How Smartsheet built a remote MCP server on AWS

This article describes how Smartsheet built a remote Model Context Protocol (MCP) server on AWS to enable AI agents and assistants to interact with Smartsheet's work management platform through natural language.

• MCP server runs on AWS Fargate with Amazon ECS, using auto-scaling to handle bursty agent traffic patterns
• Security implemented through AWS WAF, AWS Shield, OAuth2 proxy, and layered rate limiting per user identity
• Architecture includes Amazon Kinesis, Apache Flink, and Amazon Neptune for change-event ingestion and cross-project insights
• Progressive disclosure caps token consumption per response by dynamically sampling data based on column count
• Proprietary serialization format reduces token count by 35-47% compared to JSON for data-heavy responses
• Schema-driven tool contracts with Pydantic models prevent LLM hallucinations and parameter errors
• End-to-end testing includes LLM in the loop to validate realistic business scenarios and agent workflows
• Governance built into tool framework with access tiers, audit trails, and OpenTelemetry observability signals

Smartsheet's MCP server demonstrates how to optimize AI agent infrastructure for enterprise scale, achieving 87% week-over-week user growth post-launch.