|
1
|
<p align="center"> |
|
2
|
<img src="assets/app-icon.png" alt="Siti AI" width="96" height="96"> |
|
3
|
</p> |
|
4
|
|
|
5
|
<h1 align="center">Siti AI</h1> |
|
6
|
|
|
7
|
<p align="center"> |
|
8
|
<strong>A private, on-device AI assistant for macOS, iOS, and Android.</strong><br> |
|
9
|
Your model runs locally. Your chats never leave the device. |
|
10
|
</p> |
|
11
|
|
|
12
|
<p align="center"> |
|
13
|
<a href="https://apps.apple.com/se/app/siti-ai/id6780047972"><img alt="App Store" src="https://img.shields.io/badge/App_Store-live-06a561?style=flat-square&labelColor=1a1a1a"></a> |
|
14
|
<a href="https://play.google.com/store/apps/details?id=ai.siti.Siti"><img alt="Google Play" src="https://img.shields.io/badge/Google_Play-live-06a561?style=flat-square&labelColor=1a1a1a"></a> |
|
15
|
<a href="https://github.com/ondeinference/onde"><img alt="Built with Onde" src="https://img.shields.io/badge/built_with-Onde-06a561?style=flat-square&labelColor=1a1a1a"></a> |
|
16
|
<a href="LICENSE"><img alt="License: Apache-2.0" src="https://img.shields.io/badge/license-Apache_2.0-blue?style=flat-square"></a> |
|
17
|
<a href="https://github.com/sponsors/ondeinference"><img alt="Sponsor" src="https://img.shields.io/badge/♥-Sponsor-db61a2?style=flat-square"></a> |
|
18
|
</p> |
|
19
|
|
|
20
|
<p align="center"> |
|
21
|
<a href="https://getsiti.5mb.app">Website</a> · |
|
22
|
<a href="#quick-start">Quick start</a> · |
|
23
|
<a href="#architecture">Architecture</a> · |
|
24
|
<a href="CONTRIBUTING.md">Contributing</a> · |
|
25
|
<a href="SECURITY.md">Security</a> |
|
26
|
</p> |
|
27
|
|
|
28
|
--- |
|
29
|
|
|
30
|
Siti AI is the assistant you already know, rebuilt around privacy. It loads an |
|
31
|
open-weights language model onto your device and runs inference entirely |
|
32
|
locally — no server round-trip, no account, nothing to leak. The name is a wink |
|
33
|
at the assistant on your phone; the difference is where your words go. |
|
34
|
|
|
35
|
**Siti is also the flagship open reference app for [Onde](https://ondeinference.com)** — |
|
36
|
the on-device inference engine that powers it. If you want to see Onde driving a |
|
37
|
real, shipping app on the App Store and Google Play, this is the codebase to read. |
|
38
|
|
|
39
|
## Why it exists |
|
40
|
|
|
41
|
The device you keep closest to you should be the one that keeps your data |
|
42
|
private. Most assistants send everything you type to a remote server; Siti keeps |
|
43
|
it on your device instead. Using AI shouldn't be the price you pay to give that |
|
44
|
up. |
|
45
|
|
|
46
|
## Features |
|
47
|
|
|
48
|
- **Fully on-device inference** — chat runs locally via [Onde](https://crates.io/crates/onde): Metal-accelerated on Apple silicon, CPU on Android. |
|
49
|
- **Bring your own model** — pick from on-device open-weights models; download and swap them in-app. |
|
50
|
- **No account, no telemetry of your chats** — conversations stay on the device. |
|
51
|
- **Native and small** — Rust core + Tauri, a real app rather than a web wrapper. |
|
52
|
|
|
53
|
## Quick start |
|
54
|
|
|
55
|
**Prerequisites:** [Rust](https://rustup.rs) (stable), [pnpm](https://pnpm.io), |
|
56
|
and the [Tauri prerequisites](https://tauri.app/start/prerequisites/) for your |
|
57
|
platform (Xcode for macOS/iOS; Android SDK + NDK for Android). |
|
58
|
|
|
59
|
```bash |
|
60
|
git clone https://github.com/ondeinference/siti.git |
|
61
|
cd siti |
|
62
|
pnpm install |
|
63
|
make dev # or: pnpm tauri dev |
|
64
|
``` |
|
65
|
|
|
66
|
The first launch downloads a model; after that, inference is offline. The Onde |
|
67
|
engine is pulled from crates.io — no separate SDK checkout needed. To develop |
|
68
|
against a local Onde checkout, see the commented `[patch.crates-io]` block in |
|
69
|
`src-tauri/Cargo.toml`. |
|
70
|
|
|
71
|
Common tasks: |
|
72
|
|
|
73
|
```bash |
|
74
|
make dev # start the Tauri dev server |
|
75
|
make build # debug build |
|
76
|
make fmt # cargo fmt |
|
77
|
make lint # clippy + tsc |
|
78
|
``` |
|
79
|
|
|
80
|
## Architecture |
|
81
|
|
|
82
|
Siti is a thin, native shell around the Onde inference engine. |
|
83
|
|
|
84
|
``` |
|
85
|
┌──────────────────────────────┐ |
|
86
|
│ React + Vite UI (src/) │ chat view, model settings |
|
87
|
├──────────────────────────────┤ |
|
88
|
│ Tauri command layer │ src-tauri/src/chat/command_*.rs |
|
89
|
│ (load/unload model, send, │ typed IPC between UI and Rust |
|
90
|
│ stream tokens, history) │ |
|
91
|
├──────────────────────────────┤ |
|
92
|
│ Onde engine (crates.io) │ GGUF load + streaming generation; |
|
93
|
│ Metal on Apple, CPU on │ Metal-accelerated on Apple silicon |
|
94
|
│ Android │ |
|
95
|
└──────────────────────────────┘ |
|
96
|
``` |
|
97
|
|
|
98
|
Things worth reading if you're evaluating the engineering: |
|
99
|
|
|
100
|
- `src-tauri/src/chat/` — the command surface: model lifecycle, streaming token |
|
101
|
events, and history. This is the cleanest example of wiring Onde into an app. |
|
102
|
- `src-tauri/src/chat/mod.rs` — token-budget heuristics (reasoning models burn |
|
103
|
a large `<think>` budget before the visible reply begins). |
|
104
|
- `src-tauri/src/setup/` — redirecting the model cache into the App Group |
|
105
|
container so the sandbox doesn't fight the download. |
|
106
|
- `src-tauri/.cargo/config.toml` — the `+fullfp16` target-feature notes for |
|
107
|
Apple silicon fp16 SIMD. Non-obvious, and load-bearing. |
|
108
|
|
|
109
|
## Repository layout |
|
110
|
|
|
111
|
| Path | What | |
|
112
|
| ---------- | ----------------------------------------------- | |
|
113
|
| `/` | Tauri + Vite desktop/mobile app (the product) | |
|
114
|
| `src/` | React UI | |
|
115
|
| `src-tauri/` | Rust core, Tauri commands, platform config | |
|
116
|
| `web/` | Next.js marketing + legal site (getsiti.5mb.app) | |
|
117
|
|
|
118
|
## The Onde SDK |
|
119
|
|
|
120
|
The inference engine is developed separately and is open source under |
|
121
|
MIT OR Apache-2.0, with SDKs for several ecosystems: |
|
122
|
|
|
123
|
[Rust](https://crates.io/crates/onde) · |
|
124
|
[Swift](https://github.com/ondeinference/onde-swift) · |
|
125
|
[Kotlin Multiplatform](https://central.sonatype.com/artifact/com.ondeinference/onde-inference) · |
|
126
|
[Flutter](https://pub.dev/packages/onde_inference) · |
|
127
|
[React Native](https://www.npmjs.com/package/@ondeinference/react-native) · |
|
128
|
[CLI](https://github.com/ondeinference/onde-cli) · |
|
129
|
[Docs](https://ondeinference.com/sdk) |
|
130
|
|
|
131
|
If Siti is useful to you, the engine underneath it probably is too. |
|
132
|
|
|
133
|
## Contributing |
|
134
|
|
|
135
|
Contributions are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md). We ask |
|
136
|
contributors to sign a lightweight [CLA](CLA.md) so the project can keep both |
|
137
|
its open-source and commercial licensing options open. Start with issues labeled |
|
138
|
[`good first issue`](https://github.com/ondeinference/siti/labels/good%20first%20issue). |
|
139
|
|
|
140
|
## Sponsor |
|
141
|
|
|
142
|
Siti and Onde are built by a small team ([Splitfire AB](https://ondeinference.com)). |
|
143
|
On-device AI that keeps user data private is worth funding — if you agree, |
|
144
|
[**sponsor us on GitHub**](https://github.com/sponsors/ondeinference). Sponsorship |
|
145
|
funds the engine, the SDKs, and this app. |
|
146
|
|
|
147
|
## Security & privacy |
|
148
|
|
|
149
|
Chat content is processed on-device and is not transmitted to us. To report a |
|
150
|
vulnerability, see [SECURITY.md](SECURITY.md) — please do not open a public |
|
151
|
issue for security reports. |
|
152
|
|
|
153
|
## License |
|
154
|
|
|
155
|
Code is licensed under [Apache-2.0](LICENSE). The "Siti AI" name, logos, and app |
|
156
|
icons are trademarks of Splitfire AB and are **not** covered by the code license — |
|
157
|
see [TRADEMARK.md](TRADEMARK.md). |