Building ChatGPT and Claude Apps

In this video, Brandon Mathis and Ben Lesh break down the emerging MCP Apps protocol and what it means for the future of AI-powered applications. Drawing from hands-on experimentation with tools like ChatGPT Apps, Claude, Codex, and MCP servers, they explore how developers can build interactive applications directly inside AI chat experiences using standardized protocols, iframe-based UI rendering, and tool integrations. The conversation walks through the technical architecture behind MCP Apps, including app tools, metadata configuration, input schemas, UI resource hosting, and how models decide when and how to invoke tools. Brandon and Ben also discuss the realities of building against rapidly evolving AI standards, covering everything from TypeScript and Zod validation to local development workflows with ngrok, caching issues, integration testing challenges, and content security policies.Along the way, they unpack larger themes shaping AI development right now: the growing importance of open standards, the tradeoffs of non-deterministic systems, security concerns around MCP tooling, and how AI interfaces may reshape the future of application development itself. The episode also explores the parallels between today’s AI tooling ecosystem and the early days of the web and mobile app platforms, including the risks, opportunities, and maintenance challenges developers should expect as these protocols mature. What You Will Learn: \- How the MCP Apps protocol allows developers to build interactive applications directly inside AI chat platforms like ChatGPT and Claude \- Why open standards are becoming important for creating AI tools that work across multiple ecosystems and models \- The practical realities of building MCP Apps, including tool registration, UI hosting, schemas, caching, and local development workflows \- The security and privacy risks involved with MCP tools and why developers need to carefully manage tool permissions and data exposureWhy testing AI-powered systems is more difficult than traditional software testing due to non-deterministic model behavior and evolving protocols Chapters 00:00 Introduction to MCP apps and AI chat integrations 04:00 How MCP apps work inside ChatGPT and Claude 06:40 Building MCP apps under the hood 13:20 Local development, ngrok, and testing workflows 16:40 Ideal use cases and limitations of MCP apps 20:20 MCP app marketplaces, approvals, and discovery 21:40 Security risks, trust, and data leakage concerns 24:40 Why MCP apps require ongoing maintenance 27:10 AI tooling and plugins for building MCP apps 28:45 Testing non-deterministic AI applications 32:20 The return of iframes in modern AI apps 35:00 Final thoughts on the future of MCP apps Brandon Mathis on Linkedin: https://www.linkedin.com/in/mathisbrandon/ Ben Lesh on Linkedin: https://www.linkedin.com/in/blesh/ This Dot Labs Twitter: https://x.com/ThisDotLabs This Dot Media Twitter: https://x.com/ThisDotMedia This Dot Labs Instagram: https://www.instagram.com/thisdotlabs/ This Dot Labs Facebook: https://www.facebook.com/thisdot/ Sponsored by This Dot: https://ai.thisdot.co/ AI Workshop Series from This Dot Labs: https://ai.thisdot.co/workshops Use Code THISDOTX at Checkout for $50 tickets!