AWS Shows How to Build Interactive MCP Apps on Amazon Bedrock AgentCore

Amazon Web Services published a detailed guide showing developers how to build MCP Apps—interactive HTML widgets that render inside AI chat hosts—using Amazon Bedrock AgentCore. The post addresses a specific problem: as users increasingly interact with services through AI hosts like ChatGPT and Claude, businesses need a way to deliver rich interfaces rather than plain text, without locking into a single host's ecosystem.

The solution combines two technologies. MCP Apps extends the Model Context Protocol (MCP) with a standard for delivering interactive HTML widgets from MCP servers, rendered inside sandboxed iframes within any compatible AI host. Amazon Bedrock AgentCore supplies the infrastructure: AgentCore runtime acts as a secure, serverless, session-isolated host with native MCP support, while AgentCore Gateway exposes that runtime through a single secure endpoint that MCP Apps-compatible hosts can call.

AWS demonstrates the pattern with a reference application called Unicorn Rentals. Users can ask an AI host in natural language to "show all unicorns," and the host renders interactive cards with images, names, hourly rates, and availability rather than a text list. Booking a unicorn returns a confirmation card with booking ID, date, and rate. Checking active bookings returns plain text instead of a widget—AWS notes that not every response requires a rich interface. Returning a unicorn calculates and reports total cost in the conversation. The same MCP server, according to AWS, delivers an identical experience whether accessed through ChatGPT, Claude, or any other host supporting the MCP Apps extension.

Architecturally, the MCP server is a TypeScript application built on the official @modelcontextprotocol/sdk with the @modelcontextprotocol/ext-apps extension, running as an Express.js HTTP server managed by AgentCore runtime. Tools such as list_unicorns, book_unicorn, view_bookings, and return_unicorn are registered via registerAppTool, each paired with a handler containing the business logic. A dedicated AWS Lambda function executes that logic against Amazon DynamoDB for persistence. Widget HTML is served as MCP resources through registerAppResource, and images load from Amazon CloudFront backed by Amazon S3.

Request flow runs through AWS WAF for IP allowlisting and managed rule screening before reaching AgentCore Gateway, which invokes AgentCore runtime using an IAM execution role. When a tool call has an associated resource URI, the AI host separately requests the widget's HTML via an MCP resources/read call, then injects structured data from the tool response into the rendered iframe. Hosts may cache both tool/resource listings and widget HTML for performance.

AWS provides no new pricing details specific to this pattern beyond standard Bedrock AgentCore runtime and Gateway costs, which are billed separately based on usage.

What this means

This is a reference architecture, not a new model or product launch—AWS is packaging existing AgentCore runtime and Gateway capabilities to support the emerging MCP Apps extension, positioning Bedrock as infrastructure for host-agnostic AI-native applications. The practical implication for developers is a documented path to build one MCP server that serves rich UI across multiple AI hosts without duplicating integration work for each one. Its significance depends on how quickly hosts beyond ChatGPT and Claude adopt the MCP Apps extension; until adoption broadens, the host-agnostic promise remains partially theoretical.