Skip to main content
Superwire separates provider instances and model profiles. 
provider instance -> model profile -> agent

​
Provider declarations

A provider declaration creates a named provider instance. 
secrets {
    api_key: string
}

provider llm from openai {
    endpoint: "https://api.openai.com/v1"
    api_key: secrets.api_key
}
openai is the provider type. llm is the provider instance name. At the moment, the executor supports OpenAI-compatible providers through the /responses API.

​
Model declarations

A model declaration creates a reusable model profile from a provider instance. 
model fast from llm {
    id: "gpt-4.1-mini"
}
The id is the provider-specific model identifier sent to the configured provider. Models that accept non-text inputs should declare those supported asset kinds: 
model vision from llm {
    id: "qwen-plus"
    assets: ["image", "video"]
}
The supported asset kinds are image, video, and document. Superwire validates asset usage against this list before sending a model request. Agents reference model profiles through the model namespace: 
agent reply {
    model: model.fast
    instruction: "Reply to {{ input.message }}"

    output {
        message: string
    }
}

​
Inference defaults

Inference configuration belongs on the model profile when it should apply to every agent using that profile. 
model fast from llm {
    id: "gpt-4.1-mini"

    inference {
        temperature: 0.2
        max_tokens: 4_000
    }
}
An agent can override inference for that specific model usage by opening a block after the model reference: 
agent creative_reply {
    model: model.fast {
        inference {
            temperature: 0.8
        }
    }

    instruction: "Write a creative reply to {{ input.message }}"

    output {
        message: string
    }
}
The usage block specializes the model for that agent. It does not create a new named model profile.

​
Multiple model profiles

Define multiple profiles when different agents should use different model IDs or inference defaults. 
model fast from llm {
    id: "gpt-4.1-mini"
}

model smart from llm {
    id: "gpt-4.1"
}

model cheap from llm {
    id: "gpt-4.1-nano"
}
Then each agent chooses the profile that matches its job: 
agent quick_reply {
    model: model.fast
    instruction: "Reply briefly to {{ input.message }}"

    output {
        message: string
    }
}

agent careful_review {
    model: model.smart
    instruction: "Review this carefully: {{ input.message }}"

    output {
        review: string
    }
}

​
Provider-level defaults

Provider-level defaults are optional and should be reserved for settings that truly apply to all model calls through that provider, such as timeout or retry configuration. 
provider llm from openai {
    endpoint: "https://api.openai.com/v1"
    api_key: secrets.api_key

    inference {
        timeout_seconds: 60
        retry_attempts: 2
    }
}