The sourced knowledge graph approach is a game-changer for reducing token waste. Most AI teams are basically paying for hallucination overhead when they could just curate facts from the start. I like that you're building this for Git versioning too - that's how you maintain governance around what your agents actually know. The MCP integration is smart positioning since agents are becoming the primary UX. Quick question: how does the performance scale when you're querying graphs with thousands of nodes? And for teams building RAG systems, what's the migration path from vector-based retrieval to deterministic knowledge graphs?