Andee reaches 2,277 stars for an on-device Android agent
Andee, an Android app that runs an AI agent on the device, taps and types in any other app through the accessibility service and keeps its memory on the phone, carries 2,277 stars and ranks 23 on Trendshift's daily list.
Image: GitHub
Why it mattersAn agent that reads the accessibility tree instead of asking each app to add an SDK can reach every installed app at once, which is the fastest path to a general phone agent and the hardest one to audit.
A phone assistant that lives on the phone and reads the accessibility tree can reach every installed app at once, with no SDK on the far side. Andee, an open-source Android app from kuafuai, is one of the first to ship this shape with a real user base. The repository ranks 23 on Trendshift's daily list on 22 October 2026 and has 2,277 GitHub stars, up from zero on 17 September when the repository was created.
What the project says it does
Andee's README calls the app "the soul of the device" and sells it as a resident rather than an assistant. The practical description underneath is that one APK is installed on an Android phone, the user grants the accessibility and screen-capture permissions, and from then on Andee can read the screen, tap buttons, type into fields and move between apps the user already has. The app runs on Android 11 and higher, is written in Kotlin, and the project is licensed Apache 2.0 with conditions, according to the badges in the README.
The repository says the brain runs on the device and talks to any OpenAI-compatible model the user points it at. Andee pairs with OpenAI by default on an English UI and DeepSeek on a Chinese one, and the README says the only data the app sends to kuafuai is the version number and phone model when it checks for an update. The notebook the app keeps about the user is marked localOnly and filtered out of any shared payload.
What the number does and does not say
2,277 stars in 34 days is a real signal from developers. Trendshift's daily rank 23 is a public list of what has gained attention in the past day. Together they say that developers found the project and starred it. They do not say how reliably Andee completes a task on a real phone, which is the number that matters for a daily user. The README does not yet report a success rate on held-out tasks against the model Andee ships with, which is the figure a buyer should ask for before trusting the app with a real phone.
The audit question
The accessibility tree is the fastest way to control a phone and the one that gives the agent the most reach. It is also the attack surface. Any notification, any opened link, any text that lands in a field becomes an input the model has to decide against. A team evaluating Andee for its staff or product will want to see what the model refuses to do, what it logs, and whether a stranger's message can push an action the user did not ask for. The README names a hardened mode and a vault, but a public write-up of a red-team pass against the shipped build is the thing a buyer should wait for.
Source
GitHub: kuafuai/andee. Trending rank from Trendshift.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.