balaji-md/bmass

BMASS, short for Bootable Model As System, is an experimental project that boots a computer directly into an AI controlled environment instead of starting a normal desktop and then opening an AI app on top of it. The idea behind it is simple: what if the AI model itself was the main way you interact with the computer, rather than just another program running inside one. When you boot from the BMASS USB drive, the computer starts a lightweight Linux environment called Alpine Linux with no graphical desktop, then automatically launches a local language model using a runtime called llama-server. From there you can talk to it in plain language, and it can inspect the machine, run Linux commands through a restricted, non-root user account, and respond using the real output it observed rather than just guessing. Commands that only inspect the system run automatically, while anything that would modify the system requires the operator to approve it first, and commands that try to gain elevated privileges are blocked outright. The current prototype, labeled Seed v0.5, runs on very modest hardware: an entry-level laptop with 4 GB of memory, booting from an 8 GB USB flash drive, using a small local model called Qwen3 0.6B. Once installed, the whole thing works offline with no internet connection needed. The project describes its goals as private AI by default, offline-first computing, portable AI environments, and giving people ownership of their own models and data, rather than trying to build the largest or most powerful AI system. The README is honest about current limitations: the security model is still a prototype, installation is manual, configuration is hard coded, and there is no persistent memory, package installer, or automatic update mechanism yet. Large models, compiled binaries, and operating system images are intentionally left out of the repository, which instead documents how the pieces fit together. The project was independently reviewed by a technical writer at Korben, and the author has since made several safety improvements based on that feedback, including requiring explicit approval before running any command that changes the system.