a2ui-starter-swiftui

Reference client-side A2UI starter app. SwiftUI, macOS/iOS, Claude-powered, no server.

Built on a2ui-swiftui (renderer) + a2ui-skills-swiftui (client-side skill runtime). One swift run away from a working multi-turn agent that streams Claude's output into progressively-rendering A2UI surfaces, with all three progressive-rendering RFC primitives implemented end-to-end.

What this demonstrates

• Client-side skills. Four skills (greeting, planner, decider, critic) load as inline SKILL.md strings. The client routes events to the right skill based on the user's tap.
• Four distinct UI shapes. One per skill — OptionsGrid (planner's numbered steps), two stacked Cards (decider), single RichMessageCard (critic), plus the static greeting.
• Skeleton-first streaming. Each skill's first-turn UI renders instantly from the bundled skeleton (no LLM delay). Claude fills only the text values.
• Progressive-rendering RFC primitives:
  • Proposal 1 — pending state: shimmer placeholders while path bindings are unresolved.
  • Proposal 2 — streaming: true/false flag: typewriter caret appears inline on bound text fields during streaming, disappears on finalization.
  • Proposal 3 — append patch op: each token is appended via append, not resent as full accumulated value (~50× wire-efficient).
• Direct Anthropic streaming: AnthropicLLMProvider calls the Messages API directly from the app, parsing SSE text deltas into per-field streams via FieldParser. No server required.

Run

git clone https://github.com/vpm238/a2ui-starter-swiftui
cd a2ui-starter-swiftui
export ANTHROPIC_API_KEY="sk-ant-..."
swift run A2UIStarter

A native macOS window opens with the greeting surface. Tap any option to see the skill-specific layout render instantly, then watch the LLM-written text stream in character-by-character.

Architecture

┌────────────────────────────────────────┐
│ A2UIStarter (SwiftUI app)             │
│                                        │
│  @main App → ChatView                 │
│                                        │
│  State: SkillRuntime                  │
│    skills: [greeting, planner,        │
│             decider, critic]           │
│                                        │
│  Render: A2UISurfaceView per turn     │
│  (from a2ui-swiftui)                    │
└──────────┬────────────────────────────┘
           │ LLM calls (client-side)
           ▼
   api.anthropic.com/v1/messages

Zero backend. Skills are bundled as Swift strings. Events route locally in the runtime. LLM calls go from the client to Anthropic directly over HTTPS/SSE.

For production, wrap the API key with a thin proxy (conform any type to LLMProvider and point it at your proxy endpoint) so the key isn't on the device. The web starter ships a Cloudflare Worker proxy you can adapt.

Project structure

a2ui-starter-swiftui/
├── Package.swift
├── skill.manifest.json         # experimental host manifest (see below)
├── Sources/A2UIStarter/
│   ├── StarterApp.swift        # @main + app delegate
│   ├── StarterChatView.swift   # transcript + composer + shim client
│   └── BundledSkills.swift     # SKILL.md as Swift strings
└── README.md

About skill.manifest.json

Experimental proposal — no host reads it yet. It's here as a concrete draft of how A2UI host apps might eventually declare their capabilities, bundled skills, and LLM backends in a discoverable way. Analogous to package.json for npm or Cargo.toml for Rust, but for A2UI.

See skill.manifest.json in this repo for the proposed shape.

Dependencies

• a2ui-swiftui — SwiftUI A2UI renderer
• a2ui-skills-swiftui — client-side skill runtime

Requires Swift 6.0, macOS 14+.

License

MIT. See LICENSE.

Related

• a2ui-swiftui — renderer library used here
• a2ui-skills-swiftui — skill runtime used here
• a2ui-progressive-rendering-rfc — RFC and live demo for the streaming primitives this starter implements