Every question you send to a cloud AI can become someone else's record. Whoever runs the server holds a switch: off, reprice, filter, watch. None of that needs a villain. It's what concentration does. And when the signal drops, the tool is gone.
Some tools are built the other way: open, with no owner, so that depending on them can't be turned against you. Sanctuary technology. BOAR is that idea, on the phone you already have.
This isn't about hiding. It's about not having to ask.
Privacy
No server receives your questions. No account. No Google Play Services. BOAR uses the internet in exactly two places, both started by you: the model and pack downloads, and searching Hugging Face for another model. Built for de-Googled phones like GrapheneOS.
Data sovereignty
No owner, nothing to switch off. The code is MIT, and you can check the APK's SHA-256 yourself. Your documents stay on your phone. Knowledge packs are plain files you can export and pass on.
Local AI on your phone
One download at setup, about 1 GB. After that, airplane mode is fine. It runs on Android phones today (any 64-bit ARM), and more platforms are coming. Phones with more memory can run bigger models. Next: a 35B tier for phones with 12 GB, not shipped or measured yet.
Receipts, under all of it
Every answer shows the sources it was built from, so you can check it. Every answer records how it ran: model, time to first token, tokens per second, memory. We publish the failures with the wins.
What we don't claim yet
- That it matches internet search plus a frontier model. Our internal test (2026-09-30) put it at about half.
- Encrypted storage beyond the phone's own protection.
- A library in any language but English.
This is the first release of an idea, and it's growing in public. Useful now beats perfect later.
Sovereign AI, in your pocket. 🐗
boarapp.com · github.com/rferrari/boar-app · Open source (MIT) · On your phone, Android today
Join us
- Test it on your phone. Run the benchmark in the app, or with one command from a computer (benchmark your phone), and share the result. A result from your device is data nobody else has.
- Try it on GrapheneOS. A run on a de-Googled Pixel is the result we're missing (how).
- Break it. Ask the questions a small model fails on. Find where retrieval pulls in nonsense. Open an issue with the output.
- Build the missing pieces. Bigger knowledge packs, better routing between small and large models, mixture-of-experts models streaming from storage. See contribute.