Self-hostable, OpenAI-compatible LLM proxy: routes to the cheapest healthy provider, caches responses exactly and semantically, tracks cost and latency per call.