I rebuilt Siri AI from scratch and open sourced it

A voice-first, open-source AI assistant for your desktop. Wake it, talk to it, and a tool-using agent talks back — in a cinematic interface you choose.

Or browse every version on the Releases page.

The app auto-updates: it checks GitHub Releases on launch (and hourly), downloads new versions in the background, and prompts you to restart when one is ready.

OpenDex is a desktop app that turns any LLM into a hands-free, Iron-Man-style voice assistant. Say the wake word (or push to talk), speak naturally, and the agent thinks, uses tools, and replies out loud — streaming its answer into a live visualization.

It's a harness, not a single bot: the model, the voice, the wake/transcription engines, the on-screen theme, the greeting, and the agent's skills are all configurable, and it can run fully offline and free (local speech in, local speech recognition, system voice out) — and on a Mac with Apple Intelligence, even the model runs on-device, so the whole loop is free with no key.

• 🎙️ Voice-first loop— wake word → listen → think (with tools) → speak, plus natural follow-ups and opt-in barge-in (interrupt mid-reply). - 🧠 Bring any model— pick your provider in setup:** Apple Intelligence**(on-device, free, no key — macOS), your own** OpenAIor Anthropickey, or the Vercel AI Gateway**(one key → Claude, GPT, Gemini, and more).(An OpenDex hosted subscription — sign in, no keys, cloud-synced — is coming soon.) - 🆓 Free & offline option— Vosk wake word + local Whisper transcription (WASM, no signup) and your OS's built-in voice. No data leaves the machine except the LLM call. - 🔌 Pluggable voice I/O— wake via push-to-talk, Vosk, or Web Speech; transcribe via local Whisper/Vosk, OpenAI, or Web Speech; speak via ElevenLabs or system TTS. - 🎨 Full-interface themes— the themeisthe whole UI: a cinematicJarvis HUD with an animated arc reactor, a minimalTalking Dot, or a** Typing Cursorterminal. All react to your voice. - 🛠️ Agentic skills with a permission gate— the agent can take real actions (e.g. open apps & URLs); sensitive actions pop an Allow once / Always / Deny**prompt that's remembered per skill. - 🖥️ Computer-use (opt-in)— let itsee the screen and drive the mouse & keyboardto operate apps for you. Works with any vision model (screenshots stream back as images), and stays behind the permission gate. - 🔐 Secure by design— API keys are encrypted with your OS keychain and live only in the main process, never in the UI.

OpenDex is fully theamable, you can change anything about the user interface and make it yours.

git clone https://github.com/wassgha/opendex.git
cd opendex
pnpm install
pnpm dev            # launches the OpenDex desktop window

On first launch a short onboarding wizard walks you through choosing a model provider, voice, wake/transcription engine, theme, and greeting. Everything is changeable later from the Settings gear (⚙).

Pick where the thinking happens — every part of the loop can be free/offline:

Model:Apple Intelligence(on-device, free, no key — macOS only), your own OpenAI/** Anthropickey, or the Vercel AI Gateway**(one key, any provider).** Voice out:"System voice" (free) or ElevenLabs (key). Voice in:local Whisper**/** Vosk**(free, offline, one-time model download) or OpenAI Whisper (key).** Wake:**push-to-talk / Vosk (free, offline) or Web Speech (browser).

On a Mac with Apple Intelligence enabled, the whole loop (model + speech + voice) runs locally with no keys at all.

Keys are normally entered in-app and stored encrypted. For development you can seed them via .env

(used only as a fallback):

cp .env.local.example .env

The chat provider needs

oneofAI_GATEWAY_API_KEY

/OPENAI_API_KEY

/ANTHROPIC_API_KEY

— matching the provider you select. Apple Intelligence needs none.

The agent's capabilities are skills — declarative tool bundles. Sensitive ones run behind a permission gate: when the model wants to act, OpenDex s and asks, and your choice (Allow once / Always / Never) is remembered. Built-in skills today: Open apps & URLs, and Control the computer (screen capture + mouse/keyboard — opt-in, off by default).

Computer-use setup (macOS):enableControl the computerinSettings → Skills & tools, then grant OpenDexScreen RecordingandAccessibilitypermission inSystem Settings → Privacy & Security(without them, screenshots come back blank and clicks do nothing). It's powerful — keep the permission onAsk, and "Allow once" covers the whole task it's working on.

• Electron shell + secure agent/TTS-over-IPC
• Config, onboarding & OS-keychain key storage
• Full-interface themes (Jarvis HUD · Talking Dot · Typing Cursor)
• Pluggable wake-word + speech-to-text (incl. free offline Whisper & Vosk)
• Skills + permission gate (Open apps & URLs) - Computer-use — screen capture + mouse/keyboard control, gated & opt-in
• Pluggable model providers — Apple on-device, OpenAI/Anthropic keys, AI Gateway
• OpenDex hosted subscription — sign in, no keys, cloud-synced settings & history
• MCP servers + more built-in skills (shell, filesystem, …)
• Signed GitHub releases + auto-update

Electron · electron-vite · React 19 · Tailwind CSS 4 · Vercel AI SDK v6 · ElevenLabs · Vosk · transformers.js (Whisper) — all local speech engines are WASM. The only native module is nut.js (computer-use input control); it ships prebuilt N-API binaries per platform.

MIT — contributions welcome.