Every model purpose — image, video, text-to-speech, speech-to-text, embed, and rerank — has its own binding in ModelBindings. Capabilities work standalone, with or without AgentCore.

Standalone Calls

from cliver_core import ModelBinding
from cliver_core.capabilities import embed, generate_image

vectors = await embed(
    texts=["hello", "world"],
    binding=ModelBinding.embed(model="text-embedding-3-small", api_key="...", base_url="..."),
)

image = await generate_image(
    prompt="a red panda on a skateboard",
    binding=ModelBinding.image(model="gpt-image-1", api_key="...", base_url="..."),
)

The capability functions are generate_image, generate_video, text_to_speech, transcribe_audio, embed, and rerank.

Through AgentCore

Bind the purpose and call the matching method — image(), video(), tts(), asr(), embed(), or rerank(). Long-running generation also has submit_* variants that return a MediaJob:

from cliver_core import AgentCore, ModelBindings

agent = AgentCore(
    ModelBindings(
        llm=...,
        video=...,
    )
)

job = agent.submit_video("A hummingbird hovering over a flower")
response = await job.result(timeout=600)

The Media Pipeline

Protocol providers share one media pipeline: bindings describe the endpoint (base URL, protocol, and explicit path= where a vendor deviates), and the provider handles request building, payload parsing, and asynchronous task polling. generate_media(...) on a provider is public: pass purpose=, and path= for custom purposes, plus an optional MediaOutputHandler to take over output handling.

Local Models

Embedding and rerank can run in-process with the optional local backends:

from cliver_core.capabilities import embed
from cliver_core.local import LocalEmbeddingProvider, local_embed

embedder = LocalEmbeddingProvider(
    model="BAAI/bge-base-zh-v1.5",
    backend="sentence-transformers",  # required; never inferred
)
vectors = await embed(texts=["hello"], binding=local_embed(provider=embedder))

await embedder.close()

A local provider owns one loaded model and is meant to be long-lived: create it once, share it, and close it at shutdown. There is no process-global model cache.