Ruma

Ruma lives in the menu bar and comes up on a hotkey. It answers, it summarises, it works with what is on your screen. All of it happens on the laptop. There is no API key, no account, and nothing leaves the machine.

I wanted to find out whether a local model could feel instant rather than merely private. Privacy is easy to promise and easy to deliver. Speed is the part people actually judge you on, and a model that takes six seconds to start answering is one people stop opening.

Most of the work went into the parts that are not the model: keeping weights resident so the first token does not wait on a cold load, budgeting memory so the machine stays usable while it runs, and moving work onto the Neural Engine and Metal instead of the CPU. MLX and Core ML each turned out to be better at different halves of that.

Five beta releases between June and July 2025, MIT licensed, 31 stars.

I stopped maintaining it in August 2025. It had answered the question I built it to answer, and the thing I wanted to build next needed all of my time. I would rather say that plainly than leave it looking abandoned.