🧠 AI-native Android automation via MCP. Give Claude, GPT, Gemini or any AI agent full control over Android — tap, swipe, type, screenshot — with ~6ms latency. 100x faster than Appium. 43 tools.
概要
AI-native Android automation via the Model Context Protocol Give any AI agent — Claude, GPT, Gemini, or your own — full control over an Android device. Observe the UI. Tap. Swipe. Type. Manage apps. Capture screenshots. All with sub-100ms latency . No root required. Existing Android automation falls into two camps — and both share the same bottleneck: 1. (Appium, Maestro) — built for QA, not AI agents. Every action routes through ADB or UIAutomator IPC, costing 200-1000ms per operation. 2. (mobile-mcp, droidrun) — AI-aware, but thin ADB wrappers underneath. 200-2000ms per action. NeuralBridge takes a different approach: an on-device AccessibilityService that executes actions , eliminating IPC overhead entirely. — NeuralBridge averages per action vs for Appium and for mobile-mcp. See full benchmarks → Your AI agent speaks MCP over HTTP directly to the companion app — no middleware, no ADB, no intermediate server.
README
🧠 NeuralBridge
AI-native Android automation via the Model Context Protocol
Give any AI agent — Claude, GPT, Gemini, or your own — full control over an Android device. Observe the UI. Tap. Swipe. Type. Manage apps. Capture screenshots. All with sub-100ms latency. No root required.
🤔 Why NeuralBridge?
Existing Android automation falls into two camps — and both share the same bottleneck:
- Test frameworks (Appium, Maestro) — built for QA, not AI agents. Every action routes through ADB or UIAutomator IPC, costing 200-1000ms per operation.
- MCP wrappers (mobile-mcp, droidrun) — AI-aware, but thin ADB wrappers underneath. 200-2000ms per action.
NeuralBridge takes a different approach: an on-device AccessibilityService that executes actions in-process, eliminating IPC overhead entirely.
| NeuralBridge | Appium | Maestro | mobile-mcp | droidrun | ADB Shell | |
|---|---|---|---|---|---|---|
| ⚡ Tap latency | ~2ms | 300-1000ms | 750ms-2s | 500-2000ms | 200-1000ms | 300-1000ms |
| ⌨️ Text input | ~1.4ms | 500-3000ms | 750ms-2s | 500-2000ms | 200-1000ms | 500-2000ms |
| 🌳 UI tree read | 18-33ms | 500-2000ms | 750ms-2s | 1-5s | 200-500ms | 1-5s |
| 📸 Screenshot | ~60ms | 300-500ms | ~1s | 300-500ms | ~250ms | 300-500ms |
| 🔌 MCP native | Yes | Add-on | Yes (stdio) | Yes | Via adapter | No |
| 🛠️ MCP tools | 43 | 30+ (add-on) | 14+ | ~19 | ~11 | — |
| 🎯 Token optimization | Yes (73%) | No | No | No | No | No |
[!TIP] 100x faster on average — NeuralBridge averages 6.4ms per action vs ~800ms for Appium and ~1.5s for mobile-mcp. See full benchmarks →
🏗️ How It Works
Your AI agent speaks MCP over HTTP directly to the companion app — no middleware, no ADB, no intermediate server.
[!NOTE] 1 hop, not 5. Appium routes through HTTP → Appium Server → ADB → UIAutomator2 → Device. NeuralBridge goes Agent → HTTP → Companion App. The AccessibilityService runs in-process — like the difference between a phone call and a note on your own desk. Deep dive →
🚀 Quickstart
Prerequisites: Android SDK (API 24+), Java JDK 17, Android device or emulator (7.0+)
1️⃣ Clone & Build
git clone https://github.com/dondetir/NeuralBridge_mcp.git
cd NeuralBridge_mcp/android
./gradlew assembleDebug
2️⃣ Install & Enable
# Install the companion app
adb install -r app/build/outputs/apk/debug/app-debug.apk
# Enable AccessibilityService
adb shell settings put secure enabled_accessibility_services \
com.neuralbridge.companion/.service.NeuralBridgeAccessibilityService
adb shell settings put secure accessibility_enabled 1
[!IMPORTANT] Android 15+ requires an extra step: Settings → Apps → NeuralBridge → Enable “Allow restricted settings”
3️⃣ Connect Your AI Agent
The app shows its IP on the main screen. Add it to your agent’s MCP config:
# Claude Code
claude mcp add neuralbridge http://:7474/mcp --transport http
4️⃣ Verify
Ask your AI agent: “Take a screenshot of the Android device” — if you see a screenshot, you’re connected! 🎉
📚 Documentation
| Document | What’s inside | |
|---|---|---|
| 🏗️ | Architecture | Data flow, command paths, selector system, limitations |
| 🛠️ | Tools Reference | All 32 MCP tools with descriptions and latencies |
| ⚡ | Performance | Benchmarks, architecture comparison, token optimization |
| 🔧 | Troubleshooting | Connection issues, permissions, screenshots, crashes |
| 🔨 | Development | Building from source, project structure, protobuf, logs |
| 📱 | Android App | App-specific setup and permissions |
| 🤝 | Contributing | Code style, PR process, guidelines |
| 🔒 | Security | Vulnerability reporting policy |
🗺️ Roadmap
- [x]
Core MVP — 16 tools, TCP protocol, basic gestures - [x]
Advanced gestures, selectors, event streaming, notifications - [x]
Semantic resolver, scroll-to-element, accessibility audit, screenshot diff - [ ] Multi-device, WebView tools, CI/CD integration, visual regression
🤝 Contributing
See CONTRIBUTING.md for guidelines. High-priority areas: additional MCP tools, performance optimizations, cross-platform support (iOS), and example demos.
📄 License
Apache 2.0 — Copyright 2026 dondetir
“Powered by NeuralBridge”
インストール
This server does not publish a one-line install command.
Open the repository installation guide設定
{
"mcpServers": {
"neuralbridge": {
"type": "http",
"url": "http://<device-ip>:7474/mcp"
}
}
}