abod707/sting

Sting is a command line tool that lets you control your Android phone's hardware from Termux by typing plain English requests, such as turning on the flashlight, sending a notification, or reading a sensor. Instead of remembering exact command syntax, you type what you want and a small on-device AI model translates it into the correct Termux command with the right arguments. The tool is written in Rust and runs a 26 million parameter model entirely on the phone. There is no cloud service and no server involved, everything happens locally after you install a single binary plus about 52 megabytes of model weights. The model was trained specifically on the set of commands the Termux:API package provides, so it understands requests like adjusting brightness, checking battery level, or vibrating the device for a set time. Safety is built into how commands are executed. Every command is built as a plain list of arguments rather than run through a shell, which avoids a common class of security bugs. The tool also asks for confirmation before running a command unless you explicitly skip that step, and sensitive actions like reading text messages or contacts are not included by default. You can teach sting new tools by adding entries to a tools.json configuration file, describing a command's name, what it does, and its parameters, and the model will generalize to new tools reasonably well. For heavier use of custom tools, the project explains how to further train the model on your own examples. Beyond direct phone control, sting is also designed to work as a building block for other AI agents. A larger AI system can hand off a plain English instruction, and sting turns it into a validated command with no invalid options possible, since the decoding process is constrained to only produce commands and arguments that actually exist. The project also doubles as a teaching example for people learning Rust, with a companion guide mapping parts of the code to standard Rust learning material.