OpenDroid
1.0.9 · GitHub
More about this app
Open-source autonomous on-device AI agent that plans and executes multi-step tasks via screen automation
OpenDroid
๐ค The Open-Source Autonomous AI Agent for Android
Your phone. Your rules. Your AI.
Features โข Architecture โข Get Started โข Providers โข License
๐ฏ What is OpenDroid?
OpenDroid isn't just another chatbot. It's a fully autonomous AI agent that lives on your Android phone and actually does things for you.
"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."
OpenDroid will plan this as 3 steps, execute each one, verify the results, and adapt if anything fails โ all without you lifting a finger.
โจ Features
๐ง Autonomous Agent Engine
| Capability | Description |
|---|---|
| Self-Planning | Breaks complex commands into sequential steps with dependency tracking |
| Habit & Routine Detection | Mines recurring daily app sequences (e.g. Gmail \(\to\) Calendar \(\to\) Slack at 9 AM) and offers one-click automated routines |
| Re-Evaluation | Monitors execution results and dynamically replans when steps fail |
| Compound Intent Guard | Smart detection of multi-action commands (e.g. "open WhatsApp and send message") |
| Contact Disambiguation | 4-tier contact resolution with fuzzy matching and relationship aliases ("call dad") |
๐ ๏ธ On-Device Model Manager (LiteRT-LM)
| Capability | Description |
|---|---|
| Background Downloader | Real network downloads (via WorkManager) with Pause/Resume, cellular support, speed tracking, and ETA |
| Secure Authentication | Direct Android Keystore AES-GCM token storage to securely fetch gated Hugging Face models |
| Integrity Verification | Computes SHA-256 hashes and verifies LiteRT engine loading compatibility before marking READY |
| Local Model Import | Direct offline importing of catalog or freestanding custom .task / .litertlm files with JNI verification checks |
๐ฑ Full Device Control
| Action | Examples |
|---|---|
| System | Brightness, WiFi, Bluetooth, Flashlight, DND, Volume, Screenshot |
| Communication | WhatsApp messages, Telegram messages/channels (@user), Calls, SMS, Email drafts |
| Routines | Morning briefings, agenda summaries, automated macro sequences |
| Productivity | Screen extraction ("Read & Remember"), Alarms, Timers, Reminders, Calendar events, Notes |
| Navigation | Google Maps directions, Uber/Ola booking |
| Media | Play/pause music, YouTube search, camera |
| Finance | UPI payments, bill splitting, currency conversion |
| Smart Home | Google Home device control |
๐๏ธ Vision Engine & Screen Understanding
Captures screenshots via Accessibility API and feeds them to vision-capable LLMs for real-time screen analysis and "Read & Remember" multimodal meeting/note extraction. Falls back to accessibility tree text-scraping on older devices.
๐๏ธ 4-Tier Personal Knowledge Graph
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ Personal Knowledge Graph โ
โโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโค
โ โก Level 1 โ ๐ง Level 2 โ ๐ Level 3 โ ๐ Level 4 โ
โ Temporary โ Long-Term โ Learned โ Sensitive โ
โ (Active โ (Explicit โ Patterns โ (Keystore โ
โ Plan) โ Facts) โ (Inferences) โ Encrypted) โ
โโโโโโโโโโโโโโโโดโโโโโโโโโโโโโโโดโโโโโโโโโโโโโโโโโดโโโโโโโโโโโโโโโ
๐๏ธ Voice Interface
- Offline wake word detection โ say "OpenDroid" to activate
- Speech-to-text for hands-free commands
- Text-to-speech with ElevenLabs premium voice support
๐จ Premium UI
Built with Jetpack Compose featuring a futuristic glassmorphic design:
- Deep navy (
#080C10) + Neon green (#00FF88) color system - Pulsing audio orb animation during listening
- Live latency benchmarks for each provider
- Dark mode by default
๐ฌ Meet OpenDroid in 3D
OpenDroid saying hi โ rendered in 3D!
Note: If the video doesn't play inline on GitHub, click here to download and watch it.
๐ธ Screenshots
Chat โข Plan Engine โข Memory System โข Alarm Control
๐๏ธ Architecture
Clean architecture with Dagger-Hilt dependency injection:
com.opendroid.ai
โ
โโโ ๐ค accessibility/ App automators (WhatsApp, SMS, Calls)
โโโ โก actions/ 60+ action executors across 10 modules
โโโ ๐ง core/
โ โโโ agent/ AgentLoop, PlanManager, IntentClassifier, VisionEngine
โ โโโ llm/ 13 LLM providers, fallback chain, prompt engine
โ โโโ memory/ 4-tier memory system + notification intelligence
โ โโโ security/ Direct Keystore provider credentials + legacy plaintext migration
โ โโโ service/ Foreground service, notification listener, boot receiver
โ โโโ voice/ Wake word, speech recognition, TTS engine
โ
โโโ ๐พ data/
โ โโโ db/ Room database (7 DAOs, 7 entities, 3 migrations)
โ โโโ models/ Unified data models (Plan, Memory, ChatMessage)
โ โโโ repository/ Repositories backed by Room & DataStore
โ
โโโ ๐ di/ Hilt modules (App, Database, LLM)
โโโ ๐จ ui/
โโโ theme/ Glassmorphic design system
โโโ screens/ 16 screens (Chat, Plan, Memory, Settings, etc.)
โโโ viewmodel/ 8 ViewModels
โโโ components/ Reusable Compose components
๐ Supported LLM Providers
OpenDroid supports 13 LLM providers with automatic failover:
| Provider | Models | Type |
|---|---|---|
| ๐ข Google Gemini | Gemini 2.0 Flash, Pro, Nano | Cloud + On-device |
| ๐ฃ Anthropic Claude | Claude Sonnet 4, Opus 4 | Cloud |
| ๐ต OpenAI | GPT-4o, GPT-4.1, o3 | Cloud |
| โก Groq | LLaMA 3, Mixtral (ultra-fast) | Cloud |
| ๐ท DeepSeek | DeepSeek V3, R1 | Cloud |
| ๐ Mistral AI | Mistral Large, Medium | Cloud |
| ๐ OpenRouter | 200+ models via unified API | Cloud |
| ๐ค Together AI | Open-source model hosting | Cloud |
| ๐ด Cohere | Command R+ | Cloud |
| ๐ GitHub Copilot | GPT-4.1, Claude via Copilot API | Cloud |
| ๐ฉ Opencode-Go | Beam Free and other OpenCode Go models | Cloud |
| ๐ Ollama | Any local model (LLaMA, Phi, etc.) | Local |
| ๐ง Custom OpenAI | Any OpenAI-compatible endpoint | Self-hosted |
Smart Fallback: If your primary provider fails, OpenDroid automatically tries the next available provider in the chain.
โก๏ธ Getting Started
Prerequisites
- JDK 21 โ not newer. The project compiles against Java 21 (
jvmToolchain(21)).gradle/gradle-daemon-jvm.propertiesmakes./gradlewselect an installed JDK 21 automatically, so you do not have to changeJAVA_HOMEโ but a JDK 21 must be installed. - Android SDK 35 (Android 15)
Build & Install
# Clone the repository
git clone https://github.com/JMAN730/opendroid.git
cd opendroid
# Build debug APK
./gradlew assembleDebug
# APK output location
# โ app/build/outputs/apk/debug/app-debug.apk
Required Permissions
On first launch, OpenDroid will guide you through granting:
| Permission | Why |
|---|---|
| ๐ Accessibility Service | UI automation, screen reading, app control |
| โ๏ธ Write Settings | Toggle WiFi, Bluetooth, brightness |
| ๐ค Record Audio | Wake word detection & voice commands |
| ๐ Notification Access | Smart notification reading & auto-reply |
| ๐ฑ Post Notifications | Foreground service status |
Configure LLM
In Settings, add your API key for any supported provider. OpenDroid works best with:
- Gemini (free tier available)
- Groq (fastest inference)
- Ollama (fully offline)
๐ Security
Found a vulnerability? Please report it responsibly. See SECURITY.md for details.
โญ Star History
๐ License
Copyright 2026 OpenDroid Contributors
Last Updated: August 18, 2026
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
Made with โค๏ธ by JMAN730
โญ Star this repo if OpenDroid has helped you!
How Shizuku is used
Can run arbitrary shell commands and persistent terminal sessions via `sh` through Shizuku
This is an AI-assisted analysis of Shizuku-related usages in the app's public source code. It is best effort, so it may not catch every single usage.
How this app uses Shizuku
Shizuku is used to run shell commands with elevated shell privileges for the agent's local automation interface.
- Run arbitrary commands: A command string supplied through the local automation interface is executed with
sh -cthrough Shizuku and its output and exit code are returned. - Use persistent terminals: Long lived interactive shell sessions can be created through Shizuku, then any typed command can be sent to a session, its available output read back, and sessions listed or closed.
Android APIs or commands used
sh -csh
Notable details
Without Shizuku the same command and terminal features run in the app's normal shell with fewer privileges.
Changelog
What's new for version 1.0.9
Important
v1.0.9 is signed with a new release key (certificate SHA-256 45e7e11c6476b4f21387db1809c66c769878416f3f311e71b286f2d77db1e165). If you have an earlier version installed, Android will refuse the update, so the old build must be uninstalled first. Uninstalling deletes all local app data, including conversations, memories, routines and saved API keys, and app backups are disabled, so there is no automatic way to restore it. Install this build only once you accept losing that data.
Current release. Sideload the APK for direct install, or use the AAB for Play Store distribution.
v1.0.9 is a stability, privacy and safety release: 50 fixes across actions, auto-reply, storage, on-device models, social, memory and MCP. v1.0.8 was never published as a GitHub release, so everything in the v1.0.8 notes also ships for the first time in this build.
Highlights since v1.0.8
๐ Privacy & Security
- Backups off again:
android:allowBackup="false"is restored, and app files are excluded from device-to-device transfer (#241). - Level 4 secrets kept out of saved plans: sensitive plan parameters are kept out of the
planstable (#240), and Level 4 label nodes are cleared, with collector errors passed through the stream wrapper (#237). - Workspace confinement: file actions are confined to the workspace, refuse its root (#243), stay on the chosen custom SAF folder (#244), and refuse copy, move or zip when the source and destination overlap (#242).
- MCP server hardening: request headers are capped, and requests get a deadline and a bounded worker pool (#249). The MCP store no longer resets when decryption fails (#248).
๐ฒ Actions & Messaging
- PAY_UPI is no longer reported as sent, and its recipient and amount are validated (#206).
- WhatsApp: links carry the number's country code in E.164 (#208), the empty-input check uses WhatsApp's own hint, and fallbacks no longer re-send unconfirmed messages (#209). An open SMS composer counts as a hand-off, so the WhatsApp fallback can't send the message twice (#212).
- Calls & contacts: video calls go through the contact picker, saved picks are re-checked, and
#is kept in tel URIs (#207). - TYPE_TEXT no longer wipes a field it found by its own text (#210).
- Do Not Disturb turns on with the priority filter instead of total silence (#213).
๐ฌ Auto-Reply & Notifications
- Never replies through another conversation's notification (#214) and never replies into group conversations (#217).
- Pending replies are cancelled when you answer by hand (#218). Private context is kept out of the auto-reply prompt (#220).
AUTO_REPLY_TOGGLEapplies to every channel and the global switch (#223). The SMS toggle is limited to SMS apps (#215). Contact lookups are keyed on the app, not the bare name (#222).- A pending settings edit is saved when you leave the screen (#219).
- Notification history is pruned after 30 days and the history screen is capped (#216). Learned patterns stay in fixed rows and are cleared with the history (#225). The learned profile is kept out of
READ_NOTIFICATIONSresults (#224).
๐ง On-Device & Cloud Models
- On-device engine use is serialized under one mutex (#232). Decoding stops at the output cap (#231). The token estimate is more conservative, though still a heuristic (#230).
- The on-device planning prompt goes to the model unrewritten (#233). Only requests marked as planning are rewritten, not ones matched by prompt text (#221).
- RAM tiers are judged by what phones report, not advertised GB (#234).
- A stopped download keeps its partial file and releases the LiteRT engine (#236). Delete Unused Models keeps the selected on-device model (#235).
- Cloud provider responses are parsed without trusting their shape (#227). A huge
Retry-Afterno longer overflows into an immediate retry (#226). An unknown saved provider fails instead of routing to Gemini (#229). - Settings: a model pick is saved for the provider it was picked for (#238), and non-secret settings still save while credentials are unreadable (#228).
๐ Social
- YouTube publish and YouTube/LinkedIn replies fail visibly instead of faking success (#245).
- Publish and reply are idempotent across cancellation (#239).
- Send permissions are enforced and the right account is picked (#246). Refresh, re-sync triage, inbox replies and smaller bugs are fixed (#247).
- Room migrations
MIGRATION_9_10andMIGRATION_10_11: the database schema goes from version 9 to 11, addingsocial_comments.accountIdandsocial_interactions.replyTargetIdand storing unreported metrics as unknown instead of 0.
๐๏ธ Memory, Routines & UI
- Ordinary mentions of a spouse no longer overwrite the stored spouse name (#255). Contact picks are kept out of the prompt's user facts (#211). Startup cleanup no longer deletes user facts (#248).
- Each detected routine keeps one memory (#253). Paused routines stay on the Routines screen (#250). Routine sections come from one flow, with geocoding kept off the main thread (#254).
- The service stops instead of crashing when a foreground start is refused at boot (#252).
- Stale contact pickers are closed, and speech stops on cancel (#251).
Release Assets
app-debug.apkโ Debug build APK for developer testing & logging.app-release.apkโ Release APK (signed with V1+V2 signatures, ready to sideload).app-debug.aabโ Debug Android App Bundle.app-release.aabโ Release Android App Bundle.
Checksums (SHA-256)
app-debug.apk:1ddf7efb22ed44e21f5ba60a48584207d9025370a4478a2d038bfca5c55443bbapp-release.apk:cafe14e8f36c13fb99233f5fbd82bc6a4f6f0f5a0d977242acb9387b5b015615app-debug.aab:61a7452a4e545a14829b8605eeaffdf04275e0b1e1cbadab6e339e3b3a5b59a5app-release.aab:98ef49f2fd7039441837ec9269a1b71fa4519fe93503bbfa812b0b3966d6c98f
Build Configuration
- Package:
com.opendroid.aiagent - Version Code: 10
- Version Name: 1.0.9
- Min SDK: 26 (Android 8.0)
- Target SDK: 36 (Android 16)
Permissions
47 permissions requested
Sources
4 sources
Sources
4 sources