kuafuai/andee
AI EngineeringAn on-device agent for Android and Mac that operates any app through accessibility, operating apps without integrations or root.
What it is
Andee is a personal agent that lives on an Android phone/tablet or a Mac and controls other apps the way a user would. On Android it reads the accessibility tree or a screenshot, taps, swipes, types and reads notifications; on the Mac it adds accessibility access, real mouse/keyboard events, screen capture and a sandboxed terminal. The shared agent loop, prompt, notebook format, scenes and safety rules live in a common core/ module, and each platform implements a BrainHost interface. The project targets people who want an assistant that works across arbitrary apps without SDKs, integrations or root.
The model backend is any OpenAI-compatible endpoint; device memory and the password vault are described as local-only. It also supports pairing two devices over a QR code so one can delegate work to the other. For first-time users, the README recommends building an installable package through CodeFlying, a separate commercial service that preconfigures model and voice keys, with a free allowance followed by paid tokens; alternatively you can use release packages and supply your own keys, or build from source.
Quick start
Fastest route for Android, per the install section:
- Open the CodeFlying build page, build, and scan the QR code to download the package (build online).
- Install the APK.
- Grant the Overlay permission, then enable the Andee Accessibility service. The floating ball appears.
- Tap the ball and allow the microphone prompt, then talk.
If you prefer a release package from GitHub, download andee-<version>-en-android-arm64-v8a.apk, install it, and put your own CodeFlying key or an OpenAI-compatible model key under ⚙ → Backend, or switch to On-device mode (latest release).
To build from source on Android you need JDK 17–20, Gradle 8.4, Android SDK Platform 34, Build-Tools 34 and local.properties with sdk.dir=; then ./gradlew assembleDebug and adb install -r app/build/outputs/apk/debug/app-arm64-v8a-debug.apk (build steps). For the Mac desktop app, run ./gradlew :desktop:createDistributable and copy the resulting Andee.app into Applications; macOS will ask for Accessibility and Screen Recording.
Permissions and platform requirements
Android 8.1 (API 27) is enough to install and chat, but screen reading, visual tap targets and screenshots require Android 11 (API 30) or newer. The Mac build requires macOS 14+ and is architecture-specific: an Apple silicon build will not start on an Intel Mac, and vice versa.
The permission set is unusual and the README says so: overlay, accessibility, notification access, SMS, call log, contacts and installed-app list are requested per call and degrade gracefully when refused, but some capabilities are all-or-nothing. Screen actions run entirely through the AccessibilityService, and on a non-rooted device that is the only route to real touch events. ASR and TTS go over the public internet; the wake word is fully offline.
Known limitations
- The robot-dog dongle firmware and wiring diagram are not in the repository, and if the dongle loses power mid-motion the dog finishes its last action because it has no lost-contact stop.
- The debug WebSocket server binds
0.0.0.0:9008without client authentication on the current network (SECURITY.md). - Building the Mac app loses the Accessibility and Screen Recording grants on every unsigned rebuild; the repo documents fixing a local signing identity (signing notes).
- The project is pre-1.0 (
v0.15.0), and the license is Apache-2.0 with additional conditions: no multi-tenant SaaS, no white-labelling and no embedding in another product without written permission (license). Bundled ML Kit andplay-services-*components are not open source.
HotRepo's take
There is no recorded daily-rank, snapshot or discussion record for this repository in the available HotRepo data, so the 6,150 stars and 291 forks can only be read as a current total, not as a trend. What is distinctive in the material is the engineering posture: a non-Compose Android UI built on WindowManager and AccessibilityService, a shared core/ with both bodies, and guard rails placed in the dispatcher rather than the prompt. That is also why the repository needs careful handling: an agent that asks for accessibility, SMS and contacts permissions is functionally indistinguishable from malware at the manifest level, and the project itself points this out. Readers should treat the installable APK and DMG as the main trust decision, and the published source and permission documentation as what they can actually check.
Further reading
The README documents the call architecture, notification triage, scenes, on-device notebook and friend pairing in detail (project README). A Chinese permissions page is also available (permissions.zh-CN).
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- phone-harness/mcpA hosted MCP server that lets AI agents control real iPhones and Android emulators in the cloud.
- kilolonion/agent-vscreenRun Android apps on an invisible virtual display so phone AI agents can operate them while you keep using your own screen.
- droidrun/mobilerunAn LLM-agnostic framework that lets AI agents inspect, tap, swipe, and type on real Android and iOS devices via CLI or Python.
- KnifeLemon/DroidlineAutomate and test real Android phones from Python, Node.js, C#, CLI, HTTP or Appium without ADB, USB or root.
- X-PLUG/MobileAgentA family of GUI agents and GUI-Owl models that automate desktop, mobile, and web apps via screenshots and instructions.