Policy.canvas(anthropic:)Wire any frontier model in one line
Anthropic, OpenAI, on-device — same API. The policy owns the round-trip; you own the surface.
Give your AI
assistant
a body.
The native iOS canvas where AI assistants come alive — drop into your chat, no server, no iframe.
Get started
$ swift package add ainativeui9:41
Maya
I'm bored.
Sure — want a crossword, music, or a quick game?
Quick crossword
1
2
3
Across
1Common house pet (3)
3Buzzing insect (3)
Down
1First letter of "car" (1)
2First letter of "apple" (1)
A canvas, a policy, a session.
AICanvasView(session:)Drop into any SwiftUI view
Renders alongside whatever chat surface you already ship. The canvas is a view, not a window.
suggestion_cluster recipeLet the AI propose, the user picks
Manifest a cluster of choices; the user's first tap is local — no model round-trip, no spinner.
Where it fits, where it doesn't.
| Vs. | Native iOS | Same chat | No backend | Frontier model | Whole-app build |
|---|---|---|---|---|---|
| AINativeUI | |||||
| OpenAI Apps SDK | |||||
| Anthropic Artifacts | |||||
| MCP Apps (SEP-1865) | |||||
| CopilotKit / AG-UI | |||||
| Apple Intelligence |
We're not a web framework and not a tool for building a whole app from a prompt. The columns where we say no are honest about the shape of the wedge.
Three frontier models, all near-ceiling.
100.0%
100.0%
97.9%
Three price points, three vendors, one substrate. The wedge is real.
One policy, one view, one input.
import AINativeUIAgent
import AINativeUIAnthropic
import AINativeUIWidgets
import AINativeUIRecipes
@State private var session = Session(
policy: .canvas(anthropic: ProcessInfo.processInfo.environment["ANTHROPIC_API_KEY"]!)
)
var body: some View {
VStack {
AICanvasView(session: session)
.aiBuiltinWidgets()
.aiBuiltinRecipes()
ChatInputView(onSend: { prompt in Task { try? await session.send(prompt) } })
}
}The canvas is a SwiftUI view, the policy owns the model round-trip, and your existing input field stays exactly where it was.
Read the full guide