The Model Context Protocol (MCP) is a great base protocol that allows AI agents to connect to 3rd party systems, and has gained rapid adoption across the industry. However, it requires designing APIs with specific patterns that make them easy for AI agents to use.

Your organisation likely already has pre-existing APIs with machine readable schemas, like GraphQL or OpenAPI (REST). While they expose useful business data and capabilities, these schemas are not optimised for consumption by AI agents. Still, they have mature frameworks around them and importantly – are already implemented.

In this talk we'll share what we learned at Isometric when building a translation layer between GraphQL and MCP. We optimised for avoiding duplicate work and making the most of existing tooling built into the GraphQL server, while at the same time ensuring that the end result follows all the MCP server design best practices and can be effectively used by AI agents.

Our internal GraphQL server has more than 800 types, 100 top-level query fields and 200 mutations. We'll show how we leveraged GraphQL schema introspection, query templates and complexity heuristics to distill this large API surface into a small, but scalable, set of MCP tools.

You’ll learn about practical strategies for exposing your existing APIs to AI agents without rewriting the entire API layer. We'll discuss specific challenges like handling GraphQL's infinite nesting problem and demonstrate how complex GraphQL queries get transformed into simpler, AI-friendly MCP tools.

Konrad Komorowski is a software engineer at Isometric, working on carbon removal credit verification. He's spent his career between big tech and startups – building probabilistic ads measurement systems as a tech lead at Meta, and helping scale engineering organisations at companies like Glovo from 150 to 500 engineers. He's passionate about finding simple & robust solutions to complex, real-world problems, believing the best engineering happens where theory meets practice.