It can talk about YouTube. It can guess what's popular. It can hallucinate a "comprehensive overview" of trends it's never actually seen. But it cannot go in, run multiple search queries at once, jump to the exact minute a topic is discussed in a 3-hour lecture, fetch transcripts for many videos in parallel, mine 500 comments, and come back with structured intelligence.

Not without this.

See it in a few seconds

Ask your agent: "Pull the comments from my last 3 uploads, group them by theme, and rank by likes plus reply count so I know what to produce next."

It comes back with comments grouped by theme — requests, complaints, questions — each one tagged with the video it came from and the commenter's name, sorted by engagement. You know exactly what your audience is asking for, and which video to follow up under when you answer them.

That's get_channel_videos + get_video_comments doing structured comment mining across multiple videos in one pass. (Swap in your own channel URL — the agent needs it to find your uploads.)

Ask your agent: "Pull the chapter list of this 2-hour ML lecture, then give me the transcript of just the chapter on backpropagation."

It comes back with the chapter titles, picks the right one, fetches the transcript for that slice only. No 12,000-line dump. No re-watching the video. The agent cites the chapter title and [MM:SS] timestamps for every claim.

That's get_video_chapters + get_chapter_transcript doing windowed transcript extraction. Works on any video. Any chapter. Any language.

What your agent can do the moment you connect

Find the 10 most-viewed videos on any topic from the last 6 months — multi-query parallel search, deduplicated, sorted by views, with inline enrichment (subscriber count, verified badge, live status) and an opt-in enrich=True flag that batches full metadata for the top candidates
Pull full transcripts with timestamps — any available language, grouped by chapters, windowed via start_seconds / end_seconds so the agent fetches just the slice it needs (no burning context on a 3-hour lecture when chapter 2 is all it wants)
Navigate long courses in two calls — fetch the chapter list (small, fast), pick the chapter by 1-based index OR case-insensitive title substring, get just that chapter's text
Ask "what do the top 200 comments say about X?" — returns structured fields (like_count, reply_count, is_pinned, text, author) for up to 500 comments per call, auto-paginated, with get_more_comments for the long tail. The agent analyzes the structured fields directly with its own model.
Mine any comment section for content ideas — "scan these 500 comments and extract every suggestion, request, complaint, or question — group by theme, rank by frequency, surface 2–3 direct quotes per theme as evidence, format as a content calendar brief"
Get the business email, sponsorship status, and all description links of any channel — subscriber count, open_to_sponsorships flag, all description links categorized by platform (Instagram, TikTok, X, personal site, affiliate) — one call, ready for outreach
Batch-fetch metadata for 30 videos in one parallel call — title, description, tags, chapters, engagement metrics
Batch-fetch transcripts in parallel — up to 5 concurrent requests, also supports start_seconds / end_seconds windowing so each call can target the right slice
Discover trending content before it's everywhere — All, Gaming, News, Movies, Sports, Music, Science & Technology — up to 50 videos per call
Get the full content history of a YouTuber — newest, oldest, or client-side re-sorted popular — with continuation tokens for the long tail
Extract any public playlist — courses, conference talks, compilations — with multi-path metadata extraction for the edge cases YouTube's renderer change broke

The gap this fills

YouTube's UI lets you watch videos. It doesn't let you research them at scale.

Manually watching 20 competitor videos to extract positioning takes a full workday. Reading 500 comments to understand what customers actually complain about takes longer. Finding every channel in your niche that has a business email and is open to sponsorships — without this server — means clicking through profiles one by one, forever.

Agents with this MCP do all of it in a single conversation. The server handles provider rotation, retries, residential-proxy routing for the YouTube endpoints that reject cloud IPs, and bounded in-process caching so repeat calls cost nothing. You get clean JSON and markdown.

What this could look like in practice

These aren't case studies. They're the kind of workflows this server is built to make possible — so you can judge whether the math works for your situation.

Competitive intel brief

A founder searches "Linear vs Jira" across the last 90 days, enriches the top 10 with full metadata, then batch_get_transcripts(start_seconds=0, end_seconds=300) for the first 5 minutes of each (where reviews/comparisons state the context). The agent extracts brand mentions, sentiment, and which tool the YouTuber recommends. Output: a positioning table ready for a board meeting, built in under 2 minutes instead of 4 hours of video-watching.

YouTuber deal hunting

A software company's growth team wants to sponsor YouTubers in the productivity or indie hacking space. They search for channels covering "solopreneur tools" and "build in public," pull get_channel_info on the top 30, filter for those with a business email and open_to_sponsorships flag, then get_chapter_transcript on the intro chapter of each (sponsor reads happen early). Output: a ranked shortlist ready to send to the partnerships team. One activated deal at market rates covers the Pro tier for a year.

Investment research

An analyst searches the last 2 weeks of YouTube for earnings commentary on a specific ticker, enriches the top candidates, then fetches the first 5 minutes of each analyst video where they state the thesis (no need to transcribe a 40-minute show — the first 5 minutes is where the call lives). The agent clusters by sentiment, flags outliers, cites the [MM:SS] timestamp for every claim. The same sweep via paid research platforms costs multiples more per month.

Churn research from your customers' own words

A PM searches for "why I switched away from [competitor]" across 10 query variations, fetches the transcripts, scans for switching signals, surfaces direct quotes with timestamps. Output: a churn-reason taxonomy in users' actual language, without a single customer interview. One insight from this can redirect a roadmap.

Long-course audit before you enroll

A learner pulls a 40-lecture playlist, walks the chapter list of the first 3 lectures to extract topic and prerequisites, gets just those sections' transcripts. Verifies difficulty level (beginner / intermediate / advanced) before committing 40 hours. Total transcript bytes burned: a fraction of a full dump per lecture.

14 tools, no API key, no quota, no LLM pass-through cost

Search & Discovery — multi-query search with dedup, inline enrichment, enrich=True opt-in, days_back auto-fallback, trending videos by 7 categories, related-video clusters always enriched with metadata, full metadata on up to 30 videos in one parallel call

Transcripts — full text with [MM:SS] timestamps, multi-language, chapter grouping, time-windowed via start_seconds / end_seconds, batch fetch with up to 5 concurrent requests

Chapter Navigation — chapter list (small payload) + single-chapter transcript fetch by 1-based index or title substring. Designed for long courses and lectures.

Channel Intelligence — subscriber count, business email (when listed), primary social link, all description links categorized by platform, open_to_sponsorships flag, full video history sorted newest / oldest / popular

Comment Mining — up to 500 comments per call with auto-pagination, continuation tokens for the long tail, optional inline replies. The calling agent analyzes the structured fields directly with its own model — no LLM middleman, no extra cost.

Playlists — full playlist extraction with multi-path metadata extraction that handles the renderer changes YouTube ships without notice

Zero LLM pass-through cost. The server is a structured data extractor. The calling agent already has an LLM — it does the synthesis on the raw transcripts and metadata this server returns. No chained providers. No per-call tokens. No surprise bills.

Honestly, what it cannot do

No official like counts — YouTube removed them from public APIs. No historical subscriber growth curves. No real-time alerts or persistent monitoring. No outreach automation — this is research, not a sending tool. No AI-extraction focus parameter — by design, the agent does its own extraction on the structured data we return. No coffee or bagel included either.

What it is: a structured research layer that makes your agent genuinely capable of YouTube intelligence workflows, instead of pretending.

For more case studies, use cases, demos, suggested workflows, visit youtuberesearchmcp.com

YouTube Research

How to pay

Subscribe

Your AI agent is flying blind on YouTube.