JMAN730/opendroid

What it solves

Most AI assistants on phones are just chatbots—they talk but don't act. OpenDroid is a fully autonomous agent that actually does things on your Android device. It takes a complex request like "Check if it's going to rain tomorrow, and if so, text my wife that I'll be late and set an alarm for 6 PM," breaks it into a step-by-step plan, executes each step, verifies the result, and re-plans if anything fails—all without you touching the screen.

How it works

OpenDroid runs on your phone and uses several layers of capability:

• Agent engine: It self-plans by breaking commands into sequential steps with dependency tracking, executes them via action executors, monitors results, and dynamically replans when steps fail. It also detects recurring app patterns (like checking Gmail, then Calendar, then Slack at 9 AM) and offers one-click routines.
• Full device control: It connects to Android's Accessibility Service and other system APIs to perform actions across system settings (brightness, WiFi, Bluetooth, volume), communication (WhatsApp, Telegram, SMS, calls, email), productivity (alarms, timers, calendar events, notes), navigation (Google Maps, Uber), media, finance (UPI payments, currency conversion), and smart home.
• Vision and screen understanding: It captures screenshots and sends them to vision-capable LLMs for real-time screen analysis, with a fallback to accessibility-tree text extraction on older devices.
• Four-tier memory: A personal knowledge graph stores temporary plan context, long-term explicit facts, learned patterns, and sensitive data (keystore-encrypted).
• Voice interface: Offline wake word ("OpenDroid"), speech-to-text for hands-free commands, and text-to-speech with premium voice support.
• Model flexibility: Works with 13 LLM providers (Gemini, Claude, GPT, Groq, Ollama, and more) and automatically falls back to the next provider if one fails. On-device model manager (LiteRT-LM) supports downloading and importing local models.

Who it's for

OpenDroid is for Android users who want a genuinely autonomous AI assistant—not just a chat bot—that can manage their device, automate repetitive tasks, and understand complex multi-action commands. It's also suited for developers and privacy-conscious users who want to run models locally via Ollama or import their own .task or .litertlm files.

Highlights

• Fully autonomous planning, execution, verification, and re-planning of multi-step tasks
• 60+ action executors across 10 modules, covering system, communication, productivity, navigation, media, finance, and smart home
• Supports 13 LLM providers with automatic fallback, including cloud and fully offline local models
• Built-in vision engine for screen understanding and multimodal "Read & Remember" extraction
• Four-tier personal knowledge graph for persistent, encrypted memory
• Voice interface with offline wake-word detection, speech-to-text, and TTS
• On-device model manager with pause/resume downloads, secure token storage, and integrity checking
• Premium Jetpack Compose UI with glassmorphic dark design and live latency benchmarks