AWS describes a serverless MCP server architecture for AI agent access to the data lake
Amazon Web Services has described a serverless Model Context Protocol server on Amazon ECS Fargate that serves as a centralized gateway between AI agents and the enterprise data lake. According to the company, a new dataset can be added with just a JSON metadata file, without writing any code.
The team supporting Amazon's North America Sort Center (NASC) network has, according to the company, built a serverless Model Context Protocol (MCP) server on Amazon ECS Fargate that functions as a centralized, agent-agnostic gateway between AI agents and the company's enterprise data lake. The goal was to solve the problem where each team integrated data access separately — agents had their own SQL query generation logic, schema knowledge, and access control, so a schema change meant modifying each agent individually, and onboarding a new dataset typically required, according to the company, several days of developer work.
The architecture is divided into four layers — the client layer, MCP server, tools, and the knowledge base update layer — which can be deployed and scaled independently. The metadata catalog contains schema descriptions, column semantics, and relationships between tables; the AI agent draws on this information at query time, so table structures are not hardcoded and a schema change does not require redeployment. According to the company, exposing a new dataset therefore only requires a JSON metadata file, which the server automatically discovers at the next pipeline sync.
Technically, the solution runs on Amazon ECS Fargate, with Amazon Cognito (JWT-based authentication) and AWS WAF on the Application Load Balancer handling security, while sessions are stored in Amazon DynamoDB. The actual querying uses Amazon Bedrock as the AI agent orchestrator with tools built on the FastMCP framework; data queries are processed by Amazon Athena and Amazon Redshift, and semantic search is handled by Amazon Bedrock Knowledge Bases. According to the company, Amazon Bedrock Guardrails is applied to every interaction with the model. The company states that thanks to context mapping, which narrows the search to relevant tables, they managed in testing to keep the overall interaction latency in the range of 15–20 seconds.
Why it matters
The described pattern addresses a specific problem for companies building AI agents on top of their own data: without a centralized gateway such as an MCP server, the number of duplicate integrations grows, access control becomes fragmented, and each new dataset requires its own development. A metadata-driven approach, where the schema is loaded dynamically from a catalog, makes it possible to add data sources without touching the agents' code and without redeployment, shortening onboarding from a matter of days to creating a single metadata file.
Two audiences, two different impacts
What this means
For individuals
Developers working with MCP servers and AI agents get a concrete reference architecture — separating the client layer, MCP server, tool layer, and knowledge layer on Amazon ECS Fargate, with a metadata catalog instead of hardcoded integrations for each dataset.
For a business
Companies dealing with AI agent access to enterprise data can use a centralized MCP gateway as a way to shorten the onboarding of new datasets from several days of development to adding a metadata file, while also providing a single place to enforce access rights and auditing.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.