Models¶

In Corollary a model is a stateless proposer. It receives one prompt built by the projector and returns one response in the claim contract. It never sees earlier turns, so an adapter only needs a single-shot completion:

class Model(Protocol):
    @property
    def name(self) -> str: ...  # recorded as the source of its claims
    def complete(self, system: str, prompt: str) -> str: ...

The agent accepts a model instance, or a string:

Agent("anthropic:claude-opus-5-5", ...)
Agent("claude-sonnet-5-5", ...)  # bare claude-* ids select the Anthropic adapter
Agent("openai:<model-id>", ...)

Claude¶

pip install "corollary[anthropic]"

from corollary import AnthropicModel

model = AnthropicModel(
    "claude-opus-5-5",  # default
    max_tokens=16_000,
    effort="high",  # output_config.effort: low, medium, high, xhigh, max
    structured=False,  # constrain output to the contract's JSON schema
    fallbacks="default",  # server-side refusal fallback
    client=None,  # a configured anthropic.Anthropic(), or a platform client
)

Credentials are resolved by the SDK: ANTHROPIC_API_KEY, or a profile from ant auth login.
Effort defaults to "high", because every call carries reasoning the runtime will hold the model to. Lower it for cheaper, simpler tasks.
Fallbacks. With fallbacks="default", the request uses the server-side refusal fallback (beta server-side-fallback-2026-07-01), so a declined request can be served by a fallback model in the same call. This is available on the Claude API only, so it is left out automatically when client is an Amazon Bedrock, Google Vertex AI or Microsoft Foundry client.
Refusals that still happen raise ModelRefusalError, and truncated responses raise ModelError. Agent.run() catches both and returns its report with the error in report.error.
Other platforms. Pass a platform client, for example anthropic.AnthropicBedrockMantle(...), as client=.
model.request(system, prompt) returns the exact keyword arguments that would be sent, which is useful for logging and tests.

OpenAI-compatible endpoints¶

pip install "corollary[openai]"

import openai
from corollary import OpenAIModel

model = OpenAIModel("<model-id>")  # api.openai.com
local = OpenAIModel("llama3", client=openai.OpenAI(base_url="http://localhost:11434/v1", api_key="x"))

OpenAIModel uses the Chat Completions API, so it works with any compatible server (vLLM, Ollama, LM Studio, ...). It requests JSON mode by default. Pass json_mode=False for servers that don't support it. Extra keyword arguments (temperature=0, ...) are forwarded; an explicit response_format replaces JSON mode. Refusals (message.refusal, or finish_reason == "content_filter") raise ModelRefusalError.

Any function¶

from corollary import CallableModel


def my_model(system: str, prompt: str) -> str: ...  # call LiteLLM, a local model, an internal gateway...


agent = Agent(CallableModel(my_model, name="internal-llm"), ...)

Scripted responses, for tests and demos¶

from corollary import ScriptedModel

model = ScriptedModel(
    [
        {"actions": [...]},  # dicts and lists are serialized to JSON
        '{"actions": []}',  # strings are returned as-is
        lambda prompt: {"actions": [...]},  # callables receive the prompt
    ]
)
model.add({"actions": [...]})  # append more responses later
model.calls  # Call(system, prompt, response) for each call
model.remaining  # responses left

ScriptedModel makes agent behavior fully deterministic, which is how Corollary's own test suite covers the agent runtime without a network.

Writing an adapter¶

An adapter should:

Send system as the system prompt and prompt as a single user message. There is no history to manage.
Return the text of the response. Don't parse it: the runtime does that and reports problems back to the model.
Raise ModelError for transport failures and truncated output, and ModelRefusalError for refusals.
Expose a stable name. It is recorded as the source of every claim (model:<name>) and used by the trust policy.

Contributions of adapters for more providers are welcome; see CONTRIBUTING.md.