jacksonjp0311-gif/perci

Perci is an experimental AI assistant written in Rust that is meant to run entirely on your own machine instead of relying on a cloud AI service. The project describes itself as a small operating system for thinking, built by separating different jobs that most chat assistants hide inside one large model. It has a compact binary knowledge structure called Bitwork, about 200 megabytes in size, that handles routing and pattern matching. It has named procedures called operators that handle things like deciding whether to trust an instruction, planning a response, or writing code. It also has exact tools for math and geometry that calculate real answers instead of guessing, and a governance layer that checks results before anything is allowed to change the system permanently. When you send Perci a message, it first checks for quick reflex style responses or exact calculations, then routes the message through its Bitwork field, applies the relevant operators, and if needed builds a longer chain of reasoning that gets rewritten into plain, conversational language before being shown to you. You can inspect what happened behind a response using special commands that reveal its internal reasoning trace, and you can adjust how concise or detailed its answers are. The project places heavy emphasis on testing and measurement, publishing tables of pass and fail results from its own internal test suites that check reasoning correctness, dialogue behavior, and security boundaries, along with scripts you can run yourself to reproduce those results. Importantly, the underlying model weights can never update automatically. Any permanent change requires a human to explicitly authorize it. The full runtime package, including the Bitwork model files, is distributed through Git LFS and must be pulled down separately after cloning the repository. Perci is dual licensed under MIT or Apache 2.0.