Private. Local. Yours.
Next-generation on-device local LLM & autonomous agent client built with Kotlin Multiplatform and Google LiteRT-LM
English • 简体中文 • 繁體中文 • 日本語 • 한국어 • Deutsch • Español • Français • Русский
Agro is a pure on-device, multi-platform (Android, iOS, macOS, Windows, Linux) local Large Language Model (LLM) and autonomous agent chat application.
Unlike conventional AI tools that depend on remote cloud APIs, Agro brings the entire inference engine, context scheduling, agent tool execution, and AIGC rendering directly to your physical device. Powered by Google LiteRT-LM native C++ runtime and hardware acceleration backends (Apple Metal, WebGPU Dawn, DirectX, Vulkan, OpenCL), Agro guarantees 100% data privacy sovereignty while delivering an ultra-smooth, responsive generative experience.
- 🔒 100% Local & Offline-First: All conversations, historical records, and inference calculations remain strictly on your device—zero data sent to the cloud.
- ⚡ Native Hardware Acceleration: Low-level memory management and GPU optimizations tailored for Desktop GPUs, Apple Neural Engine / Metal, and Mobile GPUs.
- 🧩 Autonomous Agent Loop & Tool Ecosystem: Built-in structured tool-calling loop supporting local JavaScript execution, real-time web search, and URL content analysis.
- 🎨 Multimodal Structured Generation: Beyond plain text, natively synthesizes and renders SVG vector graphics, 8-bit chiptune audio, and Lottie animations on-device in real time.
| Capability | Implementation Details |
|---|---|
| Local LLMs | Loads LiteRT-LM compatible .litertlm bundles. Built-in one-click downloads for Ministral-3-3B & Gemma 4 4B, with support for custom models |
| Streaming Chat | Incremental token streaming, thinking/reasoning channels, mid-generation cancellation, auto CPU fallback upon GPU errors, runtime telemetry |
| Context Management | Segregated channels for conversational chat and structured synthesis; reads native token counts, performs budget estimation, history replay, and pruning |
| Agent Loop | Model tool call request → Host execution → Structured payload feedback → Continued inference (up to 10 iterations by default) |
| Built-in Web Tools | Pre-registered searchWeb (Bing search) and analyzeUrl (HTTP/HTTPS content fetch & analysis) |
| Creative Modes | 4 distinct session protocols: Standard Assistant, SVG Vector Graphics, 8-bit BGM Composer, and Lottie Micro-animation |
| Local Persistence | Room KMP + Bundled SQLite; safely persists conversations, messages, tool execution logs, and system prompt snapshots |
| Model Parameters | Granular control over Temperature, Top-P, Top-K, Context Window limits, Thinking mode, Speculative Decoding, and custom System Prompts |
| Chat Interface | Home Screen |
|---|---|
 |
 |
| Model Library | Settings Panel |
 |
 |
| Mobile Chat | Mobile Home | Mobile Library | Mobile Settings |
|---|---|---|---|
 |
 |
 |
 |
Download precompiled release binaries directly from the Releases Page:
| OS | Distribution Format | Download Entry | Installation & Runtime Notes |
|---|---|---|---|
| Android | apk |
Agro-Android.apk | Requires Android 10.0+ (API 29+); ARM64 devices recommended |
| iOS | ipa |
Agro-iOS.ipa | Requires iOS 16.0+; sideload via AltStore / TrollStore / Xcode |
| Windows | exe / msi / zip |
Agro-Windows-x86_64.zip | Windows 10/11 x86_64 recommended; includes bundled DirectX/WebGPU runtimes |
| macOS | dmg |
Agro-macOS-arm64.dmg | Apple Silicon (M1/M2/M3/M4) native. If blocked by Gatekeeper, run:sudo xattr -d com.apple.quarantine /Applications/Agro.app |
| Linux | AppImage / deb / rpm |
Agro-Linux-x86_64.AppImage | Major x86_64 distributions (Ubuntu, Fedora, Arch, etc.) |
Agro adheres to standard Kotlin Multiplatform (KMP) + Clean Architecture / MVVM principles:
graph TD
UI[Compose Multiplatform UI Layer] --> VM[ChatViewModel & ContextCoordinator]
VM --> CS[ContextStrategy & Token Budget]
VM --> Parser[Multimedia Engine<br>LottieSceneCompiler / MML Synth / SVG Sanitizer]
VM --> Bridge[Native Bridge Layer]
Bridge -->|JVM JNI / Direct Pointers| DesktopJNI[Desktop / Android C++ JNI]
Bridge -->|Kotlin/Native cinterop| iOSCAPI[iOS C-API Bridge]
DesktopJNI --> Engine[Google LiteRT-LM Native Engine]
iOSCAPI --> Engine
Engine --> Acc[Hardware Accelerators<br>Metal / WebGPU Dawn / OpenCL]
Engine --> LocalModel[(On-Device Task / Bin Models<br>Gemma 2 / 3 / 4, etc.)]
Agro/
├── composeApp/ # Main application entry, Compose Multiplatform UI, ViewModel, AIGC compilers, and JNI bridges
│ └── src/
│ ├── commonMain/ # Shared UI (Screens, Theme, Components, Navigation) & business pipelines
│ ├── androidMain/ # Android native configurations, Activities & JNI bindings
│ ├── desktopMain/ # Desktop JVM native integration, shared library unpacker & platform windows
│ └── iosMain/ # iOS cinterop, AVAudioPlayer & Native bridge
├── shared/ # Cross-platform shared business logic & core coordinators
├── ui-theme/ # Ethereal Minimalism design system tokens (Color, Typography, Shape, Spacing)
├── data-network/ # Ktor 3.x cross-platform network client & Sandwich response models
├── data-model/ # Pure Kotlin domain data models & serialization schemas
└── cpp/ # Native C++ workspace & LiteRT-LM precompiled runtime libraries
└── lite-rt-lm/ # Google AI Edge LiteRT-LM sources, C-API headers & Bazel build blueprints
- Core Language & Platform: Kotlin
2.4.0, Kotlin Multiplatform, Kotlinx Coroutines1.10.2, Kotlinx Serialization1.8.0 - UI Framework: Compose Multiplatform
1.11.1, Material 3 Adaptive1.1.2, Compose Rich Editor1.0.0-rc14 - Animations & Multimedia: Compottie
2.0.0-rc04(Lottie), ComposeMediaPlayer0.11.3, Coil 33.5.0(SVG support) - Dependency Injection & Architecture: Koin
4.1.1(Core, Compose, ViewModel), AndroidX Lifecycle ViewModel2.9.6, Navigation 3 UI - Data Storage: AndroidX Room
2.8.4(KMP), SQLite Bundled2.6.2, Okio3.15.0, FileKit0.14.2 - Networking & Parsers: Ktor
3.2.3, Sandwich2.1.2, Ksoup0.2.6, QuickJS-kt1.0.0-alpha13 - Low-Level AI Runtime: Google LiteRT-LM (C++ Runtime), Metal / WebGPU Dawn / OpenCL / Vulkan acceleration, Bazel
- JDK: Java 21 (Recommend JetBrains Runtime 21 or Eclipse Temurin 21)
- Android Development: Android Studio Ladybug+, Android SDK 36, Android NDK
27.0.12077973 - iOS / macOS Development: macOS environment, Xcode 15+, Command Line Tools
- Bazelisk: Install Bazelisk for building or aligning LiteRT-LM native C++ dependencies
- Git LFS: Ensure Git LFS is installed prior to cloning to pull precompiled native libraries and assets:
git lfs install
# Clone repository and submodules
git clone --recursive https://github.com/Onion99/Agro.git
cd Agro
# Fetch Git LFS binary assets
git lfs install
git lfs pull./gradlew :composeApp:runConnect your Android device or launch an emulator, then run:
./gradlew :composeApp:installDebugOpen iosApp/iosApp.xcworkspace in Xcode, select your Target/Simulator for code signing and running, or build the framework via Gradle:
./gradlew :composeApp:embedAndSignAppleFrameworkForXcode# Execute common unit tests and desktop tests
./gradlew :composeApp:desktopTest
./gradlew :composeApp:allTests# Package desktop distribution binaries for the current OS (msi / dmg / deb / rpm)
./gradlew :composeApp:packageDistributionForCurrentOS
# Build Android release APK
./gradlew :composeApp:assembleRelease- This project is open-sourced under the GNU General Public License v3.0 (GPL-3.0).
- Special thanks to the following open-source projects and communities:
- Google LiteRT-LM (LiteRT) — Ultra-fast on-device inference core
- JetBrains Compose Multiplatform — Cross-platform declarative UI framework
- Compottie — Lottie renderer for Compose Multiplatform
- ComposeMediaPlayer — Lightweight cross-platform media player
- Compose Rich Editor — Rich text editor for Compose Multiplatform
- QuickJS-kt — Embedded Kotlin JavaScript engine
- Coil — Asynchronous image loading for Kotlin
- Koin & Ktor — Modern DI & asynchronous networking stack
Made with 💚 by Onion99 and the Open Source Community