A keyless MCP server and research skill. It searches in every language, reads only the seconds that matter, sees real frames, and cites every claim with a timestamp.
MIT license · Node 20+ · Claude Code & Codex plugin · Works with any MCP client
0:01
0:03
0:00
0:02
“I do a $106 million launch, and it starts with one sheet of paper.”
“I know for a fact you're wasting half your day on tasks that belong in the trash.”
“In 1989 I was the kid behind the counter of a hobby shop.”
“If you're starting a business this year, watch this before you spend a single dollar.”
Illustrative session with AI-generated frames. Tool names, parameters and output shape are real.
WORKS WITH ANY MCP CLIENT
How it works
Four moves, each one call. The agent never wades through full transcripts it does not need.
Up to 12 queries in one call, each from a different angle (and, when you need it, another language or region). Results are merged and de-duplicated, videos found by several queries rank higher, and the top ones come with the transcript passages that actually match.
[0:00]Four AI agents you can build today that will save you hours of work every week.
[0:08]You don't need any technical skills. If you can type, you can build them.
[0:15]We'll do it together, step by step.
Read a time range, a chapter, or only the passages matching a query, for up to 25 videos at once. Every line keeps its timestamp; long transcripts tell the agent exactly where to continue.
0:01
0:03
0:00
0:02Real frames, up to 1280 px, evenly spaced over a range or at exact seconds. On-screen text, b-roll and screen demos become visible to the agent. Only the bytes around each frame are read, never the whole video.
| Quote | Type | Link |
|---|---|---|
| “This setup saved me 20 hours last week.” | Result first | 0:00 |
| “If your inbox owns your day, this is for you.” | Pain | 0:02 |
| “There's one setting nobody turns on.” | Curiosity gap | 0:05 |
Coverage: 8 queries × 2 languages · 87 unique videos · 21 reviewed · 2 skipped (no captions)
Every row links to the second it came from, quotes are verbatim, and every answer ends with what was searched, reviewed and skipped. Nothing disappears silently.
The harness
The server stays small and stateless. The research method lives in a skill: plan the queries, fan out to parallel subagents, then merge, verify and report coverage.
Search returns the transcript passages that match, so the agent often never opens a transcript at all.
Ranges, chapters and query passages keep context small, even across dozens of videos.
Subagents report videos_reviewed and what they skipped; the skill spot-checks quotes before answering.
The tools
Multi-query, multi-language search with transcript highlights and a coverage footer.
hl/gl{ "queries": ["AI productivity workflow",
{ "query": "productividad con IA", "hl": "es", "gl": "ES" }],
"uploaded": "year", "minViews": 10000 }
Batch-read up to 25 videos: exactly the slice you need, plus real frames.
{ "videos": ["TL8V41Ea6oM", "htZRCE2GgIs"],
"parts": ["transcript", "frames"],
"start": 0, "end": 30, "frames": 4 }
A channel's videos with an outlier score: views divided by the channel's median.
minOutlier{ "channel": "@aliabdaal", "limit": 60,
"minOutlier": 2 }
Checks every path from your machine, with timings, and says what to do when something fails.
✓ search: 5 results (742 ms) ✓ transcript: 61 segments (149 ms) ✓ ffmpeg: ffmpeg version 8.1
Use cases
Collect the opening lines and first frames of the best videos on a topic, typed by hook pattern.
“List 20 hooks from this year's top AI productivity videos, with hook type and what's on screen.”
Find which videos beat a channel's median by 3×, then compare their titles, topics and hooks.
“Which of @channel's videos broke out this year, and what do they have in common?”
Verbatim quotes, claims and advice across many videos, each with the exact second.
“What do creators say is the biggest mistake with AI note-taking? Quote them.”
One topic, written natively in every market's language, merged into one de-duplicated list.
“How is ‘vibe coding’ covered in English, German and Spanish YouTube?”
Frames show the setup, b-roll, on-screen text and pace of cuts: the part transcripts miss.
“How do the top 5 tutorials on Notion AI look in their first 20 seconds?”
Search the comments for a topic, filter by likes, sort by most liked and pull the replies, with a count of how many were scanned.
“Search the comments of these 5 Claude Code tutorials for questions about pricing, most liked first.”
Why it's different
| Typical YouTube MCP | YouTube Research MCP | |
|---|---|---|
| Search | One query per call, a list of links | Up to 12 queries and languages, merged, with transcript highlights |
| Transcripts | The whole transcript, every time | Range, chapter or matching passages, for 25 videos per call |
| Video content | Text only | Real frames at any timestamp |
| What worked | Raw view counts | Views/day and ×median outlier scores |
| Honesty | Failures dropped silently | Coverage footer and per-video reasons on every call |
| Rate limits | Retries until blocked | Separate caption queue, cooldown, exact wait time |
| Method | Left to the agent | Research skill with subagents, spot-checks and citations |
| Setup | Often an API key | No key, no cookies; one plugin install |
Install
Paste one prompt into your agent. It installs the MCP server and the research skill for whatever client it runs in, then checks that YouTube is reachable.
Install YouTube Research MCP for me from https://github.com/jindratilk/youtube-research-mcp. Follow the “Install instructions for AI agents” in its README for the client you are running in: the plugin (MCP server + research skill) in Claude Code or Codex, otherwise the MCP server config. Then run youtube_doctor and tell me whether everything works, and if I need to restart anything.
$ claude plugin marketplace add jindratilk/youtube-research-mcp $ claude plugin install youtube-research@youtube-research
$ codex plugin marketplace add jindratilk/youtube-research-mcp $ codex plugin add youtube-research@youtube-research
{
"mcpServers": {
"youtube": {
"command": "npx",
"args": ["-y", "github:jindratilk/youtube-research-mcp"]
}
}
}Honest limits
YouTube throttles caption downloads per IP. A burst of ~100 in a few minutes triggers a pause. The server stops, says how long to wait and keeps everything else working. For big sweeps, set YT_PROXY.
Search results carry “3 weeks ago”, so date filters are ±1 unit. youtube_fetch returns the exact publish date.
Videos without manual or automatic captions have no transcript. There is no speech-to-text in the server; frames still work.
You can search, filter and sort up to 1,000 top comments per video, with replies. YouTube no longer offers anonymous clients the “newest first” sort, so it is not every comment.
No. It talks to the same internal endpoints the YouTube website and apps use, with no API key, OAuth or cookies. yt-dlp and ffmpeg are optional and only needed for video frames.
Any client that speaks the Model Context Protocol. It ships as a Claude Code plugin and a Codex plugin (server plus skill), and runs as a plain stdio server in Claude Desktop, Cursor, Windsurf, VS Code, Zed and others.
Instructions the agent loads for YouTube research: define the result first, plan queries in every relevant language, fan out to parallel subagents, then merge, spot-check quotes and report coverage. It includes references for hooks (with a 10-type taxonomy), quotes and claims, creator outliers and the final write-up.
No. Transcripts come from caption files. For frames, ffmpeg seeks inside the stream with HTTP range requests and decodes only the frames you ask for.
Caption downloads get their own slower queue. On the first 429 the server pauses transcripts instead of retrying (retries extend the block), tells the agent exactly how many seconds to wait, and keeps search, details, frames and channels working. For very large sweeps, route requests through a proxy with YT_PROXY.
Yes. Each query can carry its own interface language and region, transcripts default to the video's original language, and highlights fold accents and match inflected word forms, so a sweep across several markets lands in one de-duplicated list.
No. It is an independent open-source project. Use it in line with YouTube's terms and respect creators' rights when you quote or reuse their content.
No. It is free and open source under the MIT license. The code is on GitHub.
Free, open source, no API key. Claude Code, Codex and any MCP client.