swift-ai-sdk

Same code, on-device or cloud. swift-ai-sdk is a provider-agnostic AI toolkit for Swift, built on the Vercel AI SDK's design. You get generateText, streamText, and generateObject, tool calling with the agentic loop handled for you, a swappable transport layer that speaks Vercel's UI message stream protocol, and an @Observable chat session for SwiftUI. Swift 6 strict concurrency, Sendable throughout.

Why

On-device and cloud sit behind one API. FoundationModelsModel() is free, private, and works offline; AnthropicModel() and OpenAIModel() are a one-line swap away, and orFallback picks for you based on availability.

If you already run a Vercel AI SDK backend, your Swift app can use it as-is. HTTPChatTransport speaks the UI message stream protocol exactly (text-delta with a delta field, tool-input-delta with inputTextDelta, data: [DONE]), so the same /api/chat route that serves your React app serves your iOS app.

And it's Swift-6-native: everything is Sendable, streaming is AsyncThrowingStream, strict concurrency is on, and there's no @unchecked in the public surface.

The one abstraction

Every provider implements a single method:

protocol LanguageModel: Sendable {
    func stream(_ request: LanguageModelRequest) async throws -> AsyncThrowingStream
}

generateText, streamText, generateObject, and the chat transports are all just consumers of that stream. One code path, every backend.

Tour

Tools + the agentic loop

let weather = Tool(
    name: "weather",
    description: "Current weather for a city",
    parameters: ["type": "object", "properties": ["city": ["type": "string"]], "required": ["city"]]
) { args in
    ["tempC": 31, "city": args["city"] ?? .null]
}

let result = try await generateText(
    model: model,
    prompt: "Weather in Mumbai?",
    tools: [weather],
    stopWhen: [stepCountIs(4)],          // bound the loop (Vercel's stopWhen)
    prepareStep: { context in            // per-step overrides (model, messages, tools)
        context.stepNumber > 2 ? PrepareStepResult(model: cheaperModel) : nil
    }
)
result.text        // final answer, after the model called the tool
result.steps       // every round-trip: text, toolCalls, toolResults, usage

The loop streams too: streamText(...).fullStream yields step boundaries, tool calls, and to