AWS Shows How to Bridge Cloud AI Agents with Local MCP Tools
AWS has published a guide on building an MCP bridge that allows an AI agent hosted on Amazon Bedrock AgentCore to access MCP tools running locally on a user's machine. The solution uses a Chrome extension and a local bridge to tunnel MCP messages over WebSocket and native messaging. The post includes architecture details, code examples, and security considerations.
AWS published a blog post describing how to build an MCP bridge that gives an AI agent hosted on Amazon Bedrock AgentCore access to local MCP tools. The architecture involves four components: the AgentCore runtime hosting the agent as an MCP client, a Chrome extension providing the chat interface and acting as a bidirectional relay, a local MCP Bridge that translates between native messaging and MCP JSON-RPC, and a local MCP server communicating over stdio. The bridge tunnels MCP messages over WebSocket and native messaging, allowing the cloud agent to call local tools like reading spreadsheets. The post includes details on the message flow, native messaging implementation, and the bridge's internal two-loop design. It also provides prerequisites, deployment steps, testing examples, and security considerations, such as native messaging origin restriction and presigned URL expiration. The solution is self-hosted on AWS with the user's own model, and the complete source code is available on GitHub.
Source: AWS ML blog —
original