Skip to main content
Glama
Sign In
deenesjakoruzh
youtube-transcript-mcp

youtube-transcript-mcp

by luigimasango-dev
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Local

youtube-transcript-mcp

MCP server that lets coding agents read, search and chapter-summarize YouTube transcripts. No API key needed.

License: MIT CI Python 3.11

get_video_metadata call and result

Quick start

Requires Python 3.11+ and uv. Tested 2026-09-11 from a fresh clone on Windows 11 (POSIX same, minus py):

git clone https://github.com/luigimasango-dev/youtube-transcript-mcp.git
cd youtube-transcript-mcp
uv sync
uv run server.py

It sits waiting for MCP stdio input, so it will appear to do nothing until a client connects. That is normal. Ctrl+C to exit.

Register it with an MCP client (Claude CLI example):

claude mcp add youtube-transcript -- uv --directory C:\Dev\youtube-transcript-mcp run server.py

Or in JSON config:

{
  "mcpServers": {
    "youtube-transcript": {
      "command": "uv",
      "args": ["--directory", "C:\\Dev\\youtube-transcript-mcp", "run", "server.py"]
    }
  }
}

Related MCP server: YouTube for AI Agents

How it works

This started as a small tool for my own use. I kept hitting conversations that contained a YouTube link and needed to know what the video actually said, and re-pasting transcripts around was wasting time. It became a reusable server so that anything running on the Model Context Protocol could just ask.

Captions come through YouTube's public endpoints; metadata comes through YouTube's no-auth oEmbed endpoint. No API key, no Google account, nothing to sign up for.

Tools

Tool

Purpose

Key args

get_transcript

Full transcript as plain text. Accepts watch, youtu.be, embed, shorts, live URLs or a bare 11-char video ID

video_url_or_id, languages=["en"], include_timestamps=False

get_video_metadata

Title, channel name, channel URL and thumbnail via oEmbed (no upload date or description — those need a Data API key)

video_url_or_id

search_transcript

Segments matching a keyword/phrase, case-insensitive, each with a [mm:ss] timestamp

video_url_or_id, query, languages=["en"]

summarize_chapters

Splits the transcript into rough time-blocked chunks at natural pauses for section-by-section summarising. A heuristic, not YouTube's real chapter data

video_url_or_id, gap_seconds=4.0, min_chunk_seconds=45.0, languages=["en"]

All tools return a plain string, including errors (for example "Error: captions are disabled for this video"). No exceptions surface to the calling agent, and no stack traces leak out.

Limitations

  • Only videos with captions (auto-generated or manual) work. Private videos, caption-less videos and live streams without captions return a clear error instead of a guess.

  • No upload dates or full descriptions without a YouTube Data API key, by design.

  • No caching. Two calls against the same video re-fetch both times. Fine for the tool's actual use; add caching if you ever need it at volume.

  • The language list matters. languages=["en"] gets the English track when one exists. If a channel uploads in another language only, pass that code.

Development

uv sync
uv run --with pytest python -m pytest tests/ -v

The suite launches the real server over stdio, calls tools/list, and asserts the four tool names are present. Pin mcp>=1.28.1 but stay below 2.0 — mcp 2.x renamed mcp.server.fastmcp and the import breaks at server start.

License

MIT. See LICENSE.

Available Tools

4 tools
get_transcriptA

Fetch the full transcript of a YouTube video as plain text.

Args:
    video_url_or_id: Full YouTube URL (watch/youtu.be/embed/shorts) or bare 11-char video ID.
    languages: Preferred language codes in priority order, e.g. ["en", "en-GB"]. Defaults to ["en"].
    include_timestamps: If true, prefix each line with a [mm:ss] timestamp.

Inputs: video_url_or_id (required), languages, include_timestamps. Returns: result

get_video_metadataA

Fetch title, channel name, and thumbnail for a YouTube video via the no-auth oEmbed endpoint.

Note: oEmbed does not expose upload date or full description (those require the official
Data API + an API key). This returns what's available without one.

Args:
    video_url_or_id: Full YouTube URL or bare 11-char video ID.

Inputs: video_url_or_id (required). Returns: result

search_transcriptA

Search a video's transcript for a keyword/phrase and return matching segments with timestamps.

Args:
    video_url_or_id: Full YouTube URL or bare 11-char video ID.
    query: Keyword or phrase to search for (case-insensitive substring match).
    languages: Preferred language codes in priority order. Defaults to ["en"].

Inputs: query (required), video_url_or_id (required), languages. Returns: result

summarize_chaptersA

Break a transcript into rough time-blocked chunks at natural pauses, for easier per-section summarizing.

This is a heuristic, not real chapter data: it splits wherever the gap between two consecutive
caption snippets exceeds `gap_seconds`, then merges any resulting chunk shorter than
`min_chunk_seconds` into its neighbor so chunks stay summarizable.

Args:
    video_url_or_id: Full YouTube URL or bare 11-char video ID.
    gap_seconds: Silence gap (seconds) between captions that triggers a new chunk boundary.
    min_chunk_seconds: Minimum chunk duration; shorter chunks get merged forward.
    languages: Preferred language codes in priority order. Defaults to ["en"].

Inputs: video_url_or_id (required), languages, gap_seconds, min_chunk_seconds. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Deploy Server
    One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

    TDQS

    A4.4/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool targets a distinct operation: fetching raw transcript text, fetching metadata, searching within a transcript, and chunking into chapters. Even though three tools operate on transcripts, their purposes are clearly separated and descriptions make the boundaries obvious.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun snake_case pattern: get_transcript, get_video_metadata, search_transcript, summarize_chapters. The verbs clearly indicate the action, and the nouns indicate the resource or output.

    Tool Count5/5

    With 4 tools, the server is well-scoped for a focused YouTube transcript utility. Each tool covers a meaningful, non-redundant capability without unnecessary bloat.

    Completeness4/5

    The set covers the core transcript workflows: fetching, searching, chunking, and basic metadata retrieval. A minor gap is the lack of a tool to list available transcript languages or caption tracks, which could help when the default language is unavailable.

    Maintenance

    ActivityStale
    ResponsivenessOutdated data

    Resources

    Related MCP Connectors

    Related MCP Servers