YouTube Data MCP (kirbah)

Install

Works in

• Claude Desktop
• Cursor
• SourceCraft
• Koda
• Coddy

npx skillfoxx add mcp/youtube-data-mcp-server-kirbah

What it does

YouTube MCP by the author kirbah gives an agent YouTube data in a trimmed form: without nested eTags, duplicate thumbnails and localizations the model does not need. Comparisons of raw and trimmed replies are in the `examples/comparisons` folder. The server installs as the npm package `@kirbah/mcp-youtube` and runs through `npx` over stdio. With no setup it already returns transcripts, and a YouTube Data API v3 key in the `YOUTUBE_API_KEY` variable opens the rest. Nine tools. `getVideoDetails` and `getChannelStatistics` return metadata, statistics and engagement ratios for several videos or channels at once. `searchVideos` searches videos and channels with filters, `getChannelTopVideos` shows a channel's best videos, `getTrendingVideos` and `getVideoCategories` work by region and category, and `getVideoComments` reads comments with a cap on their number and replies. `getTranscripts` returns transcripts of several videos in full or as key intro and outro segments. One video's transcript can also be opened as the resource `youtube://transcript/{videoId}`. The ninth tool, `findConsistentOutlierChannels`, looks for channels whose videos consistently earn more views than they have subscribers. It runs in four phases: candidate search, filtering by channel age, calculating a viral factor for recent videos and ranking. It needs MongoDB. MongoDB can also cache the other replies so that a repeated request spends no quota. For niche analysis there is a ready `analyze-niche` prompt.

Who it is for. For channel creators, marketers and analysts who study YouTube niches, compare channels and go through comments and transcripts in Claude Desktop, Cursor or another MCP client, and for those who protect the model's context and the key's quota.

Example request

Take the transcripts of these three videos, merge the main points and show which of them drew the most comments

MCP

Transport

stdio

Authentication

key is optional

Limitations

• The server is unofficial: YouTube and Google did not release it.
• The Data API quota is 10,000 units a day by default. `getVideoDetails`, `getChannelStatistics` and `getTrendingVideos` cost one unit per call, `getVideoComments` adds one more unit for each top-level comment whose replies are requested. `searchVideos` costs 100 units, `getChannelTopVideos` 101, and `getTranscripts` costs nothing.
• Without a key the server returns only transcripts, and `findConsistentOutlierChannels` is unavailable without a MongoDB connection string.
• By default `getTranscripts` returns key intro and outro segments, and the full text is requested with the `format` parameter set to `full_text`.
• The outlier search tool names its parameters differently in the table and in the deep dive, so check the exact set of fields against the schema in `src/tools`.
• For Windows users whose `npx` does not start, the documentation suggests the `cmd` command with the arguments `/k`, `npx`, `-y`.
• The license is MIT.

How to disable. Remove the youtube entry from the client config and restart the app. Delete the Data API key in the Google Cloud Console if you no longer need it; the cache lives in your MongoDB and is erased together with its database.

README in short

The documentation starts with a comparison of token use on raw and trimmed replies and an explanation of the cache, then comes installation without and with a key, the list of nine tools with parameters and quota cost, a detailed breakdown of outlier channel search, getting a Data API key and system requirements. The repository has tests with coverage, examples of raw and trimmed replies, CI and publishing configuration, an AGENTS.md file and a link to the CreatorLens skill, which builds channel recommendations on this data.