이미지: AWS ML Blog
Summary
- The AWS ML Blog has published a method for building a bridge that lets AI agents running on Bedrock AgentCore access MCP servers on a user's local computer
- It tunnels signed messages over an existing WebSocket connection via a browser extension and Chrome native messaging, requiring no additional open ports or VPN
- An internally built AI assistant for a finance team reportedly handled more than 41,000 conversations in the year since its launch
- 발행처
- AWS ML Blog
- 핵심 기술
- MCP 브리지 (WebSocket + Chrome 네이티브 메시징)
- 호스팅 환경
- Amazon Bedrock AgentCore
- 내부 활용 사례
- 재무용 AI 어시스턴트, 1년간 41,000건 이상 대화
- MCP 표준 도입
- Anthropic이 2024년 11월 발표
AWS has disclosed how it built a bridge that allows AI agents running in the cloud to access MCP (Model Context Protocol) servers located on a user's local computer. The approach addresses the problem of agents deployed on Amazon Bedrock AgentCore needing access to files—such as spreadsheets—that exist only on the local machine.
The architecture consists of four components: the AgentCore runtime, a browser extension, an MCP bridge, and the MCP server. A Strands agent running in the cloud acts as the MCP client and sends tool call requests, while the browser extension relays messages bidirectionally between WebSocket and native messaging. The MCP bridge, which runs on the local machine, is a FastMCP proxy that converts between the native messaging format and raw MCP JSON-RPC messages, communicating with the MCP server via stdio.
A key feature of this approach is that it maintains security by tunneling signed messages over an existing connection, without requiring any additional open ports or VPN configuration. AWS said the pattern is similar in structure to what products like Claude Cowork use, but noted that its version differs in being fully self-hosted on AWS, using its own models and custom tool servers.
Internally, AWS said it built a production-grade AI assistant for finance staff based on this architecture, which has handled more than 41,000 conversations in the year since its launch. What was disclosed in the blog post is a simplified version of that internal system, and the full source code has reportedly been published on GitHub.
