AWS Teams Detail Serverless MCP Architecture on ECS Fargate for Data Lakes
AI assistants · Enterprise AI
Amazon teams supporting the North America Sort Center deployed a serverless Model Context Protocol (MCP) gateway on Amazon ECS Fargate, providing AI agents with unified, configuration-driven access to enterprise data lakes.
Engineers supporting Amazon's North America Sort Center (NASC) built a serverless Model Context Protocol (MCP) server architecture on Amazon ECS Fargate to decouple AI agents from direct database integrations. The design centralizes data lake access across Amazon Athena and Amazon Redshift, replacing per-agent integration code with metadata-driven configuration files discovered automatically during pipeline syncs.
Key takeaways
- The architecture separates operations into client and MCP server layers on ECS Fargate, allowing MCP-compatible clients like IDE plugins or web chats to query tools via FastMCP JSON-RPC endpoints.
- Onboarding new datasets requires only a JSON metadata file detailing schema, business descriptions, and sample values, which AWS Lambda and Bedrock Knowledge Bases index as vector embeddings.
- Athena query tools restrict operations to read-only SELECT calls, and centralized Bedrock Guardrails enforce prompt safety policies across all connecting agents.
Source：AWS Machine Learning Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.