AWS shows how to run interactive MCP apps inside AI chat hosts
AWS published a walkthrough for building MCP Apps with interactive HTML widgets and deploying them on Amazon Bedrock AgentCore runtime.
AWS has published a technical walkthrough on building MCP Apps, servers that return interactive interfaces instead of plain text inside AI hosts. MCP Apps extends the Model Context Protocol (MCP) with HTML widgets rendered directly inside the host. Amazon Bedrock AgentCore hosts the server.
AgentCore runtime provides a serverless, session-isolated environment with native MCP support, while AgentCore Gateway exposes the server through a single secure endpoint that MCP Apps-compatible hosts can reach.
The sample app
In the sample application, Unicorn Rentals, a user asks "Can you show all unicorns?" and the host renders an interactive card per unicorn with image, name, description, hourly rate and availability. Saying "I would like to book Stardust unicorn" records the booking and returns the booking ID, date and hourly rate. Asking for active bookings returns the rental with its duration and cost so far, this time as text rather than a card, because AWS notes not every request needs a rich interface. Returning the unicorn ends the rental and the total cost is calculated from duration and hourly rate.
Because MCP Apps is a host-agnostic standard, AWS says the same server delivers the same experience in ChatGPT, Claude or any other host supporting the Apps extension.
How it is built
The app is a TypeScript project using the official @modelcontextprotocol/sdk together with the @modelcontextprotocol/ext-apps extension for delivering widgets. It runs as an Express.js HTTP server managed internally by AgentCore runtime. Business logic sits in a dedicated AWS Lambda function, with Amazon DynamoDB for persistence.
Tools are registered through registerAppTool: list_unicorns, book_unicorn, view_bookings and return_unicorn. The _meta.ui.resourceUri field in the tool config tells the host which widget to render, and structuredContent in the response carries the data injected into it. Widgets are registered as MCP resources via registerAppResource. Deployment zips the code, uploads it to an Amazon S3 bucket and creates an AgentCore runtime resource configured with the NODE_22 environment, an entry point and MCP protocol mode.
Full details and code are in the AWS blog post.
