Build interactive MCP Apps using Amazon Bedrock AgentCore
Amazon Web Services Dantis Stephen
AWS shows how to put interactive HTML widgets inside ChatGPT and Claude. The trick is one MCP server on Bedrock AgentCore, not host-specific code.
Based on reporting by Amazon Web Services, Dantis Stephen — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a cleaner way to make AI-hosted apps feel less like a wall of text and more like software. The company’s new example, Unicorn Rentals, runs inside ChatGPT, Claude, and other hosts that support the MCP Apps extension, and it keeps the experience the same across them.
The pitch is simple enough: MCP Apps adds interactive HTML widgets to Model Context Protocol, while Amazon Bedrock AgentCore supplies the runtime and gateway that let one server serve many hosts. In the demo, a user can ask for all unicorns, book Stardust, check active rentals, and return the animal without the app being rebuilt for each assistant.
Under the hood, the setup is split into a few parts. An AWS Lambda function handles the rental logic and stores data in Amazon DynamoDB. The MCP App itself is a TypeScript service built with the official SDK and the ext-apps extension. It registers tools like list_unicorns, book_unicorn, view_bookings, and return_unicorn, plus widget resources such as unicorn-list and booking-confirmation.
AgentCore runtime hosts the MCP server as a secure, serverless, session-isolated service, while AgentCore Gateway gives it a single endpoint that AI hosts can reach. AWS WAF sits in front with IP allowlisting and managed rules. When a host calls a tool, the request flows through the gateway to the runtime, then to Lambda and DynamoDB; if there’s a matching widget resource, the host reads the HTML and renders it in a sandboxed iframe.
AWS also sketches the deployment path. The code is packaged into a zip file, uploaded to Amazon S3, and wired into an AgentCore runtime configured for Node 22 and MCP protocol mode. The repo includes setup steps for ChatGPT and Claude, both using the GatewayResourceUrl output from the CDK deployment. For production, AWS points to CloudWatch, cost tuning, strict input validation, and Bedrock Guardrails for unstructured text.
My take — AI-written commentary, not fact-checked reporting
This is the right abstraction: keep the business logic where it belongs and let the MCP layer stay thin. The open-standard angle matters more than the unicorns do, because nobody needs a fresh hostage situation every time a new AI host shows up. The real test is whether teams ship widgets without turning them into a maintenance hobby.
Read more about this at: Amazon Web Services