Developer 0xchamin has released mcptube, an open-source tool that builds a persistent knowledge wiki from YouTube videos using transcripts and vision. Available under the MIT License via pip install mcptube, it turns watched videos into a searchable, cross-referenced knowledge base.
Open Source · MIT License
mcptube-vision turns YouTube videos into a persistent, cross-referenced knowledge wiki
A CLI + MCP server that reads transcripts and video frames, then has an LLM compile them once into a git-diffable wiki — instead of re-searching raw data on every query.
<1ms
Full-text search via SQLite FTS5 index
$0
LLM cost at the search stage — FTS5 is LLM-free
4
Page types built: entity, topic, concept, per-video
From URL to knowledge — the ingest flow
mcptube add <url>
1 · Retrieve
Transcript + scene-change key frames (ffmpeg)
→
2 · Describe
Vision models read slides, code & demos
→
3 · Persist
Interlinked JSON + SQLite, git-diffable
Two ways of handling knowledge
Conventional RAG re-does the work each time; the LLM Wiki approach compiles once and accumulates.
Search + generate again
Search + generate again
Search + generate again
Search + generate again
Conventional RAG
Disposable answers · repeated vector retrieval · LLM cost every query
Compile once
Cheap keyword lookup
LLM Wiki (mcptube)
Persist & accumulate · FTS5 search at zero LLM cost · contradiction detection & incremental updates
WHAT DEVELOPERS PRAISE
Captures slides, code & demos that transcripts miss
Seamless MCP integration with Claude, Cursor, VS Code & more
Great for domain KBs from lectures & technical talks
OPEN QUESTIONS
Relatively new — few large-scale usage examples yet
Token cost of heavy LLM use during ingestion
Playlists, channels & a SaaS version still "Coming Soon"
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…