
ARBR provides a single control layer for applications that use multiple AI models. By exposing an OpenAI‑compatible endpoint, it routes requests according to task, cost or policy, enforces budgets and guardrails, and records latency, cost and performance data. The tool also supports model evaluation on real workloads and safe deployment of new models. It is open source, MIT‑licensed, provider‑neutral and designed for self‑hosting.