Give your agents one controlled doorway to business data

When several agents need the same business data, separate connectors can become a maintenance problem. A schema change has to be fixed repeatedly, and permission checks can drift between implementations.

AWS describes an alternative: a shared MCP server between agents and a data lake. Its example runs on ECS Fargate and uses metadata to describe datasets, rather than requiring new integration code for each one. The useful lesson is the separation of responsibilities, not that every small project needs this infrastructure.

Start with one recurring question and one approved dataset. Document its meaning, allowed users and permitted operations. Keep authorization at the service that executes the request, rather than relying on an agent’s promise to obey. Test both an allowed request and a denied request before adding another client.

For a small business, a narrow read-only tool may be enough. Add a shared gateway when duplicated integrations become a real cost. Measure correctness, access enforcement and maintenance effort before pursuing scale.