MobileClaw-Heretic
MobileClaw
Your phone is the agent.
An open-source AI that controls your Android phone with natural language.
Quick Start • On-Device AI • Tools • Website
MobileClaw is the Android port of OpenClaw. It turns your phone into an AI agent — tell it what to do in your own language and it taps, types, reads, calls, and navigates for you. Runs with cloud AI (Claude, GPT, Gemini, 10 providers) or entirely offline with Gemma 4 on-device.
No server. No subscription. Your phone does everything.
Why MobileClaw?
| MobileClaw | Other AI Assistants | |
|---|---|---|
| Runs on your phone | Everything local. No backend. | Requires cloud servers |
| Controls any app | AccessibilityService taps, swipes, types in ANY app | Limited to their own UI |
| On-device AI | Gemma 4 with GPU — no internet needed | Always needs internet |
| Open source | Apache 2.0. Fork it, modify it, own it | Closed source |
| Multi-provider | 10 cloud + on-device. Bring your own keys | Locked to one vendor |
| Extensible | 21 skills as Markdown. Create your own in-app | Fixed capabilities |
On-Device AI
Run Gemma 4 directly on your phone via LiteRT-LM. No API key. No internet. No cost.
- E2B (2.6 GB) — 6 GB+ RAM, ~9 tok/s on GPU
- E4B (3.7 GB) — 8 GB+ RAM, higher quality responses
- GPU accelerated with NPU/CPU fallback (auto-detected)
- Tool calling works on-device via constrained decoding
- Download models from Settings, switch between cloud and local anytime
Tools
29 native tools that directly access Android APIs:
| SMS | Call Log | Contacts | Phone Call |
| Calendar | Alarms | Notifications | App Launcher |
| Navigation | UI Automation | Screen Capture | Web Browser |
| HTTP API | File System | Photos | Clipboard |
| Media Control | Volume | Brightness | Flashlight |
| System Info | Scheduled Tasks | Skill Author | Memory |
| Session History | Sub-Agent | Channel | Telegram |
| OpenAI |
Every tool works with both cloud and on-device models.
Skills
21 composable skills (Markdown + YAML frontmatter):
Built-in: morning-routine, email, messaging-apps, navigation, notification-digest, phone-basics, photos, self-learning, social-media, telegram-bot, translation, ui-fallback, weather, web-research
Vertical: finance, health, notion, real-estate, shopping, smart-home, telegram
Create your own: describe what you want, and the AI writes the skill for you (/create command).
Quick Start
git clone https://github.com/ChenKuanSun/mobileClaw.git
cd MobileClaw
./gradlew assembleDebug
Install the APK, then either:
- Cloud: Enter an API key (Anthropic, OpenAI, Google, etc.)
- On-Device: Go to Settings > AI Provider > On-Device > Download Gemma 4
Grant permissions as needed, enable Accessibility Service, and start chatting.
Architecture
┌──────────────────────────────────────────┐
│ Jetpack Compose UI │
│ ChatScreen · Skills · Settings │
├──────────────────────────────────────────┤
│ AgentRuntime │
│ Tool-use loop with streaming │
├────────────────┬─────────────────────────┤
│ 29 Android │ LiteRT-LM Engine │
│ Tools (native)│ (Gemma 4 on-device) │
├────────────────┴─────────────────────────┤
│ ClaudeApiClient · 10 Cloud Providers │
├──────────────────────────────────────────┤
│ Room · DataStore · EncryptedPrefs │
├──────────────────────────────────────────┤
│ AccessibilityService · Notifications │
└──────────────────────────────────────────┘
Stack: Kotlin 2.2 · Jetpack Compose · Hilt · Room · LiteRT-LM 0.10 · Anthropic SDK · Ktor
Contributing
See CONTRIBUTING.md. PRs welcome.
License
MIT — see LICENSE.