Moumantai is a self-hosted runtime for creating personal "mini-apps": you define the app once, the server holds its state and logic (deterministic or powered by agent), and it can be accessed and rendered natively through a thin client on any device - phone, watch, browser, or an embedded board.

Like many folks I was trying to build and customize little apps for myself, start by some web apps, but soon I realize I would want to access them across the devices I own, without opening the browser. While many agents now can generate ephemeral apps that customized to the need at the time and show in the chat interface, regenerating everything each time could be slow and costly. What I wanted is the middle ground: AI widgets/mini-apps that I can create, customize and host myself - reusable, still has the intelligence from llms when needed, and being able to access and use across devices.

So in Moumantai, an app is defined once with the schema (data), tools (actions), and faces (views). The server hosts the state, with some agent in the back as the intelligence layer, and handle all the heavy lifting. The clients just renders the UI natively, with different device-specific faces resolved on the server. The main app logic is still simple CRUD operations, with some wrapped as tools so the backend agent can provide intelligence when needed.

It is mostly a small personal experiment so designed a bit opinioned around what I use and need. Would love any feedbacks to make it more useful. In particular any security issues or performance suggestion, any recommendation to make the architecture more elegant and cut down the complexity. Any issues are welcome :)