Persona

Demo agent Aurora — general assistant Olive — bookings & concierge Max — product support Aria — agent studio (meta) Sol — daily companion

Switching resets the chat. Same engine, different hat.

Agent brain

Brain Mock (offline demo) NVIDIA NIM (prototyping) Local — in-browser (wllama) API key Endpoint WebGPU model Qwen2.5 0.5B (smallest — try on iPhone) Llama 3.2 1B (balanced) Qwen2.5 1.5B (desktop) Llama 3.2 3B (desktop, best) CPU model (fallback)

Paste a free key from build.nvidia.com. Sent via the proxy Worker (avoids CORS). Stored in this browser only — prototyping.

Runs entirely on your device, no key. Uses WebGPU (WebLLM) when available — desktop, newer Android, iPhone 15 Pro+/iOS 18+ — else falls back to CPU (wllama). First reply downloads the model once.

Appearance

Orb style Plasma (liquid) Constellation (mesh)

Voice

Test sound ▶ Beep

Tap Beep — if you hear nothing, it's the iPhone silent switch (flip the side switch off). Native apps bypass it; Safari can't.

Engine Device voice (native, fast) Kokoro (on-device, EN) Google Gemini (cloud)

Device = phone/OS voices (instant). Kokoro = on-device English. Google = cloud, 24 languages, plays reliably on iPhone (needs your free AI Studio key).

Google key Google voice Kore (F) Aoede (F) Leda (F) Puck (M) Charon (M) Fenrir (M)

Free key at aistudio.google.com → Get API key. Stored in this browser only.

Quality auto fp32 (best) q8 (small) q4 (tiny) Voice af_heart (US F) af_bella (US F) af_nicole (US F) am_michael (US M) bf_emma (UK F) bm_george (UK M)

Voice model loads on first reply.

Done