MCP solved the AI integration problem. Every SaaS vendor ships a server now. Wire
ten together and you get a different problem: auth sprawl, disconnected tool namespaces,
and response shapes each agent handles. The constraint in production AI
isn't connecting to systems. It's managing the surface area.
GraphQL federation resolved this for REST — independent services behind a single
unified schema. The same pattern applies to MCP server sprawl. Platform vendors are
already shipping production gateways that prove it.
This session shows how, live. From three real systems — Neo4j Aura, Linear, and Slack
— we build a federation gateway exposing all three as a single MCP endpoint. We cover
subgraph configuration, persisted operations that govern what an agent can call, and
the graph data model that connects the systems. The context window impact is visible:
20+ discrete tool definitions collapse to a handful of named operations.
The second half shows what a knowledge graph adds. Neo4j isn't a third data source
here — it's the layer that holds the relationships between the other two. We model
Linear issues and Slack threads as nodes, connect them with typed edges, and write
the Cypher that makes cross-system reasoning possible. The closing demo traverses all
three as a connected graph: one auth token, one log entry, one answer none of them
could produce alone.
A federated endpoint is manageable. A knowledge graph makes it intelligent.
Code and data model on GitHub before the conference.
Speaker
Jonathan Giffard
Lead Product Manager, Neo4j
Jonathan Giffard is a Lead Product Manager at Neo4j, focused on developer surfaces—the tools and interfaces that connect developers to the graph. His work spans MCP servers, platform drivers, web services, and the GraphQL library. Originally trained as a software engineer, he jokingly describes his degree as a “Drinker’s Degree.”
Outside of work, Jonathan can usually be found running, riding his Triumph motorcycle, or serving as both financier and chauffeur to his children.
