MCP UI: Extending the frontier — Liad Yosef and Ido Salomon, MCP Apps

From text-only tools to UI inside chat

Ido Salomon opens with a very current joke — they built the talk “this morning and it might already be out of date” — then gets straight to the pain point: MCP tools used to dump text into chat, and that flattened every company into the same anonymous blob. His pitch is simple and sticky: instead of Shopify, Booking, or Hugging Face being reduced to a wall of text, let each service send its own UI chunk into the assistant.

How MCPUI became the official MCP apps extension

Salomon explains that MCPUI started in May last year as a practical attempt to pass UI over MCP without throwing out everything the industry already knows about UX, branding, and interaction design. A few months ago, that work was folded into the MCP standard with Anthropic and OpenAI as the first official extension, and he rattles off the validation: Claude, ChatGPT, VS Code, Cursor, Microsoft Copilot, Postman, Goose, and more now support it.

The core mechanic: return a resource, not just text

Liad Yosef walks through the technical leap: instead of a tool returning plain text, it can return a resource — actual HTML — that a host can render as an interactive app. The important nuance is that interactivity still stays in the assistant’s loop; if you favorite a Spotify song through the widget, the event goes back through the host, not directly to Spotify, so Claude can still remember what happened later.

The PostHog demo makes the case in one glance

The demo is intentionally mundane and effective: ask Claude to analyze a funnel, and the old mode gives you a long textual explanation you have to squint through. Then Salomon says, “show me,” and the same request turns into a PostHog-built visualization that preserves the company’s identity and makes the answer instantly legible; click a funnel step, and Claude can follow up interactively instead of restarting the whole exchange.

Their real thesis: this changes how the web is organized

From there, the talk gets more ambitious. Yosef argues this is “ushering a new web,” where personal assistants compose just the relevant UI atoms from Google Calendar, Amazon, and Booking instead of making users bounce across tabs and dashboards; the human angle is an anniversary-planning example where the assistant knows your preferences, while each company contributes the part it actually knows best.

A new control model for apps, hosts, and users

They emphasize that this also breaks the old ownership model: the app no longer owns the full journey once a user enters it. Interactions now sit on a spectrum — a notification gives the UI maximum control, a tool call asks the host to act, and a prompt hands over almost everything — and they frame that not as a small API tweak but as a new philosophy for software in agent environments.

What’s shipping next: reusable views and model-clickable apps

The spec, they say, is evolving quickly through a public workgroup that meets every three weeks. Two standout next steps are reusable views, so heavy apps like Autodesk don’t have to fully rerender every time, and a standardized way for the model itself to interact with the UI — clicking buttons, filling forms, closing the loop beyond today’s stopgaps like WebMCP.

The opportunity: build once, run across billion-user hosts

They close by reframing MCP apps as a distribution channel, not just a protocol. With ChatGPT alone having gone from Sam Altman’s cited 800 million weekly users to what they say is now 1 billion, plus Claude and VS Code on top, they call this a once-in-20-years platform shift: build an MCP app once with XApps or host it with the MCPUI SDK, and it can run across the emerging ecosystem of AI-native surfaces.