The Missing Piece of the AI Puzzle: Affordances

Last summer, I wrote about a question that's been on my mind since my days studying Human Computer Interaction at Georgia Tech: what happens to user interfaces when the interface can think?

I explored the tension between conversational AI and traditional GUIs. Conversational interfaces offer infinite possibility but zero discoverability. GUIs constrain what's possible but make options explicit. At the time, I called it “the tension between capability and clarity,” and I didn’t have a great answer for how to resolve it.

Now, I think we’re looking at the first real answer: MCP Apps.

Quick Context: What’s MCP?

If you’re not familiar, the Model Context Protocol is an open standard for connecting AI models to external tools and data sources. Think of it as a universal adapter that lets your AI assistant talk to databases, APIs, file systems — whatever it needs. It’s been adopted by Claude, ChatGPT, VS Code, and a growing ecosystem of clients and servers.

MCP solved the capability problem. But it left a gap in the experience.

The Context Gap

Say you ask your AI assistant to pull sales data from your database. The tool runs, returns hundreds of rows, and the model does its best to summarize. But what if you want to sort by revenue? Filter to last quarter? Drill into a specific account?

Every interaction becomes another prompt.
“Show me just the ones from last week.”
“Sort by the third column.”
“What’s the detail on row 47?”

It works. But it’s clunky. You’re using natural language to do what a simple click-and-sort table does instantly. That disconnect between what the tools can do and what users can see and manipulate has been a significant point of friction since the dawn of AI assistants.

Enter MCP Apps

MCP Apps — the first official MCP extension — close this gap.

Tools can now return rich, interactive UI components that render directly in the conversation. Dashboards. Forms. Visualizations. Multi-step configuration wizards. Document viewers with inline annotations. Live monitoring displays that update in real time.

The user experience implications are significant. Remember the cognitive shift I wrote about last summer — from recall (command lines) to recognition (GUIs) to articulation (conversational AI)? MCP Apps adds a fourth mode: interaction.

You articulate your intent in natural language, and the assistant responds with a purpose-built interface you can manipulate directly. That’s great for users, but it also happens to be great for the assistant.

When a user clicks a checkbox, selects an option, or filters a dataset through an MCP App, that interaction flows back to the model via a standardized API. The app can call updateModelContext() to tell the model exactly what the user did:

“User selected Option B,”
“User filtered to Q4 2025,”
“User flagged clause 3.2 for review.”

The model doesn’t have to guess what you meant from an ambiguous follow-up prompt — it knows what you did, because the UI told it.

This is a massive improvement in the quality of context available to the model. Better context means better responses. Better responses mean more trust. More trust means users are willing to hand more complex workflows to their assistants. It’s a virtuous cycle.

Why This Matters Beyond UX

What excites me most is that MCP Apps is a cross-client standard. Claude, ChatGPT, VS Code, and Goose all support it today, with more coming. A tool developer can build an interactive experience once and have it work everywhere.

We haven’t had that before. Every assistant vendor was building their own bespoke UI layer, and tool developers had to choose a platform or build multiple integrations.

This is a sign of genuine maturation in the GenAI ecosystem. We’re past the “wow, it can write poetry” phase. We’re into the hard, important work of building standards, interoperability, and design patterns that make AI assistants genuinely useful for real workflows.

If you’re building tools or thinking about how AI fits into your products, MCP Apps is worth your attention. The
quickstart guide is solid, and the architecture — sandboxed iframes communicating via JSON-RPC over postMessage — is familiar territory for any web developer.

The conversational interface isn’t going away. But it’s growing up. And it’s learning that sometimes, the best response to a question isn’t more text — it’s a button.