Flutter 3.50 is live! Read the blog post

Announcing A2UI 1.0 and a2ui_flutter

A2UI 1.0 renames genui to a2ui_flutter and adds catalog guardrails, multi-catalog surfaces, and a compact format for AI-generated Flutter UI.

Bilal Fali

6 min read
Share this
Announcing A2UI 1.0 and a2ui_flutter

If you've built a chat feature backed by an LLM, you know the output is almost always a wall of text. That's fine for a conversation, but it's a bad fit the moment the agent needs to show a form, a chart, a settings panel, or anything the user is supposed to act on rather than just read. The Flutter team calls this gap "Generative UI" (GenUI): instead of the agent replying with a text description of a UI, it replies with a real, interactive Flutter widget tree.

A2UI is the protocol behind that, and on November 17, 2026, the Flutter blog announced version 1.0 — along with a renamed Flutter SDK, a2ui_flutter, replacing the earlier genui package.

If you're evaluating this for the first time, or you already have genui in a project and need to know what breaks, here's what actually changed.

What A2UI and a2ui_flutter actually are

A2UI is the protocol — the message format an agent and a client use to describe and update a UI. The docs for the protocol itself live at a2ui.org, not in the Flutter package.

a2ui_flutter is the client-side SDK: a Flutter package on pub.dev that receives A2UI messages and renders them as actual widgets, on mobile, web, and desktop. It doesn't trust the agent blindly — it validates every incoming message against a schema, and if the agent asks for an invalid widget hierarchy, the client rejects it with an error and reports that error back to the agent instead of crashing the app.

That validation is the part worth paying attention to. An LLM is generating UI structure, and LLMs make structural mistakes. The protocol's answer is a catalog: a JSON schema that declares which components exist and which nesting is legal. The model sees that schema in its prompt and is expected to follow it; the client enforces it regardless of whether the model does.

If you're coming from genui: what actually broke

The package rename is the headline, but it's not the only breaking change:

  • The transport abstraction is gone. Earlier genui shipped its own transport layer. 1.0 removes it, so you plug in whatever streaming or HTTP setup you're already using instead of working around a built-in one.
  • MessageProcessor replaces SurfaceController + Conversation. One object now handles catalog registration, validates incoming A2UI messages, and manages surface state — where that used to be split across two.
  • A surface can now use multiple catalogs. In 0.9, one Surface was locked to a single Catalog. In 1.0, a surface can mix components from several registered catalogs.
final processor = MessageProcessor(
  catalogs: [
    CommonCatalog(),
    HealthWidgetsCatalog(),
    CustomChartCatalog(),
  ],
  // ...
);

a2ui_flutter resolves component names across every registered catalog automatically, so the model works against one unified schema instead of you manually routing by catalog.

The announcement also says six separate concepts from earlier genui releases were consolidated into the refined Catalog / Surface / Data Model set — it doesn't list which six, so treat that as a signal to read the full migration notes before upgrading a non-trivial integration, not as a complete list here.

What's new in 1.0, if you're starting fresh

Dual-layer guardrails. Catalogs can declare allowedParents and allowedChildren per component:

{
  "MyButton": {
    "type": "object",
    "allowedParents": ["Row", "Column"],
    "allowedChildren": ["Text", "Icon"],
    "properties": {
      "component": { "const": "MyButton" },
      "child": {
        "$ref": "common_types.json#/$defs/Child",
        "description": "The label or icon displayed inside the button."
      }
    }
  }
}

The model sees these constraints in its prompt, and the client re-checks them before building any widget. An invalid nesting comes back as UNALLOWED_PARENT or UNALLOWED_CHILD instead of a crash — the two-layer part matters because you can't fully trust the model to honor the schema just because it saw it.

A compact "Express" format. Alongside full JSON, the model can emit a function-call-style format for the same component tree. The Flutter team's own example describes five components in 46% fewer characters than minified JSON. Solid Software, building a healthcare prototype called GenUI Clinic, reported cutting full-screen dashboard generation from 10 seconds to 5 by switching to it — one data point, not a general benchmark, but a meaningful one if generation latency is your bottleneck.

callAgentFunction. A rendered widget can call a deterministic backend function directly, without round-tripping through the LLM for another turn. Somnio Software's example: updating order weights, fees, and remaining cash in under a second — the kind of interaction that would otherwise cost you a full model turn just to recompute a number.

Extensibility metadata. Surfaces, components, catalogs, and actions can now carry arbitrary metadata for things outside the protocol's core concerns — the example given is a logging token attached to a surface at creation time.

Where this fits, and where it doesn't

The announcement is explicit that GenUI "isn't meant to replace your app's deterministic screens" — it recommends sticking with static forms where a static form is actually the right tool. That's worth taking seriously: a catalog is a curated, closed set of components you have to build and maintain, and the model needs that schema in its context on every relevant turn. For a fixed settings screen, that's overhead you don't need. Where it earns its complexity is UI that's genuinely conditional on what the agent knows — the announcement's catering-app example only hides shellfish recipes when the agent has already learned a guest is allergic, which means there's no settings checkbox or filter UI to build and maintain for that case at all.

The broader ecosystem is still assembling: agent-side SDKs exist for Dart (a2ui_agent), Python (a2ui-agent-sdk, plus a2ui-adk for Google's Agent Development Kit and a2ui-a2a for the Agent2Agent protocol), and TypeScript (@a2ui/agent for Node.js). If you're already writing agents in Dart, Genkit Dart gets a genkit_a2ui middleware package — the announcement itself flags this as "currently in preview," so don't treat it as production-ready yet.

For a working reference beyond the docs, Very Good Ventures open-sourced a smart home panel built on this stack (Gemini Live for speech-to-text, Dart Frog on the backend) with both the source on GitHub and a live demo. It's the most concrete way to see a full catalog-and-surface setup rather than isolated snippets.

Should you use it

If you're already building agent features in Flutter and your agent's responses would genuinely benefit from being interactive widgets instead of text — a dashboard that reshapes itself per user, a form the agent partially pre-fills from context — A2UI 1.0 is worth the integration cost now that the catalog model, guardrails, and multi-catalog support are in place. If your UI is mostly fixed screens with an AI feature bolted on, you don't need this yet; a text response or a small set of hand-built components the agent chooses between will get you there with far less to maintain. And if you're on genui today, budget time for a real migration — the transport removal and the MessageProcessor consolidation aren't drop-in renames.

If you're setting up agent tooling in Flutter for the first time, our guide to AI coding agents for Flutter and the MCP server explainer cover the adjacent tooling this fits into.

Sources: flutter.dev/blog/a2ui-launch, pub.dev/packages/a2ui_flutter, a2ui.org

Bilal Fali

Bilal Fali

Flutter developer. I build and ship Flutter apps and write down what I learn along the way.

// related tool

Dart Data Class Generator

Paste JSON and generate an immutable Dart data class with copyWith, toString, and equality — no JSON (de)serialization boilerplate.

Open tool

Skip the boilerplate

Production-ready Flutter starter kit with Firebase Auth, Firestore, Cloud Functions, push notifications, and Clean Architecture — ship your app in days, not months.

Get Flutter Firebase Kit — $5

Did this article save you time?

I write these for free. If it helped, a coffee keeps me going — and more articles coming.

Buy me a coffee

Comments

Comments

Leave a comment

0/2000

Comments appear after review.