| name | swift-ai-sdk |
| description | Build AI features in Swift for iOS and macOS with swift-ai-sdk — streaming text, structured output, tools, agents, embeddings, speech, transcription, images, video, realtime voice, computer use, and Apple's on-device models, across every provider (OpenAI, Anthropic, Google, xAI, Bedrock, Groq, Mistral, DeepSeek, Perplexity, Cohere, Voyage, Moonshot, Alibaba, HuggingFace, Cartesia, Sarvam, Ollama, and more). Use when writing Swift that calls generateText / streamText / generateObject, builds an Agent or tool loop, wires a ChatSession into SwiftUI, or adds realtime voice. Activate for swift-ai-sdk work even when it isn't named.
|
swift-ai-sdk
AI features for iOS and macOS apps, in plain Swift. One API across every provider; swap the
model and nothing else in your code moves.
Install
.package(url: "https://github.com/zaidmukaddam/swift-ai-sdk.git", from: "0.2.0")
Add "AI" to your target's dependencies (and "AITesting" to test targets). Requires
Swift 6 / Xcode 16+. import AI where you use it. The Foundation Models (on-device)
provider activates automatically when built against the iOS 26 / macOS 26 SDK.
The one thing to know
Every provider implements LanguageModel, and the top-level functions take any
LanguageModel — so switching providers is a one-line change:
let result = try await generateText(
model: AnthropicModel("claude-sonnet-5"),
prompt: "Why is the sky blue?"
)
print(result.text)
Keys come from the provider's conventional environment variable (ANTHROPIC_API_KEY,
OPENAI_API_KEY, XAI_API_KEY, …) or an explicit apiKey: argument.
Entry points
| You want | Call |
|---|
| A whole answer | generateText |
| Token streaming | streamText().textStream |
| Typed JSON / enum | generateObject / streamObject / generateEnum |
| A reusable agent | Agent |
| Embeddings / rerank | embed, embedMany / rerank |
| Speech, transcription, images, video | generateSpeech, transcribe, generateImage, generateVideo |
| A SwiftUI chat | ChatSession |
| Realtime voice | RealtimeSession |
References
Open only what the task needs. Start at references/index.md — a router table mapping
topic → file with a one-line "use for" per file. Then read only the references/<name>.md
files you need; don't read them all.
| Topic | File |
|---|
| Text generation, settings, callbacks | references/text-generation.md |
| Structured output + the Schema DSL | references/structured-output.md |
| Tools: function, typed, approvals, provider-executed | references/tools.md |
| Agents, loop control, prepareCall/prepareStep/toolOrder | references/agents.md |
| Providers: matrix, construction, provider tools, Sarvam | references/providers.md |
| Reasoning effort per provider | references/reasoning.md |
| Middleware: cache, extractReasoning, defaultSettings | references/middleware.md |
| Chat UI: ChatSession, transport, stream protocol | references/chat-ui.md |
| Realtime voice | references/realtime.md |
| Media: image, speech, transcription, video | references/media.md |
| On-device (Apple Intelligence) | references/on-device.md |
| Embeddings & reranking | references/embeddings.md |
| Testing & mocks | references/testing.md |
| Errors | references/errors.md |
Conventions
- Async throughout:
try await for one-shots, for try await part in stream for streams.
- Message roles:
.system, .user, .assistant, .tool. Build a prompt with system: +
prompt:, or pass messages: [Message] directly.
- The full docs mirror this skill at https://swift-ai-sdk.dev/docs — every page is also raw
markdown (append
.mdx to any URL, or start at /llms.txt).
