Execution path routing for AI agents. Kalibr automatically picks the cheapest model that works for each task, and reroutes before failures reach your users.
How you integrate Kalibr depends on who is making the routing decisions. Pick the one that matches how your system actually works.
Most agents default to one model for everything. That model is usually expensive and chosen before anyone knows which tasks it will actually handle. Kalibr changes this: it selects the model for each call based on outcome data, evaluates whether the output actually succeeded, and reroutes automatically when it doesn't.
When a provider degrades silently, Gate 1 catches it — the output fails structural validation, the failure is recorded, and the next call routes to a different model. No alert. No rollback. No human required. The dashboard calls this a heal.
This works for text LLMs, voice (TTS and STT), image generation, embeddings, classification, and any model on HuggingFace.
What Kalibr is not: observability (Langfuse, Arize), a model gateway router (LiteLLM, OpenRouter), or a prompt optimizer. It never reads or modifies prompt content. Model calls go directly to the provider.