A contract for language-model backends, the lifecycle that holds it, and the suite that proves yours conforms

Tools a model can call, written against the modelpact contract and held to it by tests

Language-model backends written against the modelpact contract — a daemon, the OpenAI dialect, the browser's own model, one in the tab, each green on the contract suite

Several models in one conversation, routed by a policy: any modelpact providers you name — Ollama, OpenAI, on-device — plus Claude from the terminal you already pay for

A tool-calling agent on storey three: a loop over sessions the library already owns