Privacy and data ownership start with the individual. We build open source software that you run on your own hardware, and we help you host open models the same way. Everything is MIT licensed, built with Nix and powered by NixOS.
We didn't set out to build a suite of apps. Every project started with a product we relied on and couldn't keep using. All of them are open source, run as a Go server on SQLite that you host yourself, and ship with a NixOS module.
There's no practical way to self-host Signal, and Telegram uploads your address book and tells your contacts the moment you join.
So we built end-to-end encrypted group chat that runs on your own server. There are no phone numbers or accounts. Your identity is a key on your phone, and the server only ever stores ciphertext.
Agent harnesses assume a hosted model with a huge context window that never goes down. None of them were built for people running models on their own GPUs.
super-badger sits in front of opencode and adds what local inference needs. It tells you when the model box is down, shows GPU temperature and tokens per second next to the chat, and pings your phone when a slow run needs you.
Life360 was caught selling its users' locations to data brokers. Where you are, all day, is about the most personal data you have.
super-atlas shows your friends on a map and does nothing else. It runs on your server, has no ads and no feed, and ghost mode hides you from everyone until you turn it off.
Open source is how we work. Underneath it is an older question: who gets to own powerful technology? We think the answer is everyone, and software is the part we can build today.
The printing press, radio, encryption and satellite navigation all started out in the hands of states and large institutions, and all of them ended up in ordinary hands. Every time that happened, people got freer. We want to help it happen sooner.
When only institutions hold the best tools, everyone else has to take their word for things. A public that's equipped as well as the people who govern it can check them, keep them honest, and look after itself.
GPUs, radios, flight controllers, sensors. The machines that compute, transmit and see should be things you can buy, open up, repair and run without asking anyone for permission.
What you write, where you go and who you talk to belongs to you. A company holding it is a custodian at best, and we'd rather it never held it at all.
Locking information up mostly decides who gets to use it. We'd rather publish the schematics, firmware, weights and research, so anyone can learn from them and build on top.
Every project ships a flake and a NixOS module. Builds are reproducible, so you can check that the binary you run came from the source you read. Deploying is a few lines of config, and rolling back is one command.
services.superchat = {
enable = true;
publicOrigin = "https://chat.example.com";
};
services.superbadger = {
enable = true;
web.enable = true;
};
Qwen, GLM and DeepSeek publish their weights, and they're good enough for real work now. The hard part is getting them to run well on your own hardware. We help you pick a model and the machines for it, set it all up with Nix, and stick around for as long as you want us to.
Sizes from a laptop up to a multi-GPU server. Good at code, and handles many languages well.
Strong at tool calling and agent work, with long context. A good fit behind super-badger.
Large mixture-of-experts models, including reasoning variants, for when you have the GPUs.
Whatever you paste into a hosted API, whether that's contracts, patient notes, source code or a private journal, ends up on someone else's servers under their retention policy. On your own machine it never leaves the room, and there's no outside processor to put in a compliance report.
Hosted models get retired, quietly swapped for new versions, repriced or rate-limited, and accounts get banned. Weights on your own disk give the same answers next year as they do today.
On a boat, at a field site, inside an air-gapped lab, or during an outage. If the network itself is part of your threat model, you can take it out of the picture.
Coding agents and batch jobs can run all day. With your own GPUs the cost is power and hardware you already own, and it doesn't go up the more you use it.
Want help self-hosting one of our apps, or help getting a model running on your own hardware? Or is there a tool that should exist in the open and doesn't yet? Send an email.