Zetla

AI-powered Android app with on-device Python execution, voice chat, web search, and tool calling.

Features

• LLM Chat - Multi-provider support (OpenCode Zen, DeepSeek, NVIDIA NIM) with SSE streaming
• On-device Python - Static CPython 3.14.6 binary with 93 modules, runs user code via run_code tool
• Voice Chat - Vosk offline speech recognition + Android TTS with waveform visualization
• Web Search - EXA AI MCP integration for real-time web search
• Tool Calling - Agent loop with max 20 iterations, reasoning streamed to UI
• File Support - Text, CSV, PDF, DOCX, PPTX, XLSX file parsing with preview chips
• C++ Native Library - Session management, encryption, SSE parsing, auto-compaction

Architecture

Zetla/                  # Android project (Kotlin/Compose)
  app/                  # Application module, APK entry point
  domain/               # Models, repository interfaces
  data/                 # Repository implementations, JNI bridge, C++ CMake build
  ui/                   # Compose UI, ViewModels, voice chat
src/zetla/              # C++ source (DLL/shared library)
  api/                  # DLL API + JNI bridge
  session/              # Session manager with auto-compaction
  search/               # Web search (EXA MCP)
  storage/              # Local encrypted storage
  providers/            # LLM provider adapters
  file_handlers/        # File parsing (text, PDF, office)
  tools/                # Tool registry

Prerequisites

• Android Studio with SDK 36 and NDK r28b
• MinGW-w64 (for building libcurl/OpenSSL native deps)
• WSL Ubuntu (for building Python static binary)
• Java 17+

Quick Start

# Full setup (builds deps, downloads assets, builds APK)
.\scripts\setup.ps1

# Or step by step:
.\scripts\setup.ps1 -SkipDeps      # Skip native deps if already built
.\scripts\setup.ps1 -BuildApkOnly  # Just build the APK

Build Steps

1. Native Dependencies (OpenSSL + libcurl)

.\build_deps.bat

Builds static OpenSSL and libcurl for arm64-v8a and armeabi-v7a.

2. Python Runtime (WSL)

The Python binary is a static build of CPython 3.14.6 for Android, built using the reproducible build system at ~/zetla_android_build/. See docs/PYTHON_ANDROID_BUILD.md for complete build documentation.

# In WSL Ubuntu - full build (deps + cpython + package)
wsl -d Ubuntu-26.04 bash -c 'cd ~/zetla_android_build && ./build_all.sh --clean'

# Or step by step:
wsl -d Ubuntu-26.04 bash -c 'cd ~/zetla_android_build && ./build_all.sh --step deps'
wsl -d Ubuntu-26.04 bash -c 'cd ~/zetla_android_build && ./build_all.sh --step cpython'
wsl -d Ubuntu-26.04 bash -c 'cd ~/zetla_android_build && ./build_all.sh --step package'

Output in ~/zetla_android_build/dist/bin/:

File	Size	Description
python	~19MB	Static Python 3.14.6t binary
python3.14t.zip	~27MB	Stdlib (zipimport)
cacert.pem	~178KB	SSL CA certificates

The build system (scripts/setup.ps1) picks up these files from WSL and copies them to:

• Zetla/app/src/main/jniLibs/arm64-v8a/libpython.so
• Zetla/app/src/main/assets/python314t.zip
• Zetla/app/src/main/assets/cacert.pem
• Zetla/data/src/main/assets/python314t.zip
• Zetla/data/src/main/assets/cacert.pem

3. Vosk Speech Model

Downloaded automatically by scripts/setup.ps1, or manually:

# Download from https://alphacephei.com/vosk/models
# Extract to Zetla/app/src/main/assets/model-en-us/

4. Build APK

cd Zetla
.\gradlew.bat assembleRelease

Output: Zetla/app/build/outputs/apk/release/app-release.apk

API Keys

Set in Zetla/ui/src/main/java/com/zetla/app/ZetlaApp.kt or via the app's settings screen.

• OPENCODE_API_KEY - OpenCode Zen
• DEEPSEEK_API_KEY - DeepSeek
• NVIDIA_API_KEY - NVIDIA NIM

TODO

• Model capabilities not shown correctly — vision, reasoning, and other capability flags from C++ providers are not reliably reflected in the Kotlin ModelCapabilities data class and ModelSelectionSheet UI.
• Thinking level selection does not work — thinking_levels are returned by providers but the reasoning effort chip selection in the UI does not apply correctly to API requests.
• Local RAG Corpus for multiple large document file support with tool definition for LLM to utilize the same.

License

MIT - Copyright 2026 Aradhya Chakrabarti