Verdict: MCP Apps represent a fundamental shift in how we interact with AI agents. By moving beyond text-only JSON responses to rich, sandboxed UI components (iframes), developers can now deliver interactive dashboards, real-time visualizations, and complex forms directly inside the chat interface of clients like Claude and ChatGPT.
What are MCP Apps?
MCP Apps are the first official extension to the Model Context Protocol (MCP), released in early 2026. While standard MCP servers return text or JSON, an MCP App returns an interactive HTML resource that the host renders in a sandboxed iframe.
Instead of an agent saying "Your sales are $1.2M," it can now show a live, interactive chart where you can drill down into specific regions or export reports—all without leaving the conversation. As Y Combinator's Paul Graham noted in July 2026, "AI apps are the new browsers," and MCP Apps are the "websites" powering this new internet.
The Interactive Primitives of 2026
Building a modern MCP App requires mastering three core primitives that allow the UI to communicate with the AI model:
- Bidirectional Communication: The UI isn't static. It can send messages back to the model using
set-stateor trigger follow-up prompts. For example, clicking a "Learn More" button in a product widget can automatically send a message to the agent to provide more details. - Streaming UI: One of the most impressive features of the 2026 spec is the ability to update the UI incrementally. As the model streams input tokens into tool arguments, the UI can react in real-time. This is already being used by tools like Excalidraw to render diagrams as the AI "thinks" them into existence.
- Nested Tool Calls: A widget can trigger other tools from within the iframe. This allows for complex, multi-step workflows like a checkout form that calls a "Payment" tool only after the user has confirmed their details in the UI.
Privacy and Data Redaction
A major advantage of the MCP App architecture is the ability to separate what the user sees from what the model sees. Developers can show full private data (like personal IDs or sensitive financials) in the sandboxed UI while sending only redacted or summarized versions to the LLM provider. This is already a key AI agent safety feature for enterprise AI deployments.
Discovery: The One-Click Install Era
The distribution of AI tools has moved from "copy-pasting JSON config files" to "one-click install stores."
- ChatGPT Apps: A dedicated store where users can browse and install vetted MCP servers and apps.
- Claude Connectors: Anthropic's registry that supports dynamic discovery—if you ask Claude a question and it doesn't have a tool, it can search the registry and suggest the right connector automatically. Learn more in our Claude Sonnet 5 guide.
- Cursor & VS Code: Integrated directories that allow developers to pull in coding-specific MCP apps directly into their IDE.
How to Build and Ship
If you are building your first MCP App in 2026, the ecosystem has matured with professional-grade tooling:
- MCP-use SDK: Pedro Zullo's popular abstraction over the official SDK that makes registering UI resources as simple as adding a React component to a folder.
- Manifold Cloud: The primary platform for hosting, testing, and publishing your MCP servers to the major stores with built-in compliance checks.
What this means for you: If you build software, the chat interface is your new front-end. Stop building standalone dashboards and start building MCP Apps that meet your users where they already are: in their AI assistants. This is a core part of the 5-layer agentic stack.
FAQ
Q: Do I need a specific framework to build MCP Apps?
A: No. While many use React with the MCP-use SDK, the protocol is framework-agnostic. You can serve plain HTML/JS resources via the ui:// scheme.
Q: Which clients support MCP Apps? A: As of mid-2026, major supporters include Claude (Desktop/Work), ChatGPT, Cursor, VS Code (via Copilot), and Goose.
Q: Are MCP Apps secure? A: Yes. They run in sandboxed iframes with a strictly defined JSON-RPC communication channel back to the host. They cannot access your local files or browser data unless explicitly granted permission via an MCP tool.
Q: Can a standard MCP server be an MCP App? A: Yes. An MCP App is just an MCP server that utilizes the UI extension. You can add UI capabilities to any existing server.
Q: How does this relate to AI agent reliability? A: Interactive UIs provide a human-in-the-loop layer that ensures verifiable AI improvement.
Discussion
0 comments