youtube-context-mcp

An MCP server that gives agents rich context about any YouTube video — what's said, what's popular, and what's on screen:

• Transcript — plain or [mm:ss]-timestamped text, with optional translation
• Deep links — watch?v=…&t=… URLs that jump straight to a moment
• Metadata — title, channel, upload date, duration, view/like counts, chapters, tags
• Most-replayed moments — the peaks of YouTube's viewer-interest heatmap
• Visuals (for multimodal models) — a still frame at any moment, or a tiled preview sheet of the whole video

Agents use it to answer questions about a video, summarize it, pull quotes, surface highlights, read what's on screen, or link you to exactly where something is said.

It builds on youtube-transcript-api (transcripts) and yt-dlp (metadata), which do the actual fetching, and shapes them into a focused set of MCP tools designed for agents.

Transcripts are a video's existing captions/subtitles — it does not transcribe audio (no Whisper/ASR). Videos without captions have no transcript to return.

Install

Run it on demand with uv (no install needed):

uvx youtube-context-mcp@latest

Or install it:

pip install youtube-context-mcp

Use it with an agent

Add it to your MCP client config:

{
  "mcpServers": {
    "youtube-context": {
      "command": "uvx",
      "args": ["youtube-context-mcp@latest"]
    }
  }
}

Or, in Claude Code:

claude mcp add youtube-context -- uvx youtube-context-mcp@latest

Running over HTTP

By default the server talks stdio (the client launches it). If your client runs on a different host — for example LM Studio on Windows while this server runs in WSL2 — run it as a long-lived HTTP server instead and point the client at a URL:

youtube-context-mcp --transport http --host 0.0.0.0 --port 8000

Then add it by URL:

{
  "mcpServers": {
    "youtube-context": { "url": "http://localhost:8000/mcp" }
  }
}

(--host 0.0.0.0 makes it reachable from the Windows side; WSL2 forwards localhost.)

Tools

In every tool, video accepts a full YouTube URL (watch, youtu.be, shorts, embed, live) or a bare 11-character video ID.

get_transcript(video, languages=["en"], include_timestamps=False, translate_to=None)

Returns the transcript as plain text. Set include_timestamps=True to group it into ~15s [mm:ss] blocks — handy for locating a topic and building a link; translate_to takes an ISO language code.

build_video_link(video, start)

Builds a watch?v=…&t=<seconds> URL that opens the video at a moment, so a user can click straight to it. start is seconds or a "mm:ss" / "h:mm:ss" string. Pairs with get_transcript(include_timestamps=True) to turn "where is X mentioned?" into a clickable link.

list_transcripts(video)

Lists available transcripts (language, code, manual vs auto-generated, translatable) plus the translation targets. Use it when get_transcript can't find your language.

get_video_metadata(video, include_description=False)

Returns the video's title, channel, upload date, duration, view/like counts, chapters and tags — answers "what's this video / who made it?" without fetching the transcript. Set include_description=True to also include the (often long) description.

get_most_replayed(video, top_n=8)

Returns the video's most-replayed moments (YouTube's viewer-interest heatmap) as up to top_n distinct high-interest content regions — each with a peak_label/region_label (mm:ss), a clickable jump url, the chapter it falls in, and a relative_intensity (0–1 within the video, 1.0 = its single most-rewatched moment — not a view count). Use it for "what are the best parts?" or to weight a summary by what viewers actually rewatch.

A peak flagged is_opening: true sits at the start (t≈0) — usually a playback-start artifact, not a real rewatch; it's returned in addition to top_n so it can't crowd out the content peaks. has_data is false when YouTube has no heatmap (common for newer/low-traffic videos and some Shorts).

get_video_frame(video, at, max_width=640)

Captures a single still frame (screenshot) at a moment and returns it as an image, so a multimodal model can answer "what's on screen here?" — read a slide, a burned-in caption, or a UI being demoed. at is seconds or a "mm:ss" / "h:mm:ss" string; the frame is the nearest keyframe at/just before it and is downscaled to max_width (clamped 64–1280) to stay cheap on the image budget. Pairs with get_most_replayed / get_transcript(include_timestamps=True) to pick the moment first. Requires ffmpeg.

get_video_preview(video, tiles=12, tile_width=320, start=None, end=None)

Returns a contact sheet: tiles frames sampled evenly across the whole video, composited into one tiled grid image, plus a text legend mapping each tile to its mm:ss timestamp — a cheap visual overview of the entire video (a 4×3 sheet costs about 3× one frame). Pass start / end (seconds or "mm:ss" / "h:mm:ss") to preview just a window of the video — e.g. zoom into the part an overview sheet or get_most_replayed flagged as interesting. Tiles are small by design: spot scene changes and pick a moment here, then zoom in further with get_video_frame at a tile's timestamp. Requires ffmpeg.

Media tools (ffmpeg)

get_video_frame and get_video_preview are the only tools that need an ffmpeg binary (everything else is metadata/transcript only). yt-dlp resolves the video stream and ffmpeg grabs downscaled JPEGs; images come back as MCP image content, so the client/model must be able to display images (e.g. a multimodal model such as Gemma in LM Studio).

The system ffmpeg is used when it's on PATH. The optional media extra installs imageio-ffmpeg, whose wheel bundles a static ffmpeg, making the install self-contained:

pip install "youtube-context-mcp[media]"   # or: uvx youtube-context-mcp[media]@latest
# or bring your own: apt install ffmpeg | brew install ffmpeg | see ffmpeg.org

If neither is available, only the two media tools are affected and they return a clear error.

Proxies (optional)

YouTube blocks most datacenter/cloud IPs, so on a server you may hit RequestBlocked / IpBlocked (transcripts) or a "Sign in to confirm you're not a bot" block (metadata). Locally this is rarely needed. The same env vars route both transcript and metadata requests through a proxy:

With no env set, requests go out directly.

Troubleshooting

• RequestBlocked / IpBlocked — YouTube blocked the IP. Set the proxy env vars above.
• No transcript found — call list_transcripts to see which languages exist for that video.
• Transcripts disabled — the uploader turned captions off; nothing can be fetched.
• "ffmpeg is not installed" — the media tools need it; install it or use the [media] extra (Media tools).
• Frame/preview returns but no image shows — the MCP client/model must support image content; text-only setups drop it silently.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run pytest
uv run mcp dev src/youtube_context_mcp/server.py --with-editable .   # interactive inspector

License

MIT

Credits

Transcript fetching is done by youtube-transcript-api by Jonas Depoix, and metadata by yt-dlp. This project is the MCP adapter that wires them together for agents.