💬
Chat UI
ChatGPT-style interface with sidebar, session history, auto-titles, date grouping, search, and math rendering via KaTeX.
Material 3KaTeX
On-deviceno cloud, no account
AGPL v3free forever
llama.cppvia JNI fork+execve
API serverOpenAI + Anthropic
Features
A full LLM workstation in your pocket.
PhonoLab is not a thin wrapper around an API. It runs the full llama.cpp inference stack on your device with a polished native Android UI.
📄
Document RAG
Attach PDF, text, or DOCX files. Text is extracted, chunked with sliding window overlap, and relevant chunks are injected into your prompt.
RAG1500-char chunks
🖼️
Vision Models
Attach images from your gallery. Auto-detects vision models and pairs mmproj files. Supports Llama 3.2 Vision, LLaVA, and more.
multimodalauto-detect
🌐
LAN API Server
Expose an OpenAI and Anthropic compatible API on your local network. SSE streaming, separate local/LAN API keys, configurable port.
OpenAIAnthropic
📦
Model Catalog
Built-in catalog of small models optimized for mobile: SmolLM2 360M, Qwen 0.5B, Llama 3.2 1B, Qwen 1.5B, TinyLlama 1.1B.
resumableauto-quant
🛡️
Crash-Proof
17-layer error handling. JNI guards, lifecycle-safe UI updates, thread-safe date formatting, JSON parsing safety, body size limits.
17 layersrestart dialog
Architecture
How it works under the hood.
PhonoLab bundles llama-server as a native binary in the APK. At install time, Android extracts it to nativeLibraryDir. At runtime, JNI fork+execve starts the server - no W^X violations, no root needed.
┌─ PhonoLabApp ────────────────────────────────────────┐
│ store, sessionManager, modelManager, runtime │
│ showError(fatal) · safeRunState() · crash handler │
├─ MainActivity (ChatFragment.Host) ─────────────────────┤
│ Toolbar + DrawerLayout · Fragments · Session sidebar │
├─ ChatFragment ─────────────────────────────────────────┤
│ ChatAdapter · Attachments · RAG · Model picker │
├─ LlamaRuntime → LlamaCppManager → JNI runner.cpp ──────┤
│ load() · generate() · abort() · nativeLoaded guard │
├─ PhonoLabApiServer ────────────────────────────────────┤
│ OpenAI + Anthropic · SSE · 1MB limit · 4-thread pool │
├─ nativeLibraryDir/ ────────────────────────────────────┤
│ libllama_server.so (bundled binary) │
│ librunner.so (JNI fork+execve wrapper) │
└───────────────────────────────────────────────────────┘
Ecosystem
NativeLab + PhonoLab. Same vision, different screens.
NativeLab runs on desktop (Linux, macOS, Windows). PhonoLab runs on Android. Both are local-first, privacy-first, and free forever under AGPL v3.
NativeLab (Desktop)
PhonoLab (Android)
PyQt6 GUI + Terminal CLI
Desktop workbench with visual pipeline builder, Labs, MCP, integrations.
Material 3 Android UI
Chat, attachments, RAG, model catalog, LAN API server.
llama.cpp CLI / server
Subprocess wrapper, optional HF Transformers backend, Ollama bridge.
llama-server via JNI
Bundled binary, fork+execve, no W^X issues, vision model support.
Visual Pipeline Builder
20+ node types, AI Builder, save/load JSON, native graph helpers.
Document RAG
PDF/text/DOCX extraction, sliding window chunking, keyword retrieval.
Python 3.10+
pip install nativelab, cross-platform.
Android 5.0+ (API 21)
APK install, arm64-v8a + armeabi-v7a.
AGPL v3
Free and open source forever.
AGPL v3
Free and open source forever.
Reliability
17 layers of error handling.
Every external call - JNI, network, file I/O, JSON parsing - is guarded. Fatal errors show a restart dialog. Non-fatal errors show a red banner that auto-dismisses. The app never crashes silently.
JNI Guards
nativeLoaded flag. All native calls check the flag first. UnsatisfiedLinkError caught gracefully.
Lifecycle Safety
runOnUi{} checks isAdded before touching views. Worker executors shut down in onDestroyView.
Thread Safety
ThreadLocal date formatters. @Volatile abort flag. Socket write guards for client disconnects.
Data Safety
optJSONObject() instead of getJSONObject(). Session logs capped at 500. renameTo() copy+delete fallback.
Get started
Build and run in 3 steps.
Open the Android project in Android Studio, or build from the command line. The APK bundles everything you need.
Clone the repo
$ git clone https://github.com/7ZoneSystems/NativeLab.git $ cd NativeLab/PhonoLab/android
Download llama.cpp binaries
$ ./setup_binaries.sh
Build and install
$ ./gradlew assembleDebug $ adb install app/build/outputs/apk/debug/app-debug.apk
License
Free and open source. Forever.
Both NativeLab and PhonoLab are licensed under the GNU Affero General Public License v3. You can use, modify, and distribute them freely. If you run a modified version as a service, you must share the source code.
AGPL v3
Full copyleft license. Guarantees the software and all derivatives remain free and open source. See the LICENSE file.
Dependencies
llama.cpp (MIT) · AndroidX (Apache 2.0) · Material 3 (Apache 2.0) · KaTeX (MIT). No proprietary dependencies.