Atlas

Atlas is my own assistant, and I use it every day. It began as a Telegram bot in front of a Python agent on a VPS. It is now a native Android app on my two phones with that agent behind it, and Telegram is switched off.

The app is organized around what I am trying to get done. A Thread is one intention with an outcome, a next step, and a history. I talk to it in Chat the way I would talk to a person, and when something I say changes a Thread, it saves the change and shows me a receipt. It takes files and voice notes, it can answer in audio, and it can turn an email that needs an answer into a Thread and a draft reply, which it does not send.

Behind the app, each piece of model work picks one of four tiers by how hard it is, and tries my Codex and Claude subscriptions before it falls back to a paid API under a spending cap. Independent jobs go out to as many as three workers at once. My Google account sits behind a broker that owns the credential and gives the models a fixed list of operations.

Atlas can also change itself. I ask for a feature in Chat, it writes a proposal, and nothing happens until I approve it with a tap. Then a coding agent builds the change in isolation on my Mac, the build goes through fifteen release stages, and the signed app installs on both phones. In September an update I approved went the whole way by that route. Approval stays with me.

It is not a product. It is a personal tool, and the code is private. I am happy to talk through how it works.