Multi-platform Apple app for discovering, broadcasting, and understanding Bonjour (mDNS/DNS-SD) services. AI chat via on-device Apple Intelligence, Anthropic Claude, or Google Gemini. Siri / App Intents. SwiftUI, Swift 6, full accessibility, 8 languages incl. RTL. Structure and agent rules from AppBootstrapAI.
저장소 README에서 가져옴
README
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
A multi-platform Apple app for discovering, broadcasting, and understanding Bonjour (mDNS/DNS-SD) network services. Rich service inspection, filter categories, a 158-entry service-type library, and AI-powered explanations — on-device by default, with optional Anthropic Claude or Google Gemini integration — on iPhone, iPad, Mac, and Apple Vision Pro.
Get KozBon running locally in under a minute:
Then ⌘R in Xcode against an iOS / macOS / visionOS destination. The first run will scan your local network and surface every Bonjour service in reach.
A multi-platform Apple app for discovering, broadcasting, and understanding Bonjour (mDNS/DNS-SD) network services. Rich service inspection, filter categories, a 158-entry service-type library, Siri and Shortcuts actions, and AI-powered Insights — on-device with Apple Intelligence by default, or with your own Anthropic Claude or Google Gemini API key — on iPhone, iPad, Mac, and Apple Vision Pro.
git clone https://github.com/kelvinkosbab/KozBon.git
cd KozBon
open KozBon.xcworkspace
Then ⌘R in Xcode against an iOS / macOS / visionOS destination. The first run will scan your local network and surface every Bonjour service in reach.
To run the SPM package tests (no simulator needed, ~1 second):
swift test --package-path KozBonPackages
For full build commands per platform, see Build below. For contribution guidelines, see CONTRIBUTING.md.
Features
Discover & broadcast
Service discovery — Live scan of your local network with filter categories (Smart Home, Apple Devices, Media & Streaming, Printers & Scanners, Remote Access)
Service broadcasting — Publish custom services with configurable port, domain, and TXT records; edit TXT records on already-published services
158 built-in service types — Pre-configured library covering HTTP, SSH, AirPlay, HomeKit, IPP, Matter, Thread, and more
Custom service types — Define and persist your own entries with IANA-format identifiers
Service details — Hostnames, IP addresses, transport layer, and TXT metadata for any discovered or published service
Context menus everywhere — Copy service names, hostnames, IP addresses, full type strings, and individual TXT keys/values with a long-press
Best-practice footnotes on the create-service-type, broadcast, and add-TXT-record forms
Sectioned Nearby list — grouped one section per device by default, or one per service type (with a short description of the type) when sorting by service name
Siri and Shortcuts
Eight App Intents, with spoken phrases localized in all 8 languages, work from Siri, the Shortcuts app, and Spotlight:
Scan for Services — counts the devices on your network ("an Apple TV is one Apple device, not three") and says when you're not on Wi-Fi or Ethernet
List Discovered Services — returns services whose name, type, hostname, and port flow into later Shortcuts steps
Find Services — "Find printers with KozBon" scans for one category, or one service type in Shortcuts
Check If Device Is Online — answers yes or no for a named device, ready for automations
Open Service in KozBon — opens the app on a service a List or Find step returned
Explain a Service Type — Siri reads a short Insights explanation from the AI selected in Settings
Broadcast a Service / Stop Broadcasting — start a TCP broadcast on a port, or stop them all
A Using Siri with KozBon row (top of Settings and Discover on iPhone, iPad, and Vision Pro) cycles through example phrases, and a Using KozBon with Shortcuts page links to the Shortcuts app with suggested automations. Siri also accepts the common mishearings "Koz Bon", "Kos Bon", and "Cause Bon".
Settings
Dedicated Settings tab for Siri phrases, display options, Insights and AI backend configuration, and library management
Ambient background switch turns off the per-tab colour wash and restores plain system styling; it also steps aside automatically for Increase Contrast, Reduce Transparency, and Smart Invert
Reset to defaults restores built-in settings and clears custom service types
AI insights (pluggable backend)
Insights — Tap the Insights button on any service or library detail page, or long-press a row, to stream a Markdown-formatted explanation of what it does, why devices advertise it, and how to interact with it from your device. The button takes on the selected AI's colours, moves into the navigation bar once it scrolls out of view, and offers sign-in or Apple Intelligence setup when the AI isn't ready
Chat (or Explore on macOS/visionOS) — Conversational assistant grounded in your live network: discovered services with their IP addresses, transport layer, and TXT records; published services; and the type library grouped by category
Query-triggered descriptions — Mention a type by name in chat and the assistant pulls in authoritative descriptions from the catalog inline
Scan freshness awareness — The assistant knows whether results are fresh, stale, or still populating and hedges answers accordingly
Prompt safety — Source-priority hierarchy (TXT > type description > model training), named uncertainty phrasing, TXT-key allowlist, client-side refusal for prompt injection and off-topic queries
Apple Intelligence (default) — Everything runs through Apple's Foundation Models on-device; no data leaves the device. Available on Apple-Intelligence-eligible hardware with the feature turned on.
Anthropic Claude (opt-in) — Sign in with your own Anthropic API key (stored in the iOS Keychain; KozBon never sees the key on a server). Requests are sent to Anthropic's API and billed to your Anthropic account. Available on any device — including older iPhones, non-M-series Macs, and devices where Apple Intelligence is disabled.
Google Gemini (opt-in) — Same shape, with your own Google AI Studio key against the Gemini Developer API.
OpenAI GPT (opt-in) — Same shape, with your own OpenAI API key against the Responses API. Requests opt out of server-side storage. API usage is billed by OpenAI, separately from a ChatGPT subscription. Store listings never name it — see .
Polish
App icon — a blue dot with broken ripples that sweep up and to the right; on visionOS the ripples float above a white back layer
Ambient background — a slow mesh-gradient colour wash that drifts behind every tab, tinted per tab and carried into the detail pages opened from it; holds still under Reduce Motion or Low Power Mode
Liquid Glass — translucent compose bars, tinted send buttons, and floating capsules, routed through one helper that uses .glassEffect(in:) on iOS / macOS and .glassBackgroundEffect() on visionOS
Haptic feedback — medium tap on send, light tap per sentence while the model streams, selection taps on sort and navigation actions
Three-dot typing indicator that ripples leading-to-trailing like iMessage, anchored to the leading edge of the streaming content
Markdown rendering for streaming responses with code, headings, and lists
Pull-to-refresh scanning on iPhone and iPad
Accessibility — VoiceOver labels and hints on every interactive element, region labels, heading traits, .accessibilityIdentifier for UI tests, Reduce Motion support, Dynamic Type throughout
Localized in English, Spanish, French, German, Japanese, Simplified Chinese, Arabic, and Hebrew — with right-to-left layout mirroring for Arabic and Hebrew
Platform support
Platform
Minimum
Platform-specific polish
iOS
26.0
Pull-to-refresh, context menus, interactive keyboard dismiss, Liquid Glass compose bar, landscape and wide layouts including the iPhone Duo inner display
Menu-bar commands, keyboard shortcuts, Settings scene, multi-window service details
visionOS
26.0
Native translucent surfaces, pointer hover, platform-appropriate tab labeling, layered app icon, Siri phrases in Settings
The minimums were raised to 26 together so no #available gating is needed for 26-era APIs — Liquid Glass, Foundation Models, and the Apple Intelligence surfaces are all unconditional.
Architecture
Swift 6.2 with strict concurrency (Sendable, @MainActor, structured concurrency, defer-guarded state resets)
SwiftUI with MVVM, @Observable view models, NavigationSplitView for adaptive list-detail layouts
Modular SPM packages in KozBonPackages/ — 15 modules, each a library product. The KozBon/ Xcode target carries only the @main shim and the resources that must live in the main bundle, so contributors can move fast under swift test without touching project settings:
AppCore — the root scene (AppCoreScene), its view model (AppCoreViewModel), macOS menu commands, and the top-level tab destinations. The Xcode app target's @main struct is a 30-line shim that just renders AppCoreScene().
BonjourCore — shared value types, constants, and utilities (Constants, TransportLayer, InternetAddress, Logger, Clipboard, HapticFeedback). Re-exports Core so downstream modules pick up Logger / Loggable without an explicit import.
BonjourStorage — all persistence in one module: the SwiftData preferences container (PreferencesStore, UserPreferences) and the legacy Core Data custom-service-type store (CustomServiceType, MyCoreDataStack, MyDataManagerObject).
BonjourLocalization — localized strings (8 languages including Arabic and Hebrew) backed by a String Catalog (.xcstrings). Type-safe Strings enum at call sites; no NSLocalizedString literals scattered through the views.
BonjourModels — domain models and the 158-entry service-type library (BonjourServiceType, BonjourService, BonjourServiceSortType); per-type icon / category / detail metadata.
BonjourScanning — Bonjour discovery and publishing (BonjourServiceScanner, MyBonjourPublishManager) with their protocol abstractions and mocks; DependencyContainer for environment-injected DI.
Where the structure comes from
The shape above — a two-file Xcode app target sitting on a 15-module local SPM package — is not ad hoc. It follows the Apple-platform conventions installed by AppBootstrapAI:
apple-modular-architecture.md sets the thin-app-target-over-local-package layout, the one-way module dependency graph, and the rule that a cross-module need moves down a layer rather than becoming a feature-to-feature edge.
apple-spm-package-conventions.md supplies the manifest pattern: one makeTargets(name:…) call per module composed with +, a single sharedSwiftSettings pinning .swiftLanguageMode(.v6), and a uniform {Module}/Sources + {Module}/Tests layout. Adding a module is a two-line change to Package.swift.
18 rules and 10 skills are committed under .claude/, so every agent session in this repo inherits the same conventions — accessibility, localization, Swift 6 strict concurrency, logging, testing strategy, and the TestFlight release path — instead of being told them again each time.
Installed 2026-06-12 from bundle commit b64ec75. .claude/.appbootstrap-manifest.json records the installed file set with checksums, so the bundle can be updated without clobbering local edits.
China mainland storefront — only Apple Intelligence is offered; the Claude, Gemini, and GPT options are withheld, and a stored cloud selection reverts to Apple Intelligence.
The Claude, Gemini, and GPT model pickers load each provider's current model list using your key, so new models appear without an app update; offline they fall back to a compiled-in list. Each provider remembers its own choice, so switching back and forth preserves both.
Configurable — Single Detail level setting (Basic / Technical) drives both vocabulary and response length, so the two settings can't drift out of sync
LocalNetworkMonitor — NWPathMonitor-backed primitive that tells the Discover tab whether the device is on a Wi-Fi or Ethernet path so it can surface a distinct empty state when scanning literally can't reach anything (cellular-only or offline). @MainActor protocol + production class + synchronous mock, driven by AsyncStream for structured cancellation.
BonjourAICore — the provider-agnostic AI layer every backend implements against: protocols (BonjourChatSessionProtocol, BonjourServiceExplainerProtocol), prompt builders (BonjourServicePromptBuilder, BonjourChatPromptBuilder), the input validator and safety rules, the intent broker that bridges tool calls to view-model side effects, shared chat UI primitives, mocks and simulator stubs, and AICloudCredentialsStore with Keychain + in-memory implementations.
BonjourAIApple — the on-device implementation on Apple's FoundationModels, plus AppleIntelligenceSupport for availability gating.
BonjourAIAnthropic — Anthropic Claude over a native URLSession + SSE streaming client (no third-party SDK), with prompt caching for the static system block and AnthropicModelCatalog fetching the live model list.
BonjourAIGemini — Google Gemini against the Gemini Developer API, with GeminiModelCatalog filtered to models advertising generateContent. Keys ride in the x-goog-api-key header, never the query string.
BonjourAIOpenAI — OpenAI GPT against the Responses API, with store: false on every request and OpenAIModelCatalog filtered to chat-capable model families.
BonjourAI — the umbrella. CloudAwareBonjourChatSessionFactory / CloudAwareBonjourServiceExplainerFactory read preferencesStore.aiBackend and route to the right provider; @_exported import BonjourAICore keeps older import BonjourAI call sites working.
BonjourUI — every SwiftUI view, view model, and the design-system primitives consumed by every screen (semantic CGFloat tokens, the Image.xxx SF-Symbol façade, the AmbientMeshBackground colour wash).
BonjourAppIntents — the eight Siri / Shortcuts App Intents (scan, list, find, is-device-online, open, explain, broadcast, stop broadcasting), the BonjourService / BonjourServiceType entity projections, and AppIntentsBridge, which AppCore fills in at launch so actions that touch app state work even when the system launches KozBon in the background. The AppShortcutsProvider itself stays in the app target, because Xcode only extracts and localizes Siri phrases there.
Dependency injection via DependencyContainer + SwiftUI environment; the shared BonjourServicesViewModel is owned by AppCoreViewModel so the Discover and Chat tabs see the same scanner delegate
FoundationModels for on-device AI, with availability gating that hides the AI surfaces on ineligible hardware rather than failing at call time
Pluggable AI backend (ADR 0005, extended by ADR 0006 and ADR 0007) — one protocol layer, four implementations, selected in Settings and swapped without an app restart. Each cloud provider's model list is fetched at runtime from its own API, falling back to a compiled-in list offline, and each remembers its own model choice. API keys live only in the Keychain (whenUnlockedThisDeviceOnly, never iCloud-synced); KozBon never operates them
Liquid Glass via a single platformGlassBackground helper that routes to .glassEffect(in:) on iOS / macOS and .glassBackgroundEffect() on visionOS
HapticFeedback provider injected via environment so view models can request haptics without direct UIKit dependencies
Swift Testing — more than 1,350 tests covering prompt-quality invariants, Siri intent responses, view-model logic, state machines (sentence haptic tracker, broadcast publish flow), design-token value pins, haptic mocks, scanner delegate flows, streaming-client wire formats, and chat-session rejection paths
SwiftLint — project-wide rules plus a custom rule forbidding literal SF Symbol strings in favor of the Image.xxx façade
CI — GitHub Actions workflows for Native CI (iOS + macOS build matrix), SPM package tests, multi-platform builds (macOS + visionOS), SwiftLint, a String Catalog validator that pins translation completeness across all 8 locales, a Markdown link checker, and a Release workflow that publishes a GitHub Release whenever a v* tag is pushed
