cliver-core has no config files. Every model is a ModelBinding; ModelBindings groups them by purpose. The only required purpose is llm — the chat/stream model.

from cliver_core import ModelBindings, ModelBinding, presets

bindings = ModelBindings(
    llm=presets.deepseek(api_key="..."),
    image=ModelBinding.image(
        model="image-01",
        api_key="...",
        base_url="https://api.example.com/v1",
    ),
)

Purpose Factories

Factory	Purpose
ModelBinding.llm(...)	Chat and streaming
ModelBinding.image(...)	Image generation
ModelBinding.video(...)	Video generation
ModelBinding.tts(...)	Text-to-speech
ModelBinding.asr(...)	Speech-to-text
ModelBinding.embed(...)	Embeddings
ModelBinding.rerank(...)	Reranking

All factories share the same endpoint fields: model, api_key, base_url, protocol ("openai" by default, or "anthropic"), provider_name, and provider (supply an existing provider instance). Media factories also accept path= to override the endpoint path.

With an exact provider_name registry hit, the default base_url comes from the registry. With no recognized brand, base_url is required — there is no implicit endpoint. Anthropic-protocol endpoints must set protocol="anthropic".

Default Request Options

Keyword arguments beyond the shared fields become the binding’s default call options. Per-call options= in chat()/stream() overlay them — the call site wins on conflict.

llm = ModelBinding.llm(
    model="deepseek-v4-flash",
    api_key="...",
    provider_name="deepseek",
    temperature=0.2,
    max_tokens=4096,
)

Lifecycle

A binding builds its provider lazily and reuses it. await bindings.aclose() closes providers the bindings built; a provider you supplied yourself stays open.