A monochrome on-device generative AI playground for Android.
HUSK is a fork of Google AI Edge Gallery rebranded with a dark monochrome theme and an expanded feature set. It runs open-source models directly on your device — fully offline, private, and without cloud dependencies — using Google AI Edge and LiteRT under the hood.
AI Chat with Thinking Mode — Engage in multi-turn conversations with supported models including the Gemma family. Toggle Thinking Mode to expose the model's step-by-step reasoning process.
Agent Skills — Augment your LLM with tools like Wikipedia for fact-grounding, interactive maps, and visual summary cards. Load modular skills from a URL or browse community contributions.
Ask Image — Use multimodal inference to identify objects, solve visual problems, or get detailed descriptions using your camera or photo gallery.
Audio Scribe — Full-featured on-device audio transcription powered by whisper.cpp. Supports universal audio formats (M4A/AAC, MP3, OGG/Vorbis, FLAC, WAV, AMR, OPUS) and video audio extraction (MP4, MKV, 3GP, WebM, MOV). Features user-selectable Whisper model size (Tiny ~75MB, Base ~150MB, Small ~500MB), no duration limit, adaptive ETA display, and transcription history. Speaker diarization uses ECAPA-TDNN embeddings with persistent voice profiles that improve through manual labeling. Recordings under 10 minutes are auto-summarized via Gemma E4B.
Notes and Brainstorm — An AI-driven brainstorming companion on the home screen. Notes persist in a Room database with auto-tagging and full-text search (FTS4).
Knowledge Base (RAG) — On-device document retrieval using the Google AI Edge Local Agents SDK. Documents are indexed with Gecko embeddings (768-dimensional) stored in a SqliteVectorStore, grounding model responses in your own content.
Prompt Lab — A focused workspace for testing prompts and single-turn use cases with granular control over parameters like temperature and top-k.
Mobile Actions — Offline device controls and automated tasks powered by a fine-tuned FunctionGemma 270m.
Tiny Garden — An experimental mini-game using natural language to plant and harvest a virtual garden, also powered by FunctionGemma 270m.
MCP Integration — Model Context Protocol client for connecting to external tool servers, bridging MCP tools into the agent chat system.
Context Memory — Hot and warm memory layers that give the AI context awareness across conversations.
Model Management and Benchmark — Download models from the built-in list or load custom models. Run benchmark tests to compare performance on your specific hardware.
All inference runs on device hardware. No internet connection is required for model inference, and your prompts, audio, images, and documents never leave the device.
A live thermal meter in the app's top bar shows device skin temperature with a color-coded severity indicator (green / yellow / orange / red). Tapping it opens a detailed popup with all sensor readings, a thermal headroom bar, battery temperature, and recommended temperature ranges for AI inference. Uses Android PowerManager.getThermalHeadroom() and BatteryManager APIs.
See docs/ARCHITECTURE.md for detailed system architecture.
- Google AI Edge — Core APIs and tools for on-device ML
- LiteRT — Lightweight runtime for optimized model execution
- whisper.cpp — Native C++ speech recognition via NDK/JNI
- ECAPA-TDNN — Speaker embedding model for diarization
- Google AI Edge Local Agents — Retrieval-augmented generation pipeline
- Room Database — Persistent local storage for notes and transcripts
- Proto DataStore — Typed settings and user preferences
- Hugging Face — Model discovery and download
- OS requirement: Android 12 or higher.
- Build from source: See DEVELOPMENT.md for local build instructions.
- Found a bug? Report it here.
- Have an idea? Suggest a feature.
Licensed under the Apache License, Version 2.0. See the LICENSE file for details.