To enable AI agent integration for invoice card payments in Winvoice, a remote MCP server was built using Amazon Bedrock AgentCore Runtime. This article explains the design process, which adopted a three-tier architecture, leveraged the existing authentication infrastructure, and used VPC isolation and PrivateLink to establish a secure communication path, with an early launch as the objective.
This article describes how Infcurion, which provides the invoice card payment platform Winvoice, came to make a remote MCP server available to enable AI agent integration. To respond to rapid technological changes in generative AI and complete the project in the approximately 1-month period from planning to release, the company reused existing assets and built the architecture primarily around AWS managed services.
The architecture has three tiers: an MCP client tier, a remote MCP server tier running on AWS Bedrock AgentCore Runtime, and the Winvoice backend tier. The MCP server integrates with Winvoice’s authentication infrastructure, while maintaining a design in which people remain the principals responsible for permissions. From a security perspective, VPC isolation and PrivateLink were used to establish a secure communication path.
Among the key design decisions, reliability and operability were prioritized in comparing the local and remote approaches, leading to the adoption of a remote MCP server. ECS and AgentCore Runtime were considered and tested as the execution environments, with AgentCore Runtime ultimately selected. The communication path was configured using VPC network mode and PrivateLink to restrict access to AWS services.
The approach for authentication and authorization was to connect the existing authentication infrastructure without modifying it. Cognito Hub-and-Spoke federation for MCP clients, a token delegation design, and a two-tier allowlist were combined to maintain a design in which people remain the principals responsible for permissions, even when access is made through AI agents. Future plans include the phased opening of the beta version and further functional enhancements.