English | 日本語

swift-a2ui

Lets an LLM agent answer with a real SwiftUI screen — cards, forms, buttons the user can tap — instead of a wall of text, and keeps the model to the components your app can actually draw.

Unofficial. Not affiliated with or endorsed by the authors of the A2UI protocol. Conforming to the specification is not a goal of this project.

Overview

Under A2UI, the agent sends JSON messages that create, update, and delete UI surfaces, and the client renders them and sends user actions back. This package is both halves in Swift.

• The palette is Swift types. The JSON Schema the model is prompted with is generated from them, so what the model is told it may draw cannot drift from what the renderer knows how to draw.
• Prompting and validation share one allowlist. Offer six components and the model gets a schema for six; anything else is rejected by the tool with an error it corrects from in the same turn, so the user never sees an unsupported-component placeholder.
• No AnyView, no string matching. A2UISurfaceView<Catalog> is generic over the catalog, so adding your own components is a type-level extension the compiler checks for exhaustiveness.
• Model output is messy; this expects that. Streaming text is parsed incrementally as it arrives, and the common malformations LLMs produce are repaired rather than dropped.
• More than one agent can share a screen. An ownership ledger routes each user action back to the agent that drew the surface, and keeps one agent's data model out of another's messages.

Usage

Render whatever the agent sends. TypedMessageProcessor holds the surfaces and A2UISurfaceView draws them:

import SwiftUI
import A2UIParser
import A2UITyped
import A2UITypedRenderer

struct AgentReply: View {
    @State private var processor = TypedMessageProcessor<BasicCatalog>()
    private let parser = A2UIStreamingParser()

    var body: some View {
        ForEach(processor.ordered) { surface in
            A2UISurfaceView(surface)
        }
    }

    func receive(_ chunk: String) {
        for part in parser.feed(chunk) {
            if let messages = part.messages {
                processor.process(messages)
            }
        }
    }
}

On the agent side, A2UIPresenterAgent carries its own prompt, tool, and agent-card extension, so the host picks only the language and the palette:

import A2UIAgent

let prompt = A2UIPresenterAgent.systemPrompt(language: "English")
let tools = A2UIPresenterAgent.tools(
    components: ["Column", "Row", "Text", "Image", "Card", "Button"]
)

Documentation

API reference for every module, including how to compose a catalog of your own components.

Installation

Add the package to your Package.swift:

.package(url: "https://github.com/no-problem-dev/swift-a2ui.git", .upToNextMinor(from: "0.28.0")),

Each module ships as its own library, so depend only on what you use — A2UICore for the message types, A2UITypedRenderer to draw them, A2UIAgent for the ready-made presenter agent:

.target(name: "MyApp", dependencies: [
    .product(name: "A2UICore", package: "swift-a2ui"),
    .product(name: "A2UITypedRenderer", package: "swift-a2ui"),
    .product(name: "A2UIAgent", package: "swift-a2ui"),
])

Requirements

Swift	6.2
Platforms	iOS 17 · macOS 14

Contributing

Bug reports and pull requests are welcome — see CONTRIBUTING.md.

License

MIT. See LICENSE.