デバイス操作はAIエージェントの時代へ。 mobile-mcpを活用したAndroid UI/E2Eテストの挑戦 DroidKaigi 2026 AI Mori Atsushi LY Corporation 2026/09/02 

Mori Atsushi Twitter: @at_sushi_at LY Corporation / LINEヤフー株式会社 Android App Engineer MANUL — Write Music with a nger Author of A Guide to Good Code / 著書: 良いコードの道しるべ fi 2019 MITOU Super Creator / 2019年度 未踏スーパークリエータ 

Get a feel for what AI agents can do through 🚀 device interaction — and where their limits are. . Demo . Use cases that work well — and those that don’t . Setup & Alternative Tools . Practical tips for e ective use ff 4 3 2 1 Today’s Goal 

1 DEMO 

MANUL - Music Score Editor Prompt: Launch the MANUL app, create a new score, and enter the rst two measures of “Kaeru no Uta” プロンプト: MANULアプリを起動し、新しい楽譜を作成し、カエルの歌 fi を最初2小節作成して。 

How to Add a Note: 1. Touch anywhere on the score 2. Slide to set the pitch and position 3. Release to place the note 

2 Use cases that work well — and those that don’t 

Case : Quick Testing After Implementation Check newly implemented features, feature changes, and bug xes Run lightweight regression tests Create an AI-driven loop: Device Test Fixes fi 1 Implementation 

Case : Bug Investigation Reproduce the issue by following the reported steps Add logs or toggle implementation ags while letting the agent interact with the app Confirm Fixes fl 2 Bug Investigation 

Use Files for Temporary Debug Switches 

Case : Store Asset Creation Capture Screenshots 3 Generate Store Assets with Scripts 

Not a good t for: AI is slower, more expensive, and less reproducible. Alternative: Human validation + E E tests Alternative: Unit Test / E E tests 📐Layout validation 🖐 Subjective UX validation May miss positioning, spacing, alignment, Hard to judge whether interactions feel or guideline issues. smooth or intuitive. Alternative: Screenshot tests + human review Alternative: Human review fi May skip steps or misinterpret results. 2 🔁 Repetitive regression testing 2 🚦Final release validation 

3 Setup & Alternative Tools 

mobile-next/mobile-mcp An MCP server for controlling mobile devices from AI agents. It can inspect on-screen elements and perform actions such as taps, swipes, and text input. Useful for letting tools like Claude Code or Codex interact with Android and iOS apps. 

Claude Code Codex 

Android CLI A command-line tool for inspecting Android devices and app UI state. Can expose properties such as selected / toggle state, beyond basic labels and coordinates. Also supports layout di s to e ciently track UI state changes. ffi ff fi "The Android robot is reproduced or modi ed from work created and shared by Google and used according to terms described in the Creative Commons 3.0 Attribution License." 

sim-use A cross-platform CLI for controlling iOS Simulators and Android devices/emulators. Provides a token-e cient UI outline and lets agents interact with elements by alias instead of coordinates. ffi https://github.com/lycorp-jp/sim-use 

mobile-mcp 

Android CLI 

sim-use 

Android CLI sim-use MCP tools CLI CLI Platforms Android / iOS Android only Android / iOS UI inspection Accessibility-based elements UI hierarchy as JSON Compact UI outline Interaction Tap, swipe, type, etc. Mainly inspection / device management Tap, swipe, type, gestures, etc. Key strength Easy integration with AI agents via MCP Android-speci c inspection, including layout di Token-e cient, agentfriendly UI representation f Interface fi ffi mobile-mcp 

ff 4 Practical tips for e ective use 

Prefer the View Tree Over Screenshots With mobile-mcp, use mobile_list_elements_on_screen to understand the current UI structure. Get the view tree Act Confirm the result 

Save Screenshots as Evidence review later whether the behavior was veri ed correctly. Get the view tree Act fi ✅ Save screenshots as evidence so humans can Confirm the result Save screenshots 

Create a reference le for each screen Include details such as how to open the screen and fi how to interact with it. 

Example: Detailed instructions mobile_swipe_on_screen in mobile-mcp does not support ne-grained swipe gestures. fi Use shell input motionevent instead, and document how to perform these interactions. 

Screen Map Create a mapping between each screen and its reference le. It is also useful to link each screen to relevant code, such as its Activity, navigation key, or Composable function. fi Keep one screen per line so the map is easy to search. 

How to Maintain Reference Files Update After Device Interaction 📝 Example prompt: Update the reference les based on what you learned during this task. Create a Skill for updating reference les fi fi fi Have a subagent review reference les 

. Run an AI agent periodically to check older les. . Verify that the documentation still matches the current implementation. fi ✅ . Add a last reviewed date to each reference le. fi 1 2 3 Review Regularly 

Friendly to Both Humans and AI Add content descriptions wherever possible. They are especially important for icon buttons and other controls without visible text. 

Expose Test Tags as Resource IDs Compose testTags can be exposed as Android resource IDs. For development tools, enabling this only in debug builds is a reasonable default. 

Make the Current Screen Observable Option : Get the Current Activity 2 1 Option : Get the Screen from Logs 

Make App Settings Controllable via ADB fl Example: Toggle feature ags 

Add a debug-only BroadcastReceiver 

Wrap ADB Commands with a Shell Script 

🎉 Enjoy exploring your app with AI agents! Design the app so that an AI agent can observe the state, understand the UI, and perform actions reliably.