Too many model context protocol servers and LLM allocations on the dance floor