A production LLM call layer for AI agents and tools: keep your existing SDKs, hot-swap models via MDA presets, and add cache, retries, circuit breakers, and key rotation.