SmartsheetがAWS上にリモートMCPサーバーを構築した方法How Smartsheet built a remote MCP server on AWS

Smartsheet, the enterprise work management platform, has published an account of how it built a remote Model Context Protocol (MCP) server on Amazon Web Services, allowing AI agents to interact with its platform's tools and data in a governed, secure way. The write-up matters because MCP has quickly become a de facto interface layer between large language models and external systems, and concrete enterprise deployments that address authentication, scaling, and operational concerns are still relatively scarce. Smartsheet's design is positioned as a practical reference for organizations attempting to move MCP from local experiments to production.

The Model Context Protocol is an open standard, originally introduced by Anthropic in late 2024, that defines how AI applications discover and call external tools, retrieve resources, and pass structured context to a model. Early adoption centered on local MCP servers that run on a developer's machine and communicate over standard input and output, which is convenient for prototyping but poorly suited to multi-user, cloud-based products. A remote MCP server, by contrast, is hosted and reachable over the network, typically using HTTP-based transports such as streamable HTTP or server-sent events. This shift introduces the same requirements any internet-facing service faces: identity, authorization, network security, observability, and horizontal scaling.

According to the account, Smartsheet's implementation wraps its existing platform capabilities—actions such as reading sheets, updating rows, or querying data—as MCP tools that an agent can invoke. Rather than exposing internal APIs directly, the MCP server acts as a controlled façade that translates agent requests into authenticated calls against Smartsheet services. This separation appears intended to keep the agent-facing contract stable while preserving existing access controls and business logic on the backend.

On the AWS side, the architecture reflects patterns common to serverless and container-based API services. A managed entry point, likely fronted by Amazon API Gateway or an application load balancer, handles inbound MCP traffic, while compute is served by AWS Lambda or containers running on Amazon ECS or AWS Fargate depending on session and latency needs. Authentication and authorization are central to the design: remote MCP increasingly relies on OAuth 2.0 flows so that an agent acts on behalf of a specific user with scoped permissions, and services such as Amazon Cognito or an enterprise identity provider are well suited to issuing and validating those tokens. The result is that every tool call can be tied back to a real principal, which is essential for enterprise governance and auditing.

The blog situates the work within the broader Amazon Bedrock ecosystem, which AWS has been building out to support agentic applications. Amazon Bedrock provides access to foundation models, while newer offerings around Bedrock Agents and the recently introduced Bedrock AgentCore aim to give developers managed building blocks for memory, tool use, gateways, and runtime hosting. A remote MCP server complements this by standardizing how those agents reach a specific vendor's tools, meaning the same Smartsheet integration can, in principle, be consumed by agents built on Bedrock or on other frameworks that speak MCP.

Several practical concerns receive attention that are often glossed over in demos. These include managing state across a session when the underlying compute is stateless, controlling which tools are exposed to which callers, rate limiting and cost management, and emitting logs and metrics through services like Amazon CloudWatch for monitoring. Security hardening—least-privilege IAM roles, input validation, and guarding against tool misuse or prompt-driven abuse—also features as a prerequisite for exposing capabilities to autonomous agents.

The context here is a rapidly forming market. Since MCP's release, model providers, developer-tool vendors, and cloud platforms have moved to support it, and remote hosting is widely viewed as the step that makes the protocol viable for software-as-a-service companies. Smartsheet's contribution is less a novel protocol advance than a documented, opinionated blueprint showing how the pieces fit together on one cloud. Readers evaluating similar projects should treat specifics as one company's choices rather than prescriptive requirements, since MCP tooling and AWS's agent services are still evolving. Even so, the example illustrates how an established SaaS provider can expose its functionality to AI agents while retaining the identity, security, and operational controls that enterprise adoption demands.