Kalibr lets agents choose the optimal model + tools for any task, across any modality, as they run in production.
Manual integration. Install the SDK, configure your router, report outcomes. ~5 minutes.
Go to Quickstart →Paste one prompt into Claude Code, Cursor, or OpenClaw. Your agent handles credentials, scanning, and instrumentation end-to-end.
Agent Onboarding →Get Kalibr working in 5 minutes.
One prompt. Your coding agent handles the entire integration autonomously.
What is a path? How does routing work? Goals, outcomes, and Thompson Sampling explained.
Proof that Kalibr routes around failures automatically. 88–100% vs 16–36% for hardcoded systems.
Statistical methods, exploration vs exploitation, the trust invariant.
Router, completion(), execute(), report(), get_policy() — full SDK reference.
Graceful degradation, trend monitoring, debugging in production.
CrewAI, LangChain, OpenAI Agents SDK, and HuggingFace.
Why isn't Kalibr routing the way I expect? Common setup issues and questions answered.
Common errors and how to fix them.