Siti AI

A private, on-device AI assistant for macOS, iOS, and Android.
Your model runs locally. Your chats never leave the device.

Website · Quick start · Architecture · Contributing · Security

---

Siti AI is the assistant you already know, rebuilt around privacy. It loads an open-weights language model onto your device and runs inference entirely locally — no server round-trip, no account, nothing to leak. The name is a wink at the assistant on your phone; the difference is where your words go.

Siti is also the flagship open reference app for Onde — the on-device inference engine that powers it. If you want to see Onde driving a real, shipping app on the App Store and Google Play, this is the codebase to read.

Why it exists

The device you keep closest to you should be the one that keeps your data private. Most assistants send everything you type to a remote server; Siti keeps it on your device instead. Using AI shouldn't be the price you pay to give that up.

Features

• Fully on-device inference — chat runs locally via Onde: Metal-accelerated on Apple silicon, CPU on Android.
• Bring your own model — pick from on-device open-weights models; download and swap them in-app.
• No account, no telemetry of your chats — conversations stay on the device.
• Native and small — Rust core + Tauri, a real app rather than a web wrapper.

Quick start

Prerequisites: Rust (stable), pnpm, and the Tauri prerequisites for your platform (Xcode for macOS/iOS; Android SDK + NDK for Android).

git clone https://github.com/ondeinference/siti.git
cd siti
pnpm install
make dev          # or: pnpm tauri dev

The first launch downloads a model; after that, inference is offline. The Onde engine is pulled from crates.io — no separate SDK checkout needed. To develop against a local Onde checkout, see the commented [patch.crates-io] block in src-tauri/Cargo.toml.

Common tasks:

make dev      # start the Tauri dev server
make build    # debug build
make fmt      # cargo fmt
make lint     # clippy + tsc

Architecture

Siti is a thin, native shell around the Onde inference engine.

┌──────────────────────────────┐
│  React + Vite UI (src/)       │  chat view, model settings
├──────────────────────────────┤
│  Tauri command layer          │  src-tauri/src/chat/command_*.rs
│  (load/unload model, send,    │  typed IPC between UI and Rust
│   stream tokens, history)     │
├──────────────────────────────┤
│  Onde engine (crates.io)      │  GGUF load + streaming generation;
│  Metal on Apple, CPU on       │  Metal-accelerated on Apple silicon
│  Android                      │
└──────────────────────────────┘

Things worth reading if you're evaluating the engineering:

• src-tauri/src/chat/ — the command surface: model lifecycle, streaming token events, and history. This is the cleanest example of wiring Onde into an app.
• src-tauri/src/chat/mod.rs — token-budget heuristics (reasoning models burn a large <think> budget before the visible reply begins).
• src-tauri/src/setup/ — redirecting the model cache into the App Group container so the sandbox doesn't fight the download.
• src-tauri/.cargo/config.toml — the +fullfp16 target-feature notes for Apple silicon fp16 SIMD. Non-obvious, and load-bearing.

Repository layout

Path	What
/	Tauri + Vite desktop/mobile app (the product)
src/	React UI
src-tauri/	Rust core, Tauri commands, platform config
web/	Next.js marketing + legal site (getsiti.5mb.app)

The Onde SDK

The inference engine is developed separately and is open source under MIT OR Apache-2.0, with SDKs for several ecosystems:

Rust · Swift · Kotlin Multiplatform · Flutter · React Native · CLI · Docs

If Siti is useful to you, the engine underneath it probably is too.

Contributing

Contributions are welcome — see CONTRIBUTING.md. We ask contributors to sign a lightweight CLA so the project can keep both its open-source and commercial licensing options open. Start with issues labeled good first issue.

Sponsor

Siti and Onde are built by a small team (Splitfire AB). On-device AI that keeps user data private is worth funding — if you agree, sponsor us on GitHub. Sponsorship funds the engine, the SDKs, and this app.

Security & privacy

Chat content is processed on-device and is not transmitted to us. To report a vulnerability, see SECURITY.md — please do not open a public issue for security reports.

License

Code is licensed under Apache-2.0. The "Siti AI" name, logos, and app icons are trademarks of Splitfire AB and are not covered by the code license — see TRADEMARK.md.