tube-bridge MCP¶

Self-hosted YouTube research for AI agents - transcripts, frames, and private semantic corpora on your own machine. tube-bridge is a 17-tool MCP server that turns YouTube into a queryable research surface: search videos and channels, read transcripts (optionally with [MM:SS] timestamps), pull comments and playlists, extract one ephemeral JPEG frame at an exact millisecond, and build private semantic-search corpora over collected transcripts. Fourteen tools need no YouTube API key; only comments and channel-search tools require an optional user-supplied key.

Server type: Self-hosted stdio (uvx) - optional self-hosted Streamable HTTP or Docker
Install: uvx tube-bridge
Tools: 17 (14 keyless, 3 optional YouTube API key)
License: MIT
Registry: io.github.TheWhiteWater/tube-bridge (PyPI v1.1.6, active)
Built by: TheWhiteWater

Why This Matters for Operators¶

Competitor videos, founder interviews, and channel research are a material research source that most agents handle badly - transcripts pulled ad hoc, no way to search across them, no way to cite the exact moment. tube-bridge fixes the workflow end to end: youtube_search finds candidates, youtube_get_transcript pulls text with timestamps, corpus_add chunks and locally embeds it (80-second windows, 20-second overlap, BGE-small-en-v1.5 via fastembed), and corpus_search returns similarity scores with canonical video URLs and timestamped deep links. youtube_get_frame returns a bounded JPEG near any transcript finding so an agent can verify visually that the quote matches the frame.

The privacy posture is the differentiator: transcripts, SQLite databases, vectors, and indexes stay on the operator's machine - there is no hosted intermediary, no account, and no managed storage. Everything is local-first and MIT-licensed.

Tools & Capabilities¶

Capability	Tools	YouTube API key
Discovery	`youtube_search`, `youtube_get_video_info`, `youtube_get_trending`	Optional (yt-dlp fallback without)
Content	`youtube_get_channel_videos`, `youtube_get_playlist`, `youtube_get_transcript`, `youtube_get_available_languages`	No
Visual evidence	`youtube_get_frame` (ephemeral JPEG at exact millisecond, needs ffmpeg)	No
Engagement	`youtube_get_comments`, `youtube_search_channels`, `youtube_get_channel_info`	Required
Local corpus	`corpus_create`, `corpus_add`, `corpus_search`, `corpus_list`, `corpus_delete`	No
Runtime docs	`tube_bridge_help`	No

Corpus results return similarity score, time span, video title, canonical URL, and timestamp URL - built for citation, not just answers.

Installation¶

uvx tube-bridge

{
  "mcpServers": {
    "tube-bridge": {
      "command": "uvx",
      "args": ["tube-bridge"]
    }
  }
}

claude mcp add --scope user tube-bridge -- uvx tube-bridge

Requires Python 3.12+. ffmpeg is only needed for youtube_get_frame; the first embedding operation downloads the local model. Alternatives: pip install tube-bridge for a persistent install (stdio, or tube-bridge --http for Streamable HTTP on port 8080), or docker run -p 8080:8080 ghcr.io/thewhitewater/tube-bridge:latest with /health and /mcp endpoints.

Configuration¶

export YOUTUBE_API_KEY="your-key"        # optional - unlocks comments + channel tools
export TUBE_BRIDGE_PROXY="http://..."    # optional - routes yt-dlp/transcript calls via proxy
export TUBE_BRIDGE_CACHE="/path/to/data" # optional - moves corpus and cache databases
export TUBE_BRIDGE_AUTH_KEY="long-random" # optional - Bearer-protects self-hosted HTTP routes

A YouTube Data API v3 key (Google Cloud Console) unlocks the three API-only tools and upgrades search reliability. Corpus storage is SQLite plus sqlite-vec under ~/.tube_bridge/corpus.db; embedding inference runs locally after a one-time model download.

Business Relevance¶

Market and competitor research - search competitor channels, corpus their uploads, ask questions across months of content with timestamp citations
Interview and earnings-call mining - transcripts with [MM:SS] links make every finding quotable with a deep link
Content teams - frame-level visual verification without building a media library
Privacy-conscious operators - all research artifacts stay on-premise; nothing is shipped to a vendor

Integration with CorpusIQ¶

tube-bridge is the research front-end; CorpusIQ is the business record layer. An agent can research a competitor's YouTube presence with tube-bridge - corpus their uploads, extract pricing mentions and product claims with timestamp URLs - then verify the claims against CorpusIQ's connectors: their web analytics via GA4/Semrush, their company record via HubSpot or QuickBooks-side vendor data, and the outreach follow-up via email. The timestamped transcript citations give the "where did we see this" audit trail that pairs with CorpusIQ's source-declared reporting.

Limitations¶

Self-hosted only - no hosted endpoint or managed storage is provided
New listing (Aug 2026), zero-star repository, single maintainer
youtube_get_frame requires ffmpeg and re-downloads a temporary section per call
ToS posture: pulls public YouTube content via yt-dlp and youtube-transcript-api; operators should respect YouTube's terms for commercial use