Wrap LLM API calls with a single block to capture token counts, latency, and cost per call — attributed by domain and tagged with arbitrary metadata. Persists to your own Postgres database. No SaaS required.