swift-llm

Native Swift Package for Claude and OpenAI — async/await, streaming, multi-turn, zero dependencies.

---

Install

Add the package in Xcode via File → Add Package Dependencies, or directly in Package.swift:

dependencies: [
    .package(url: "https://github.com/pintaste/swift-llm", from: "1.0.0")
],
targets: [
    .target(name: "YourTarget", dependencies: ["SwiftLLM"])
]

---

Usage

Single message

import SwiftLLM

let claude = ClaudeClient(apiKey: "sk-ant-...")
let reply = try await claude.message("What is the capital of France?")
print(reply) // "The capital of France is Paris."

Streaming

import SwiftLLM

let openai = OpenAIClient(apiKey: "sk-...")
for try await chunk in openai.stream("Tell me a joke") {
    print(chunk, terminator: "")
}

Multi-turn chat

import SwiftLLM

let chat = ClaudeClient(apiKey: "sk-ant-...").chat()
let first  = try await chat.send("My name is Alice.")
let second = try await chat.send("What is my name?")
// second → "Your name is Alice."

---

API Reference

Method	ClaudeClient	OpenAIClient
Single message	message(_ text:, model:, maxTokens:) async throws -> String	Same signature
Streaming	stream(_ text:, model:, maxTokens:) -> AsyncThrowingStream<String, Error>	Same signature
Multi-turn chat	chat() -> ClaudeChat	chat() -> OpenAIChat
Default model	claude-opus-4-7	gpt-4o

Both ClaudeChat and OpenAIChat expose:

Method	Description
send(_ text:, model:, maxTokens:) async throws -> String	Send a turn; history is maintained automatically

---

Error Handling

All methods throw LLMError:

public enum LLMError: Error {
    case invalidAPIKey       // HTTP 401
    case rateLimitExceeded   // HTTP 429
    case invalidResponse     // Other non-2xx
    case networkError(Error) // URLSession failure
    case decodingError(Error)// JSON decode failure
}

---

Why swift-llm?

• Zero dependencies — uses only Foundation and URLSession
• Native async/await — no callbacks, no Combine wrappers
• Streaming via AsyncSequence — for try await chunk in client.stream(...) just works
• Multi-turn history — ClaudeChat / OpenAIChat track the full conversation automatically
• Broad platform support — macOS 12+ and iOS 15+
• Tiny surface area — three types, one import, done