out" by Andrej
Karpathy, 1:56:20. He starts from an empty file and a bigram model,
derives self-attention step by step, and ends with a working GPT that...
Reading one video is rarely where the job ends. Here is how the three ways of getting YouTube
data into an agent actually compare.
This server
Local yt-dlp / scraper MCP
Google YouTube Data API
Transcripts
✅ any public video, 5 formats
⚠️ blocked on datacenter IPs, breaks when YouTube changes markup
❌ not served at all
Setup
✅ a URL and an API key
❌ local install, binaries to keep alive
❌ Cloud project, OAuth consent screens
YouTube search
✅ native, 1 credit per page
❌
⚠️ 100 quota units per search
Channels & playlists
✅ 100 videos/page, or 500 bare IDs
❌ one video at a time
⚠️ quota-metered per item
Bulk transcripts
✅ 4,000 per background job
❌
❌
RAG-ready chunking
✅ 20-5,000 chars, word-level timestamps
❌
❌
When YouTube changes
✅ fixed server-side, nothing to update
❌ you patch and redeploy
✅
Failed calls
✅ credits refund themselves
❌ your retry logic
⚠️ quota spent anyway
⚡ Quick start
1. Get an API key. Sign up at transcriptout.com and create a key
in the dashboard. New accounts receive 100 free credits and
no card is asked. Keys start with sk_ and are shown once.
2. Point your client at the server. It speaks streamable HTTP and authenticates with one
Bearer header.
One-click buttons for Cursor and VS Code sit at the top of this page. Exact snippets for the rest
live under Install in your client.
3. Paste a link.
Summarize this talk and pull the three strongest quotes.
https://www.youtube.com/watch?v=dQw4w9WgXcQ
The agent picks get_transcript on its own, reads the timed text and answers from it. Every
response carries an X-Credits-Remaining header, so the budget stays in view the whole session.
🧰 The 14 tools
All 14 tools are exposed automatically once you connect. Most calls cost 1 credit. Credits
are refunded automatically when a call fails before reaching YouTube (validation errors, rate
limits, our own capacity), so you pay for answers, not for failures. A definitive "this video has no
captions" is an answer and is billed like one.
1. get_transcript · 1 credit
Fetch the transcript of any YouTube video. format=text (default) returns plain readable text,
cheapest for a model to reason over, and format=json returns timed segments.
Parameter
Type
Default
Description
video
string
required
YouTube URL (full or short) or 11-char video ID
lang
string
"en"
Language code of the track (en, de, ...)
format
string
"text"
"text" (plain text), "json" (segments with start/duration in seconds), "srt"/"vtt" (subtitle file body), "srv3" (raw YouTube XML)
kind
string
auto-detect
"manual" or "auto". Left out, a manual track wins and auto is the fallback
segment
integer
see below
Max characters per segment. 500-1500 makes RAG-ready chunks
video_metadata
boolean
false
Add title, channel, duration and views in the SAME call, same 1 credit
Left out, segment cuts auto-generated tracks into ~180-character segments and returns manual
tracks exactly as their author broke them. Pass it whenever you need one size regardless of which
track answers.
Example output (format=json):
{
"video_id": "dQw4w9WgXcQ",
"language": "en",
"kind": "manual",
"transcript": [
{ "text": "Never gonna give you up", "start": 18.0, "duration": 4.12 },
{ "text": "Never gonna let you down", "start": 22.12, "duration": 3.85 }
]
}
srt and vtt come back as complete subtitle file bodies, ready to be written to disk by the
agent. srv3 is the raw source XML and does not combine with segment.
2. get_video_info · 1 credit
Metadata for one video (title, channel, duration, views, thumbnails) plus the list of available
transcript languages, WITHOUT downloading the subtitles.
Parameter
Type
Default
Description
id
string
required
YouTube video ID or URL
Credit hygiene: if you are going to fetch the transcript anyway, call get_transcript with
video_metadata=true instead. It returns both for one credit where these are two calls and two.
3. search_youtube · 1 credit/page
Search YouTube for videos or channels. Paginate with next_page_token. has_more tells you
whether another page exists.
Parameter
Type
Default
Description
q
string
required*
Search query (*unless paginating)
type
string
"video"
"video" or "channel"
limit
integer
20
Results per page, 1-50
next_page_token
string
Token from a previous result
4. list_channel_videos · 1 credit/page
List videos from a channel's Videos tab, newest first. Accepts an @handle, a channel name, a
UC... channel ID or a channel URL.
Parameter
Type
Default
Description
name
string
required*
@handle, channel name, UC... ID or URL
limit
integer
100
Page size, up to 500 with ids_only
ids_only
boolean
false
Return just video_ids[], up to 500 per page
next_page_token
string
Token from a previous result
ids_only=true is the cheap way to feed submit_transcripts_job.
5. search_channel_videos · 1 credit/page
Search inside one channel using YouTube's native relevance search. A result whose title lacks the
query word is normal. Results are ranked by relevance, not by substring.
Parameter
Type
Default
Description
name
string
required
@handle, channel name, UC... ID or URL
q
string
required
Query to search within the channel
limit
integer
30
Results per page, 1-100
next_page_token
string
Pagination token
6. latest_channel_videos · 1 credit
The ~15 most recent videos of a channel from its RSS feed. The fastest and cheapest way to check
what a channel published recently.
Parameter
Type
Default
Description
name
string
required
@handle, channel name, UC... ID or URL
7. list_playlist_videos · 1 credit/page
Every video of a playlist in playlist order. Accepts a PL... playlist ID or a URL with list=.
Parameter
Type
Default
Description
id
string
required*
Playlist ID or URL
limit
integer
100
Page size, up to 500 with ids_only
ids_only
boolean
false
Return just video_ids[], up to 500 per page
next_page_token
string
Pagination token
8. search_playlist_videos · 1 credit
Find videos inside a playlist by a substring of the title (case-insensitive). YouTube has no native
playlist search, so this scans up to 500 playlist items. truncated=true means there may be more
matches beyond the scanned window.
Parameter
Type
Default
Description
id
string
required
Playlist ID or URL
q
string
required
Substring to match in video titles
limit
integer
30
Max matches, 1-100
9. submit_transcripts_job · 1 credit per video
Queue transcripts for MANY videos at once (up to 4,000) and get a job_id back immediately. The
work continues in the background at your rate limit's pace. Use this instead of calling
get_transcript in a loop for more than a handful of videos.
Parameter
Type
Default
Description
videos
string[]
required
Video IDs or URLs, up to 4,000. Duplicates collapse BEFORE billing
lang
string
"en"
One language for the whole job
format
string
"text"
"text", "json", "srt", "vtt" or "srv3", one for the whole job
kind
string
auto-detect
"manual" or "auto"
segment
integer
One segment size for the whole job
video_metadata
boolean
false
Metadata per video, no extra cost
idempotency_key
string
Resubmitting the same list with the same key returns the SAME job, no double charge
Requires a user key (sk_...). Credits are charged on submit and refunded per video when a video
could not be delivered through our fault.
10. get_transcripts_job · free
Progress of a batch job: status (queued/running/done/cancelled), how many videos are ready,
failed and pending. Polling a job you already paid for costs nothing.
11. get_transcripts_results · free
Finished transcripts from a batch job, in the order submitted, paged with next_page_token
(limit 1-500, default 100). Results appear as they are fetched, so you can read before the job is
done. Each entry is exactly what get_transcript returns for that video, plus its status.
12. get_transcripts_result · free
One video's result out of a batch job, by its video id, without paging through the whole result
set. A 404 means the job does not exist or this video has not finished yet, so check
get_transcripts_job before concluding anything.
Parameter
Type
Default
Description
job_id
string
required
Job id from submit_transcripts_job
video_id
string
required
One of the video ids the job was submitted with
13. cancel_transcripts_job · free
Cancel a batch job. Credits are refunded ONLY for videos not started yet. Anything already fetched
stays in the results and stays paid for.
14. get_credits · free
The remaining credit balance of the key, with no parameters. The balance also rides in the
X-Credits-Remaining header of every response, but headers are invisible to the model, so the
number a user actually asks about needs a tool. Handy right before a large batch too, since the
batch charges 1 credit per video on submit.
🔌 Install in your client
The server is remote, so every install below is a config entry and nothing more. All of them want the same two values, the URL and the Bearer header from Quick start.
Worth doing once, a standing rule for your client
With this in your client's rules/instructions, pasting a YouTube link is enough and the word
"transcript" never has to be typed:
Whenever a YouTube link or video ID appears in my message, call the
transcriptout get_transcript tool first and answer from the transcript,
whether I asked for a summary, a quote, a translation or a question.
One-Click Install:
After installing, open the server settings and add the Authorization header with your key.
Claude's custom connectors authenticate remote servers via OAuth, which TranscriptOut does not
offer yet (API keys only). On desktop, use Claude Code (see above), which supports API-key
headers. OAuth support is on the roadmap. Watch the changelog.
Or add this to VS Code user settings (settings.json):
This repository root is a conformant Agent Plugins 1.0.0
package, the portable format supported by ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and VS Code.
One install gets you the MCP server and a bundled youtube skill that teaches your agent when
to use each tool and how not to waste credits.
plugin.json # manifest
mcp.json # hosted MCP server, streamable-http
skills/youtube/SKILL.md # when + how to use the 14 tools
VS Code. Command Palette → Chat: Install Plugin From Source, then paste:
ChatGPT, Codex, GitHub Copilot, Kiro, any other client. Point your client's plugin mechanism at
this repository, or at a local clone. Agent Plugins 1.0.0 standardizes the package format, not
installation, so each client owns its own install flow.
There are no credentials in this package, Agent Plugins 1.0.0 forbids embedded secrets. The
server authenticates with an API key you add in your client's MCP settings (see
Keys and security). Verify the package yourself:
A key is shown once, at creation. Keep it in an environment variable and out of version control.
A leaked key dies the moment you revoke it in the dashboard.
An account holds up to 20 keys, so give every machine its own.
Prefer to stay in the chat? An agent with the companion
youtube-skills installed can open the account
and mint the key for you, by email and a 6-digit code, no browser involved.
There is no OAuth flow yet, so clients whose connectors cannot send a custom header (Claude
Desktop and Claude Web) should go through Claude Code for now.
🐳 Run it locally
The hosted endpoint needs no install, but stdio-only clients, sandboxes and container platforms
sometimes want a process of their own. The repo carries one: server.js is a complete local MCP
server (official SDK, stdio transport) whose 14 tools each make one HTTPS call to the TranscriptOut
REST API — the same shape as any SaaS-backed MCP server.
# as a container
docker build -t transcriptout-mcp https://github.com/artemchuikin/youtube-mcp.git
docker run -i -e TRANSCRIPTOUT_API_KEY=sk_your_key transcriptout-mcp
# or straight from a checkout (Node 20+)
npm install && TRANSCRIPTOUT_API_KEY=sk_your_key node server.js
Without a key it still connects and lists all tools; tool calls answer with a clear 401 that says
where to get one. Tool definitions ship in tools.json and refresh from the live catalog at
startup when the network allows, so the local list never goes stale.
🍳 Recipes
Every prompt below is paste-able as written.
Use Case
Example Prompt
📝 Summarize a video
"Summarize the key points from this video: [URL]"
🔍 Research a topic
"Search YouTube for the 5 most-watched videos on neural radiance fields and summarize each."
🧠 Study notes
"Create study notes from this MIT lecture series playlist: [PLAYLIST URL]"
⚖️ Compare perspectives
"Compare arguments in these two videos: [URL1] [URL2]"
🌐 Translate
"Translate this video's transcript to Spanish: [URL]"
✍️ Repurpose content
"Turn this video into a 1,500-word blog post: [URL]"
📡 Monitor a creator
"Each morning, list new uploads from @kurzgesagt and tell me which to watch."
🏛️ Build a content database
"Pull every video ID from @3blue1brown and queue a transcript batch for all of them."
🎯 Competitor analysis
"Search inside @fireship for any video about [competitor product] and summarize the takeaways."
🧩 RAG ingestion
"Fetch this playlist's transcripts as JSON with segment=1000 and load them into the index."
The bulk recipe spelled out. "Archive a whole channel" is four tool calls, not a script:
list_channel_videos with ids_only=true: up to 500 video IDs per page
submit_transcripts_job with those IDs (up to 4,000, duplicates dropped before billing,
idempotency_key makes a retry free)
get_transcripts_job until status is done. The job paces itself inside your rate limit
get_transcripts_results page by page, readable while the job still runs
Anything the service fails to deliver is refunded per video, so the bill matches the archive.
💳 Pricing and limits
Plan
Price
Credits
Rate Limit
Free
$0
100 on signup (one-time)
200 req/min
Starter
$4.49/month
1,000/month
200 req/min
Starter Annual
$45.29/year (~$3.77/mo)
1,000/month
200 req/min
Scale
slider up to $198.99/mo
up to 100,000/month
200 req/min
Subscriptions are a slider from 1,000 to 100,000 credits/month in steps of 1,000, and the per-1,000
rate falls with volume (10,000/mo is $27.49, not $44.90). The annual discount grows with volume,
from ~16% to ~35%.
1 credit = 1 answered request. Calls that fail before reaching YouTube (validation, rate
limit, our capacity) are refunded automatically. The running balance rides in the
X-Credits-Remaining header.
One-time credit packs that never expire can be bought on top of an active subscription.
404: the video has no captions on the requested language/track, or the ID is wrong. This is a
definitive answer, retrying won't change it.
410: the video was removed.
451: age-restricted or members-only content.
Respect the Retry-After header. Both are refunded automatically
For bulk work use submit_transcripts_job: it paces itself inside your rate limit instead of
bouncing off it
Every error body is {"ok": false, "code": "...", "detail": "...", "request_id": "req_..."}.
Branch on the machine-readable code, not on the human text. Include request_id when contacting
support.
🌐 Prefer plain REST?
Building an app instead of an agent? The same backend ships as a JSON REST API, with the same
five transcript formats plus raw file download (download=true).
TranscriptOut is an independent service and is not affiliated with, endorsed by, or sponsored by
YouTube or Google LLC. "YouTube" is a trademark of Google LLC.
Chưa có danh sách công cụ. Máy chủ từ xa sẽ có danh sách sau lần kiểm tra tình trạng đầu tiên.
Câu hỏi thường gặp
Câu hỏi
Làm thế nào để cài đặt YouTube Transcript & Search MCP Server?
Sao chép cấu hình http cho Cursor, Claude Desktop, Claude Code hoặc VS Code trên trang này. Cấu hình dùng bản cài đặt của nhà phát hành, không phải proxy của PlugMCP.
YouTube Transcript & Search MCP Server có cần khóa API không?
Bản ghi này không liệt kê biến môi trường bắt buộc nào.
YouTube Transcript & Search MCP Server dùng giao thức truyền tải nào?
YouTube Transcript & Search MCP Server được truy cập qua HTTP.
YouTube Transcript & Search MCP Server được kiểm tra lần cuối khi nào?
Kiểm tra lần cuối: 2026-11-08. YouTube transcripts, search, channels, playlists and bulk transcript jobs for AI agents. 14 tools.