MCP and Tools Overview

mcp server -> import capability -> scope with uses -> execute in workflow

​

Capability types

• tool: Callable operation, often used for fetch/create/update behavior.
• resource: Read-only content source.
• prompt: Reusable prompt template managed by an MCP server.

​

Two execution modes

1. Workflow-owned call in dynamic {} using call tool.*.
   • Best for deterministic steps that must always run.
   • Best when runtime values should be bound explicitly.
2. Model-owned call inside an agent through uses: [tool.*].
   • Best when model judgment decides if a tool is needed.
   • Best for selective retrieval or exploratory reasoning.

​

Recommended authoring pattern

1. Declare MCP servers with endpoint and auth headers.
2. Import only needed tools/resources/prompts.
3. Use workflow-owned calls for required operations.
4. Give agents least-privilege uses access.
5. Keep outputs typed so downstream steps stay stable.

​

Example end-to-end shape

input {
    task_id: number
}

secrets {
    mcp_token: string
}

mcp tasks {
    endpoint: "https://tools.internal.example.com/mcp"

    headers {
        Authorization: "Bearer {{ secrets.mcp_token }}"
    }
}

tool fetch_task from mcp.tasks.tool.fetch_task
resource writing_style from mcp.tasks.resource.writing_style

dynamic {
    task: call tool.fetch_task {
        bindings {
            task_id: input.task_id
        }
    }
}

agent writer {
    model: model.fast
    uses: [resource.writing_style]
    instruction: "Write a concise update for {{ dynamic.task.title }}."

    output {
        update: string
    }
}