Summarizing a single video is useful. But the real power move is using Claude to research a whole topic across many videos — finding the right ones, reading their transcripts, and synthesizing what they collectively say. With the Scribefy MCP server, Claude can do all of that itself, in one conversation.

From one transcript to a research workflow

Most YouTube transcript tools stop at "here's the text of this video." Scribefy's MCP server gives Claude four tools, so it can run an actual research loop:

search_videos — find videos on a topic (free)
get_video_metadata — check a video's length, channel, and caption languages before committing (free)
get_related_videos — branch out from a good video into adjacent ones (free)
extract_transcript — pull the full timestamped transcript (the only step that costs credits; cached pulls are free)

Three of the four are free, so Claude can explore widely and only "pay" for the transcripts you actually want to read.

The search → triage → extract pattern

The workflow that scales is search → triage → extract → synthesize:

Search. Ask Claude to find videos on your topic. It calls search_videos and lists candidates with channel, length, and view count.
Triage. Ask it to narrow down. It uses get_video_metadata to skip the 3-hour streams or the ones without captions — all free.
Extract. Tell it which to read. It calls extract_transcript only on those, so credits go to signal, not noise.
Synthesize. Now Claude has the actual transcripts in context and can compare, summarize, and cite across all of them.

Because search and metadata are free and transcripts are cached, exploring a topic across a dozen videos stays cheap.

Example: research a topic across videos

A single prompt can kick off the whole loop:

Search YouTube for videos explaining vector databases. Pick the 3 most-viewed under 30 minutes that have English captions, read their transcripts, and give me a combined summary of the key concepts — note where the explanations disagree, with timestamps.

Claude will search, check metadata, extract the three transcripts, and hand back a synthesized answer with citations back to specific moments. That's a literature review of a topic, done in one message.

Compare and synthesize

Once Claude has multiple transcripts in context, you can push further:

Find consensus and disagreement: "Where do these three creators agree, and where do they contradict each other?"
Build a resource: "Turn these into a single study guide with sections and timestamped sources."
Spot gaps: "What important sub-topic do none of these videos cover well?"
Track a channel's thesis: start from one video, use related videos to pull the creator's other talks, and ask Claude to trace how their argument evolved.

The timestamps matter here — every claim Claude makes can point back to the exact moment in the source video, so the synthesis stays verifiable. (See summarizing a single video for prompt patterns that carry over.)

Keeping costs predictable

Search + triage are free, so let Claude explore as much as it wants before extracting.
Extraction is 1–8 credits by length, and cached transcripts are free — popular videos and re-reads cost nothing.
Ask Claude to confirm which videos it's about to extract before it does, if you want a credit checkpoint.

See pricing for the credit details, and the developer guide if you'd rather drive the same tools from your own code via the REST API.

Frequently asked questions

Do I need to set anything up?

You need the Scribefy MCP server configured in Claude (Desktop, Cursor, or Windsurf) with an API key from the API + MCP plan. After that, you just talk to Claude.

How many videos can Claude handle at once?

As many as fit in its context window. Long transcripts are large, so for a broad survey, ask for a structured summary of each first, then a combined synthesis — rather than dumping ten full transcripts at once.

Is the search the same as YouTube's?

search_videos uses the same query syntax as YouTube's own search bar and returns the same kind of results — title, channel, duration, views — so Claude can reason about which to pick.

Ready to turn Claude into a YouTube research assistant? Set up the MCP server, then ask it to research your first topic.