LangChain’s tool routing is a bloated mess. So we built a <10ms local semantic registry to replace it at US Neural.

US Neural built Mycelium, a decoupled semantic tool registry that replaces LangChain's AgentExecutor for sub-10ms local tool routing. The system achieves 9.6ms cold discovery latency and 70.7% family-level intent matching using all-MiniLM-L6-v2 and ChromaDB, all running 100% locally.

Look, LangChain was great to get the ecosystem started 0 to 1. But let's be honest: taking LangChain to production with 50+ tools is a nightmare of abstraction hell and brittle if/else logic. The AgentExecutor is slow, and relying on it to route complex tools feels like fighting the framework. Tool discovery shouldn't be an LLM's job, and it shouldn't require a heavy python wrapper. It’s a distributed networking problem. What we built at US Neural: We built Mycelium — a decoupled, semantic tool registry. Instead of chaining tools in code, you register them to a local mesh. When a user has a task, the engine routes it semantically no keyword matching . The Benchmarks No BS : We ran this against a synthetic corpus of 100,000 agents to test if local semantic routing is actually viable for production. Latency: 9.6ms cold discovery vs 194ms for BM25 . Accuracy: 70.7% family-level intent matching. Infra: 100% local all-MiniLM-L6-v2 + ChromaDB . If you are tired of bloated frameworks and want deterministic, sub-10ms tool routing, check out our methodology. GitHub & Reproducible Scripts: https://github.com/udaysaai/mycelium https://github.com/udaysaai/mycelium Visual Dashboard: https://mycelium-agents.netlify.app/ https://mycelium-agents.netlify.app/ I’m ready to be roasted. Tell me why this architecture wouldn't scale for your stack.