AWS Launches Serverless MCP Server on ECS Fargate to Streamline AI Agent Access to Enterprise Data Lakes

Organizations building AI-powered applications often face duplicated effort in granting multiple AI agents access to large enterprise data lakes. Each agent typically implements custom integration, SQL generation, and access controls, resulting in fragmentation and multi-day onboarding for each new dataset.

AWS addresses this challenge through a serverless MCP server deployed on Amazon Elastic Container Service (ECS) with AWS Fargate. This centralized gateway decouples AI agents from the underlying data lake, enabling agent-agnostic access driven by metadata configuration rather than bespoke coding.

The MCP server architecture separates the system into four layers: client, MCP server, tools, and knowledge-base refresh. A metadata catalog stores dataset schemas and semantic descriptions; AI agents reference this catalog at query time to generate accurate SQL without hardcoding table structures.

Client requests flow through secure, scalable components including Route 53, Application Load Balancers, AWS WAF, and Amazon Cognito for authentication. The MCP server layer orchestrates queries, invoking tools like Amazon Athena and Redshift to run validated SQL against data lakes. Session state is persisted in DynamoDB to support stateless scaling.

Adding new datasets involves simply uploading a JSON metadata file describing the schema and semantic context, which the system indexes automatically into the knowledge base. This approach eliminates per-dataset integration code and accelerates onboarding. Policy enforcement—such as read-only SQL queries and content safety—is centralized at the MCP server layer to ensure consistent, auditable permissions.

The serverless infrastructure leverages ECS Fargate to scale seamlessly with fluctuating AI workloads, supporting multiple concurrent agents and tenants. By exposing a standard MCP protocol endpoint, the system enables diverse AI clients to connect uniformly without custom adapters.

While the architecture significantly simplifies data lake access for AI agents and centralizes security, it currently enforces query restrictions like SELECT-only operations, potentially limiting some use cases. Planned enhancements include finer-grained column-level access control to exclude sensitive data and feedback loops for improving dataset metadata based on query outcomes.