AI agents are here, and they're hungry for context. But if you just throw your massive API at them with hundreds of endpoints, inconsistent patterns, and cryptic parameters, the agent will stumble, burn tokens, and make mistakes.

We ran into this problem as we designed the Square MCP server. With a very intentional approach, we managed to cover 30 APIs and 200 endpoints… with just three tools.

In this talk, I'll show you the layered design pattern we used to keep our tools structured, how we avoided "tool bloat" while still covering a large API surface, and how we baked in safety, auditability, and recovery from the start.

If your API feels too big or messy for agents to handle, this session will show you how to build a lightweight, agent-friendly MCP server without rewriting your APIs.