Smartsheet's Enterprise MCP Architecture

AWS ML Blog · rss · 2026-07-18

Smartsheet built a remote **MCP server** on AWS, allowing AI clients and its own smart assistant to access work management data and capabilities via a unified tool layer. The article focuses not on "concept introduction," but on how to make MCP a deployable enterprise-grade infrastructure. ### Core Capabilities - Enables AI assistants to analyze project data, update tasks, create sheets, and manage workspaces via natural language - Allows enterprise-built agents to execute longer-chain workflows, such as gathering requirements, claiming tasks, attaching test results, and generating docs - The same MCP layer serves both internal and external agents, achieving "build once, reuse across all clients" ### Architecture & Optimization - Compute layer uses **AWS Fargate / ECS** to host stateless containers - Event streaming uses **Kinesis + Flink + S3** to feed into the intelligence layer - Uses **Bedrock + Neptune + Databricks** for inference and cross-project insights - Ensures security via API Gateway, WAF, Shield, OAuth, mTLS, etc. - Implemented AI-friendly serialization and interface optimizations for agent access patterns, claiming to have saved **30 亿+ tokens** since launch ### Operations & Governance - Elastic auto-scaling and load testing based on bursty agent traffic - Uses ECS deployment circuit breakers, regional canary deployments, and a 15-minute canary workflow to check release risks - Logs, traces, metrics, and audit info are recorded for every tool call to facilitate compliance and troubleshooting

Original post →

More from coding & agent

coding & agent channel →