TouchPilot
Android-native AI agent runtime
for safe, observable phone control.
TouchPilot lets AI agents perceive and operate Android devices through explicit, permissioned, and inspectable tools.
See
Observe UI semantics via AccessibilityService
Act
Tap, type, scroll, open apps, and more
Safe
Permissioned tools with risk levels
Observable
Every decision and action is logged
Chat
OpenAI-compatible models today, local models later
Built for developers
# Build debug APK
$ ./gradlew assembleDebug
# Build release APK
$ ./gradlew assembleRelease- Runs natively on Android
- OpenAI-compatible providers
- Local inference (ExecuTorch, llama.cpp) coming
- Full audit logs and approvals
How it works
From a downloaded APK to a working on-device agent in under five minutes.
Every step is reversible. No root, no Play Store, no account.
Sideload the APK
Download TouchPilot from GitHub Releases. No Play Store needed. Works on real phones and emulators.
Grant Accessibility
Standard Accessibility service — how the agent reads your screen and performs taps. No root, no overlay hacks.
Chat your task
Open Wi-Fi settings, draft a text, change volume. The agent proposes a typed tool call.
Approve what matters
LOW tools run silent. MEDIUM asks once per session. HIGH always asks. You see every argument.
Chat with the agent
Inspect every tool
Read every action
Modular
Pluggable tools, skills, and providers
Secure by design
Approvals, risk levels, and policy controls
Local first
Logs, skills, and sessions stored on-device
Early stage
MVP: AccessibilityService spike and agent runtime
Try TouchPilot on your phone.
Sideload the APK, grant Accessibility, and start a chat. You'll see every tool the agent runs and approve the ones that matter.