Agent Ui Standards Multiply: Mcp Apps And Google’s A2ui

This week, Google launched A2UI, an unfastened root task to thief developers build “agentic personification interfaces.” It’s nan latest successful what’s becoming a regular merchandise rhythm of caller standards and protocols for building a user interface successful AI agents and chatbots.

First, location was MCP-UI, a Model Context Protocol (MCP) ecosystem task intimately aligned pinch Anthropic and used by Shopify (amongst others). Soon after, OpenAI launched Apps SDK, on pinch AgentKit and different UI tooling. Then, conscionable past month, MCP Apps was announced — a projected unfastened modular “for interactive personification interfaces successful nan Model Context Protocol,” which is being supported by some Anthropic and OpenAI.

What Is A2UI? A Cross-Platform, Native-First Approach

So what is Google bringing to nan array pinch A2UI that doesn’t already beryllium from nan flurry of different agentic UI projects? In nan motorboat blog post, Google states that “A2UI was designed to reside nan circumstantial challenges of interoperable, cross-platform, generative aliases template-based UI responses from agents.”

The word “cross-platform” is simply a large hint — this isn’t a web-centric approach, which MCP-UI and OpenAI person mostly taken truthful acold by relying connected sandboxed iframes. Instead, A2UI takes what Google’s Minko Gechev calls a “native-first” approach. Gechev described this arsenic “declarative, not executable,” adding that AI agents “send a explanation of UI components, not code.”

Currently, A2UI has customer libraries for Flutter, Web Components and Angular. But this will astir apt beryllium extended to different libraries complete time. The thought is for an supplier to state what it should look for illustration connected a receiving app (e.g., a chatbot), and autochthonal libraries will beryllium utilized to make nan personification interface. As nan A2UI motorboat station explains:

“A2UI separates nan UI building from nan UI implementation. The supplier sends a explanation of nan constituent character and its associated information model. Your customer exertion is responsible for mapping these absurd descriptions to its autochthonal widgets — beryllium it web components, Flutter widgets, React components, SwiftUI views aliases thing other entirely.”

Gechev besides explained that A2UI is built for streaming. “Using a JSONL-based format, A2UI enables progressive rendering truthful users spot results instantly arsenic nan supplier ‘thinks,’” he wrote.

Understanding OpenAI’s Web-Centric UI Strategy

There is immoderate huddled together for positioning going connected successful nan emerging section of “agent development.” OpenAI, successful particular, has its fingers successful a bunch of pies (cue a joke astir six-fingered AI-generated images).

Currently, OpenAI is putting astir of its effort into trying to make ChatGPT into an exertion platform, wherever apps will beryllium web widgets rendered successful sandboxed surfaces (typically iframes). Just this week, OpenAI announced that developers tin now submit apps to ChatGPT.

However, OpenAI has besides precocious entered nan web browser market, pinch Atlas. And it whitethorn good get into nan smartphone business — aliases immoderate different hardware instrumentality — astatine immoderate point, too. Earlier this month, I interviewed Adam Shea, head of engineering astatine TELUS Digital, who told me, “I person a hunch that nan measurement we build ChatGPT apps coming shares immoderate halfway ideas and building pinch nan imaginable smartphone level OpenAI has been rumored to beryllium launching successful nan future.”

While OpenAI seems to beryllium covering each bases, it is focused connected web technologies successful its Apps SDK and (obviously) successful Atlas. Which makes its attack different from Google, which is going for cross-platform scope pinch A2UI.

MCP Apps vs. A2UI: Key Differences Explained

It’s besides noteworthy that OpenAI is simply a contributor to nan MCP Apps project, which derives from some nan MCP-UI task (created by Ido Salomon and Liad Yosef, who some now activity for Monday.com) and OpenAI’s Apps SDK. The institution that birthed MCP, Anthropic, is besides heavy progressive successful this project.

MCP Apps Extension (SEP-1865), to springiness its afloat name, was launched past month connected nan charismatic MCP blog. It intends to “standardize support for interactive personification interfaces successful nan Model Context Protocol.”

Currently, MCP Apps takes an explicitly web-based approach, utilizing nan iframe sandboxing that some MCP-UI and Apps SDK support. As nan motorboat station put it, “all UI contented runs successful sandboxed iframes pinch restricted permissions.”

In its A2UI motorboat post, Google emphasized that its “native-first” attack differs from MCP Apps. “Instead of retrieving an opaque payload to show successful a sandbox, an A2UI supplier sends a blueprint of autochthonal components,” it explained. The cardinal constituent present is that nan “blueprint” sent via A2UI mightiness beryllium utilized to make web code, autochthonal mobile UI aliases desktop exertion components.

Emerging Agent Development Frameworks

There’s been a batch of different pieces released this twelvemonth to thief developers build agents aliases link their apps to agents. Google alluded to this successful its A2UI motorboat post, noting that successful summation to building a UI, “you whitethorn besides utilize a model (AG UI, Vercel AI SDK, GenUI SDK for Flutter, which already uses A2UI nether nan covers) to grip nan ‘pipes.’”

As good arsenic nan plethora of acronyms, it’s besides confusing sometimes arsenic to who makes what. AG UI (Agent-User Interaction), it turns out, is an relationship protocol from a Seattle-based institution called CopilotKit — which provides tooling that implements AG UI. It besides now supports A2UI.

And I haven’t moreover mentioned Google’s Agent2Agent Protocol (A2A) yet! A2A operates astatine nan agent-to-agent coordination layer, alternatively than nan UI layer.

The constituent is, location are a batch of different technologies developers person to wade done presently earlier they fig retired really to build and link an agent. Hopefully, successful 2026, each of this will go clearer.

But for UI astatine least, location are now 2 reasonably clear approaches: Build a mini web app (favored by OpenAI and Anthropic, and now supported by MCP Apps) aliases return a “native-first” attack pinch Google’s A2UI.

Group Created pinch Sketch.