MCP Night 2.0 Demo Recap: Block's Goose - The Layered Tool Pattern

At MCP Night 2.0, Richard Moot from Block showcased an innovative solution for creating an MCP server for a massive API platform, specifically their Square API ecosystem, which includes over 30 APIs and more than 200 endpoints. Block developed a "layered tool pattern" to streamline the platform into three conceptual layers: Discovery, Planning, and Execution. This approach allows AI agents to dynamically discover available services, plan workflows, and execute API calls without needing to understand hundreds of specific endpoints upfront. Moot demonstrated this with a practical example, where an AI agent named Goose successfully created an invoice through natural language commands by leveraging self-directed exploration and the layered tools. This pattern not only reduces errors and context window clutter but also adapts to API changes without requiring tool updates, offering a scalable solution for complex API integrations. The presentation highlighted that higher-level abstractions and self-discovery processes can improve reliability and flexibility, suggesting the pattern's potential application beyond API platforms.