Smartsheet, an enterprise work management platform, has developed a remote Model Context Protocol (MCP) server on AWS to enable AI agents to access its data and capabilities directly. This server allows AI assistants like Amazon Quick and Claude Desktop to interact with Smartsheet’s features through natural language, performing tasks such as analyzing project data, updating tasks, and managing workspaces. Enterprises are also creating custom AI agents that operate autonomously, coordinating through Smartsheet using MCP. These agents handle tasks like capturing requirements and attaching test results, streamlining workflows that previously took weeks into days or hours. The MCP server integrates with Smartsheet’s existing APIs and central intelligence layer, adding an AI-optimized interface to reduce token costs and prevent hallucination. The server also ensures reliable interactions between large language models and enterprise data. Source: awsml
The MCP server architecture on AWS includes AWS Fargate for stateless server containers, Amazon Kinesis Data Streams and Amazon Managed Service for Apache Flink for change-event ingestion into Amazon S3, and Amazon Bedrock and Amazon Neptune for LLM inference and knowledge graph capabilities. AI clients send requests through an API gateway layer before reaching the MCP server on AWS Fargate. The server then interacts with Smartsheet’s domain services and intelligence layer built on Amazon Neptune and Databricks for cross-project insights. Change events stream through Kinesis and Flink into the S3-backed intelligence layer, which follows a medallion architecture. The architecture ensures that all services, including edge protection, observability, and secrets management, are integrated into the system. Source: awsml
Smartsheet designed the MCP server to handle the unique scaling demands of AI traffic, which includes sudden spikes and sustained throughput. The server runs on AWS Fargate for Amazon ECS with Auto Scaling policies that combine traffic volume with compute utilization. Load testing under production-like traffic patterns confirmed the infrastructure can absorb agent bursts without performance degradation. Deployments use Amazon ECR and a CI/CD pipeline with safety nets to ensure updates do not disrupt active agent sessions. Rollouts start in the smallest region to minimize impact, followed by automated end-to-end tests and canary tests every 15 minutes. Source: awsml
