sleepy
8ab2e661ee
New Features: - RemoteAgent: Full server-as-agent mode (bypasses local model) - DelegationAgent: Local model decides when to ask server for help - Server mode selector in Settings: Local Only / Full Remote / Smart - Updated MainViewModel to support all three modes for text, audio, and images - SettingsUiState and SettingsViewModel updated with serverAgentMode Modes: - local_only: Use only local Gemma 4 model (default) - full_remote: All queries go to delegate server (OpenAI-compatible API) - delegation: Local model classifies queries, delegates complex ones to server
Sleepy Agent
A fully local AI assistant for Android powered by Google's Gemma 4 models via LiteRT-LM. Your conversations stay on your device - no cloud required. Can search the web when you need up-to-date information.
Features
🔒 Fully Local Inference
- Voice, image, and text processing all happens on-device
- No internet connection required for inference (except for web search tool)
- Conversations stay private - no data sent to external AI services
🎙️ Voice Input
- Tap the mic button and speak naturally
- Voice Activity Detection (VAD) automatically stops recording after you finish speaking
- Optional TTS (Text-to-Speech) responses when using voice input
🖼️ Image Understanding
- Send images from your gallery or take a photo
- Ask questions about what's in the image
- Works with text prompts alongside images
📝 Text Chat
- Full markdown support including tables and code blocks
- Persistent conversation history
- Navigate between multiple chat sessions
🧠 Gemma 4 via LiteRT-LM
- Powered by Google's official LiteRT-LM SDK
- Choose between E2B (2B params, ~2.7GB, faster) or E4B (4B params, ~4.5GB, higher quality)
- 16K token context window
- KV cache reuse for faster multi-turn conversations
- Performance: E2B runs at ~25-30 tokens/sec on a Samsung Galaxy Z Fold 5 (personal testing)
📥 Easy Model Setup
- Download directly in the app: Settings → Download Gemma 4 E2B/E4B
- Or select your own model: Use any
.litertlmfile from HuggingFace LiteRT Community - Device info card shows your RAM to help choose the right model
💾 Session Management
- Navigation drawer shows all your past conversations
- Continue previous chats or start fresh
- Auto-saved conversation history
🔊 Smart TTS
- Optional text-to-speech for responses
- Auto-detect mode: speaks when you use voice input, silent for text input
Work in Progress
- Floating Bubble: Quick access overlay (requires additional permissions)
- Home Server Delegation: Optionally route requests to your own server
Requirements
- Android 8.0+ (API 26)
- 4GB+ RAM recommended
- ~3GB free storage for E2B model (~5GB for E4B)
Building
See DEVELOPMENT.md for detailed build instructions and how to configure your own SearXNG server.
Quick build:
./gradlew :app:assembleDebug
The release APK is ~50MB (arm64-v8a only).
Web Search Setup
The app can search the web using a SearXNG server. To set up your own, see DEVELOPMENT.md.
Model Sources
Download .litertlm models from:
- HuggingFace LiteRT Community
- Gemma 4 E2B:
gemma-4-E2B-it-litert-lm - Gemma 4 E4B:
gemma-4-E4B-it-litert-lm
License
MIT License - See LICENSE file