Smartsheet deploys remote Model Context Protocol server on AWS to connect AI agents to enterprise data

Smartsheet, an enterprise work management platform, built a remote Model Context Protocol (MCP) server on AWS to give AI agents structured access to its data and capabilities. The MCP server enables AI assistants such as Amazon Q and Claude Desktop to interact with Smartsheet using natural language, performing tasks like analyzing project data, updating tasks, creating sheets, and managing workspaces.

The remote MCP server connects to Smartsheet’s existing APIs and central intelligence layer, adding an AI-optimized interface designed to minimize token cost, reduce hallucinations, and help LLMs work reliably with enterprise data. According to internal telemetry, the optimizations have saved over 3 billion tokens since launch.

The architecture uses AWS Fargate for Amazon Elastic Container Service (Amazon ECS) to run stateless MCP server containers, with Amazon Kinesis Data Streams and Amazon Managed Service for Apache Flink ingesting change events into Amazon Simple Storage Service (Amazon S3). Amazon Bedrock and Amazon Neptune power LLM inference and knowledge graph–based cross-project insights.

Requests flow from AI clients through an API gateway layer that includes AWS WAF, AWS Shield, AWS Application Load Balancer, and OAuth validation before reaching the MCP server on AWS Fargate. The MCP server then calls Smartsheet’s domain services via APIs and queries the intelligence layer built on Amazon Neptune and Databricks for agentic insights.

To handle bursty agent traffic, Smartsheet deployed the MCP server on AWS Fargate with ECS Auto Scaling using target-tracking policies that combine traffic volume and compute utilization. Load testing under production-like patterns validated that the infrastructure absorbs agent bursts without degradation.

Deployments use Amazon Elastic Container Registry (Amazon ECR) and a CI/CD pipeline with circuit breakers that automatically revert to the last stable version if failures are detected. Canary tests run every 15 minutes, executing multi-step MCP workflows through the full authentication and gateway path to validate tool behavior before customer impact.

For governance, the MCP server enforces tiered access control per organization, with options to enable AI access globally, restrict to non-destructive operations, or allow full write and destructive capabilities. Tools carry MCP protocol annotations like readOnlyHint and destructiveHint to guide AI client confirmation flows.

Observability includes OpenTelemetry signals across the request lifecycle, with logs streaming through Amazon Kinesis to Amazon OpenSearch Service and infrastructure metrics surfaced via Amazon CloudWatch. Datadog provides per-tool APM visibility and PagerDuty handles incident routing, while structured analytics events emitted via Amazon SQS feed back into the intelligence layer.