An AI agent that prefers system calls over screenshots
Eta turns Xiaomi and OPPO phones into hosts for a system-level AI agent that controls devices through structured APIs, GUI automation, and root shells instead of blindly tapping buttons.

What it does
Eta is a system-level AI agent runtime for Android 14+ devices running ColorOS or HyperOS. It uses system assistants like Breeno and Xiao AI as entry points, but carries its own model configuration and execution loop: the model issues commands, the system executes them, and results feed back into context until the task is done. The runtime offers three ways to act—structured device tools that call Android system interfaces directly, a GUI automation layer that falls back to accessibility events and screenshots when no API exists, and a terminal that runs shell commands in either the native Android environment or an optional Alpine Linux chroot.
The interesting bit
While most phone agents treat the OS as a black box they can only poke through the screen, Eta exposes schema-validated device tools—set_alarm, set_volume, send_message, device_status, and more—so the model can set an alarm or send a WeChat message without ever opening the Clock or WeChat UI. The GUI layer is still there for cross-app scraping and comparison shopping, but it is the fallback, not the default. When APIs cannot reach deep enough, the agent can drop into a root shell or a Linux tool environment to read logs, tweak configs, or run Python scripts.
Key highlights
- Structured device tools with parameter schemas and permission groups let models call system functions directly instead of simulating taps.
- A built-in browser agent loads JavaScript pages in the background, extracts structured content, and can hand off to a user-visible WebView when a captcha appears.
- Terminal access spans both native Android shell (with auto-discovered BusyBox) and an optional Alpine chroot for Bash, Python, Git, and SQLite.
- Skills system supports installing extensions from public GitHub repos or local ZIPs, with conflict protection and lazy loading.
- Sensitive data—SMS verification codes, Wi-Fi passwords, notification text, and message parameters—are kept in-memory for the current turn only and never written to persistent session storage.
Caveats
- Full functionality demands root access and an LSPosed framework installation; without them, the agent loses its deepest system hooks.
- The project is tightly bound to ColorOS and HyperOS, so owners of Samsung or Pixel phones are out of luck.
- WeChat messaging and other sensitive operations rely on runtime safeguards—duplicate contact names halt execution and the send button is clicked at most once—but these are runtime checks, not guarantees against a confused or insistent model.
Verdict
Eta is worth a look if you run a rooted Xiaomi or OPPO device and want an AI that can actually manipulate system settings, send messages, and run shell commands rather than just chat about them. If you are on a different OEM skin, unrooted, or prefer your Xposed modules in the past tense, this is not your project.
Frequently asked
- What is Mangi-11/Eta?
- Eta turns Xiaomi and OPPO phones into hosts for a system-level AI agent that controls devices through structured APIs, GUI automation, and root shells instead of blindly tapping buttons.
- Is Eta open source?
- Yes — Mangi-11/Eta is an open-source project tracked on heatdrop.
- What language is Eta written in?
- Mangi-11/Eta is primarily written in Kotlin.
- How popular is Eta?
- Mangi-11/Eta has 551 stars on GitHub.
- Where can I find Eta?
- Mangi-11/Eta is on GitHub at https://github.com/Mangi-11/Eta.