LLMkit

A lightweight Swift package that provides stateless API clients for cloud transcription, LLM chat completion, and real-time streaming transcription services. Built for VoiceInk, but usable in any Swift project.

What It Does

LLMkit is a pure networking layer — no UI, no persistence, no app state. It handles HTTP requests, multipart uploads, WebSocket connections, error mapping, and automatic retries with exponential backoff on transient failures (429, 5xx).

Clients

Transcription (audio file → text):

• DeepgramClient, ElevenLabsClient, SonioxClient, MistralTranscriptionClient, GeminiTranscriptionClient, OpenAITranscriptionClient

LLM Chat Completion (messages → response):

• AnthropicLLMClient, OpenAILLMClient, OllamaClient, OpenRouterClient

Streaming Transcription (live audio → real-time text via WebSocket):

• ElevenLabsStreamingClient, DeepgramStreamingClient, MistralStreamingClient, SonioxStreamingClient

Usage

All batch clients use static methods — no initialization needed:

import LLMkit

let text = try await DeepgramClient.transcribe(
    apiKey: "...", audioData: data, model: "nova-3", language: "en"
)

let response = try await AnthropicLLMClient.chatCompletion(
    apiKey: "...", model: "claude-sonnet-4-5-20250929", messages: [.user("Hello")]
)

Streaming clients are class-based with AsyncStream event delivery:

let client = DeepgramStreamingClient()
try await client.connect(apiKey: "...", model: "nova-3", language: "en")
for await event in client.transcriptionEvents { /* .partial, .committed, .error */ }

Requirements

• Swift 6.2+, macOS 14+, iOS 17+
• Zero external dependencies — uses only Foundation and URLSession

Error Handling

All clients throw LLMKitError with cases for missing keys, HTTP errors, network failures, decoding issues, and timeouts.