BYOKit

Bring Your Own Key, beautifully.

2.0 — Provider Command Center

BYOKit 2.0 turns the provider list into a compact command center: the active route, secure-key posture, and add-provider action are visible before the first row. The underlying components remain independently reusable and the base package remains dependency-free.

A Swift Package that gives any iOS / iPadOS / macOS app a production-grade BYOK (Bring Your Own Key) LLM configuration experience — in one line.

🏠 Site  ·  📖 Full guide →

---

import BYOKit

BYOKSettingsView()
    .environmentObject(store)            // a ConfigurationStore
    .byokClient(DefaultLLMClient())    // or your own LLMClient

Why

The hard part of supporting multiple LLM providers isn't calling them — there are many good libraries for that (AnyLanguageModel, SwiftOpenAI, …). It's the configuration UX: per-provider onboarding with deep links, key validation, one-tap connection testing, custom base URLs, and multi-key management. Every app rebuilds it by hand. BYOKit ships it.

Features

• One-line settings center — list, add, edit, reorder, delete, choose the active provider.
• Per-provider onboarding — structured "get a key" guide with steps, notes, and deep links (console / sign-up / docs / pricing).
• Secure by default — keys go to the Keychain; the UI masks them; "Test Connection" uses a minimal request.
• Data-driven catalog — add a provider by editing providers.json; no UI code. Ships built-in, optionally overridden by a remote OTA JSON.
• Swappable engine — DefaultLLMClient (URLSession, zero deps) speaks OpenAI, Anthropic, Gemini, and Ollama, with streaming built in. Conform to LLMClient to use anything else.
• Adaptive & themeable — looks right on iPhone, iPad, and Mac; matches your app's style.
• Localized — UI, validation hints, and error messages in English + Simplified Chinese, following the host app's language.

Built-in providers

OpenAI · Anthropic · Google Gemini · DeepSeek · OpenRouter · Groq · Mistral · xAI (Grok) · Ollama (local) · Custom (any OpenAI-compatible endpoint).

Each ships with brand styling, key validation, model presets, and a full onboarding guide.

Install

Xcode → File ▸ Add Package Dependencies… → paste https://github.com/everettjf/BYOKit → choose Up to Next Major Version (Xcode fills in the latest release for you).

Package.swift

.package(url: "https://github.com/everettjf/BYOKit", from: "2.0.0")

from: "2.0.0" is a floor, not a pin — SwiftPM always resolves the newest compatible release, so this line never needs editing when a new version ships.

Add the BYOKit product (umbrella). For just the data layer, depend on BYOKitCore; for just the UI, BYOKitUI.

Requirements: iOS 17 / iPadOS 17 / macOS 14, Swift 6.1 toolchain.

Usage

@main
struct MyApp: App {
    @StateObject private var store = ConfigurationStore()   // Keychain + UserDefaults

    var body: some Scene {
        WindowGroup {
            BYOKSettingsView()
                .environmentObject(store)
                .byokClient(DefaultLLMClient())
        }
    }
}

Send a request with the active configuration:

if let config = store.activeConfiguration,
   let provider = await ProviderRegistry.shared.provider(config.providerID) {
    let resolved = ResolvedConfiguration(
        provider: provider,
        configuration: config,
        apiKey: store.apiKey(for: config.id)
    )
    let response = try await DefaultLLMClient()
        .complete(.text("Hello"), with: resolved)
    print(response.text)
}

Streaming

Every LLMClient can stream incremental text. DefaultLLMClient parses the provider's wire stream (OpenAI/Anthropic/Gemini SSE, Ollama NDJSON); other clients fall back to a single complete call automatically.

for try await delta in client.streamComplete(.text("Write a haiku"), with: resolved) {
    print(delta, terminator: "")   // each chunk is a delta — concatenate for the full text
}

Localization

UI chrome, key-validation hints, and connection/error messages ship in English and Simplified Chinese (zh-Hans), resolved from the package's own bundle. The host app's language selection drives which is shown — no setup required.

Customization

BYOKSettingsView()
    .byokProviders(.only(.openAI, .anthropic, .ollama))   // limit providers
    .byokTheme(BYOKTheme(accent: .pink, cornerRadius: 16))
    .byokShowsOnboarding(true)
    .byokClient(myCustomClient)

Reusable pieces

ProviderPickerView, ProviderConfigForm, OnboardingGuideView, ConnectionTestButton, KeyField, and ProviderBadge are all public and usable on their own.

On-device models (Apple Intelligence)

The optional BYOKitClientAnyLanguageModel product adds on-device Apple Foundation Models via AnyLanguageModel, delegating every cloud/HTTP provider to a wrapped fallback (so custom BYOK endpoints keep working through DefaultLLMClient):

import BYOKitClientAnyLanguageModel

let builtins = await ProviderRegistry.shared.all()

BYOKSettingsView()
    .byokClient(AnyLanguageModelClient())                  // adds on-device support
    .byokProviders(builtins + [.appleFoundationModels])    // expose the Apple provider

Only depend on this product if you want local models — the base BYOKit library stays dependency-free.

MLX & llama.cpp (opt-in traits)

The adapter also services on-device MLX (Apple Silicon) and llama.cpp / GGUF models — gated behind BYOKit traits so the heavy backends are only pulled in when you ask for them:

.package(url: "https://github.com/everettjf/BYOKit", from: "1.0.0",
         traits: ["MLX", "Llama"])     // enable the backends you want

BYOKSettingsView()
    .byokClient(AnyLanguageModelClient())
    .byokProviders(builtins + [.appleFoundationModels, .mlx, .llama])

• .mlx — the selected model id is a Hugging Face MLX repo (e.g. mlx-community/Qwen3-0.6B-4bit), downloaded on first use.
• .llama — point the modelPath field at a local .gguf file.

With neither trait enabled (the default), these providers report that the backend isn't compiled in, and nothing heavy is added to your graph.

Architecture

Target	Role	Dependencies
BYOKitCore	Models, provider registry, onboarding metadata, providers.json	none
BYOKitStore	Keychain credential store + configuration persistence	Core
BYOKitClient	LLMClient protocol + DefaultLLMClient (URLSession)	Core
BYOKitUI	SwiftUI components	Core, Store, Client
BYOKit	Umbrella re-export	all
BYOKitClientAnyLanguageModel	Optional — on-device Apple Foundation Models (+ MLX / llama.cpp via traits)	Core, Client, AnyLanguageModel

BYOKitCore has zero third-party dependencies, and the base BYOKit library pulls in nothing third-party. Only the optional adapter above adds a dependency.

Example app

Example/ contains a multi-platform SwiftUI demo.

cd Example
xcodegen generate           # needs `brew install xcodegen`
open BYOKitDemo.xcodeproj

Testing

swift test

The test suite covers the registry, key validation, Keychain round-trips, the store, all four API formats (via a stubbed URLProtocol), streaming, and the AnyLanguageModel adapter's delegation. Set ROCKY_OPENAI_APIKEY to also run two live OpenAI smoke tests (skipped otherwise). The package is build-verified on macOS and the iOS/iPadOS simulator SDKs.

Roadmap

• M1/M2 — Core models, Keychain store, URLSession client, full SwiftUI configuration UI.
• M3 — Optional AnyLanguageModel adapter (on-device Apple Foundation Models) behind LLMClient.
• MLX / llama.cpp local models via opt-in BYOKit traits.
• M4 — Streaming completions + English / Simplified Chinese localization.
• Usage / quota hints.
• DocC API reference on the docs site.

Contributing

Issues and PRs welcome. Adding a provider is usually just an entry in Sources/BYOKitCore/Resources/providers.json — no code required.

License

MIT © Everett. See LICENSE.