Mana

The goal was presence rather than convenience. Not a faster way to run a command, but something you talk to: it listens, replies out loud, remembers what you told it last week, and has a face on screen next to the chat window rather than instead of it.

The reference point I keep coming back to is Alice from Sword Art Online, and the useful part is not the science fiction. That series separates an AI scripted to respond to stimuli from one that grows out of what it has actually lived through, and the character who matters is the one shaped by a relationship with a single person rather than by more training. Mana is not an attempt at the second thing, and models available today cannot be. It is pointed in that direction, which in practice means one instance whose memory is built from one person's conversations instead of a service answering everybody at once.

A hosted assistant is the same assistant for everybody. Claude and ChatGPT answer millions of people at once, and whatever they come to know about you sits inside a service built for all of them. I wanted one instance belonging to one person, shaped by their conversations and nobody else's. I also did not want to hand my own information to a company to keep, and I did not want a monthly bill for the privilege. Running the models locally settles all three at once: it is mine, nothing said to it leaves the machine, and there is no subscription and no per-token cost.

Running it locally answers that and creates a harder engineering problem, because the transcription, the model, the voice and the vision all now have to fit and stay responsive on a single consumer machine.

Two Electron applications, a launcher and a desktop client, share one Node backend. Any feature has to land in both, which is a constraint the codebase enforces on itself.

Speech in runs through whisper.cpp, replies through llama.cpp, and speech out through a local synthesiser. A vision model plus OCR handle screen awareness, and local web search runs through a self-hosted SearXNG instance.

The avatar is Live2D, rendered in a transparent always-on-top overlay window.

Code is Apache-2.0. The artwork and avatar assets are not, and the Live2D runtime is fetched at setup under its own licence rather than vendored into the repository.

The project is a working developer preview rather than a finished product. Voice conversation, local memory, screen reading and the avatar all function on a real Windows setup, and the release notes are honest about which parts are still rough.