Runic

An AI-native browser for iOS and macOS. Runic wraps a multi-tab WebKit browser around two agents: a voice agent that acts while you are still speaking, and a browser agent that clicks, types and navigates the web to complete tasks on its own.

Built with SwiftUI and targeting iOS 26 / macOS 26.

Features

Voice agent: faster than finishing your sentence

Streams microphone audio to Deepgram for live transcription and dispatches actions incrementally, before you stop talking:

1. Interim transcript: instant navigation to known sites via regex fast-path
2. Interim stable for 300ms: LLM evaluates whether the partial command is already actionable
3. Final transcript: LLM catches anything not yet dispatched

A rolling session context (last 20 actions) is hydrated into each LLM call, so follow-ups like "open the second one" resolve against what just happened.

Browser agent: autonomous task completion

A tool-calling LLM loop that operates the browser like a user would. The page's interactive elements (buttons, links, inputs) are extracted from the DOM with bounding boxes and indexed, then the agent acts through a fixed tool set:

Tool	What it does
click	Tap an indexed element
type_text	Fill an input, optionally press Enter
scroll	Scroll the page up or down
navigate / go_back	URL navigation
new_tab / switch_tab	Multi-tab control
wait	Pause for slow pages
evaluate_js	Run JavaScript directly
fork_tasks	Spawn parallel subtasks across tabs (e.g. compare prices on three sites at once)

Built-in platform shortcuts route obvious tasks straight to YouTube, Amazon, Reddit, GitHub and others instead of detouring through Google. A site cache remembers which URLs satisfied past tasks. A live agent log shows every thought and action as it happens.

On-device task classification

On iOS 26+ devices, Apple's FoundationModels framework classifies task complexity (simple, moderate or complex) on-device to route between fast paths and full agent runs, with no server round-trip. Older devices fall back to sub-millisecond heuristics.

Full browser underneath

• Multi-tab browsing on WKWebView with a Safari user agent
• Start page, navigation bar and a unified bottom input bar for both typed and spoken commands
• Scroll-direction-aware chrome (UI collapses as you scroll)

Accounts and subscriptions

• Backend auth (register / login) with token persistence
• StoreKit 2 monthly and yearly premium subscriptions with a transaction listener and restore support

Architecture

RunicApp
└── ContentView
    ├── TabManager          WKWebView pool, DOM extraction, JS bridge
    ├── BrowserAgent        tool-calling agent loop (OpenAI-compatible API)
    │   ├── OpenAIService   chat completions and tool definitions
    │   ├── SiteCache       task-to-URL memory
    │   └── BackendService  auth, chat proxy, task history
    ├── VoiceAgent          Deepgram streaming and incremental dispatch
    │   └── SessionContext  rolling action history for LLM hydration
    ├── OnDeviceClassifier  FoundationModels complexity routing
    └── StoreManager        StoreKit 2 subscriptions

Everything is SwiftUI + @Observable, with @MainActor isolation where it touches WebKit.

Getting started

Requirements

• Xcode 26+
• iOS 26 / macOS 26 deployment target
• An OpenAI API key (or any OpenAI-compatible endpoint, the base URL is configurable)
• A Deepgram API key (for the voice agent)

Setup

1. Clone the repo:

   git clone https://github.com/unixers/runic.git
   cd runic

2. Create your secrets file:

   cp Runic/Secrets.swift.example Runic/Secrets.swift

   Then fill in your keys:

   enum Secrets {
       static let openAIAPIKey  = "sk-..."
       static let deepgramAPIKey = "..."
   }

   Secrets.swift is gitignored. It never leaves your machine.

3. Open Runic.xcodeproj in Xcode and run.

Usage

• Type a task in the bottom input bar ("find the cheapest flight from SFO to Tokyo in November") and watch the agent work in the live log.
• Hold to speak. Commands dispatch as you talk: "Open YouTube and search for..." starts navigating before you finish the sentence.
• Parallel tasks: ask for comparisons ("compare this laptop's price on Amazon and eBay") and the agent forks subtasks across tabs.

Project layout

Runic/
├── RunicApp.swift            App entry point
├── ContentView.swift         Main browser UI
├── TabManager.swift          WKWebView, tab management, JS bridge
├── BrowserAgent.swift        Agent loop and action execution
├── VoiceAgent.swift          Deepgram streaming and incremental dispatch
├── OnDeviceClassifier.swift  FoundationModels task routing
├── OpenAIService.swift       LLM API client and tool schemas
├── BackendService.swift      Auth, chat proxy, tasks
├── SiteCacheImpl.swift       Task-to-URL memory
├── StoreManager.swift        StoreKit 2 subscriptions
├── Models.swift              Shared data models
└── *View.swift               SwiftUI components

License

All rights reserved.