Skip to main content
Superwire workflows are built from declaration blocks and references between those blocks. 
input {
    topic: string
}

secrets {
    api_key: string
}

provider llm from openai {
    endpoint: "https://api.openai.com/v1"
    api_key: secrets.api_key
}

model fast from llm {
    id: "gpt-4.1-mini"
}

agent summarize {
    model: model.fast
    instruction: "Summarize {{ input.topic }}."

    output {
        summary: string
    }
}

​
Declaration blocks

Use these blocks to define what a workflow needs, how it runs, and what it returns.
BlockWhat it definesWhy it exists
input { ... }Public request data for the workflow.Makes runtime inputs explicit and typed.
secrets { ... }Sensitive request data such as API keys.Keeps credentials separate from normal input.
schema name { ... }Reusable structured types.Avoids repeating the same object shapes across blocks.
provider name from openai { ... }A configured backend endpoint and auth settings.Centralizes provider connection details for models.
model name from provider { ... }A named model profile (id, inference defaults).Lets multiple agents share model settings.
mcp name { ... }An MCP server connection.Declares external capability endpoints once.
tool local_name from mcp.server.tool.remote_nameA tool imported from an MCP server.Exposes deterministic actions to agents.
resource local_name from mcp.server.resource.remote_nameA resource imported from an MCP server.Exposes read-only external content.
prompt local_name from mcp.server.prompt.remote_nameA prompt template imported from MCP.Reuses centrally managed prompt assets.
dynamic { ... }Computed values derived from references and expressions.Prepares runtime values before agent execution.
agent name { ... }One model-backed execution step.Encapsulates instruction, file attachments, tools, context, and typed output.
output { ... }Final response payload of the workflow.Defines exactly what /execute returns.

​
Typical flow

Most workflows follow this shape: 
input + secrets -> provider -> model -> agent(s) -> output
When MCP is used, tools/resources/prompts become additional inputs available to agents.

​
Common expressions

"text"
123
true
input.topic
secrets.api_key
agent.summarize.summary
schema.shared.status
call tool.fetch_context { bindings { task_id: input.task_id } }
read resource.project_readme
render prompt.instructions
context: agent.investigate_task
References are how data moves through the workflow. For example, agent.summarize.summary uses a previous agent result, while secrets.api_key injects a credential into provider config.

​
Naming

Prefer lowercase snake_case for schemas, agents, tools, resources, prompts, providers, models, and fields.