MindOS: Controlling a Computer with Silent Speech Using Low-Cost Muscle Sensors and AI Agents

Some people, such as those with ALS or after a stroke, cannot type or speak, but they can still move the muscles of the jaw and throat slightly. A silent speech interface picks up these small muscle signals when a person mouths words without making a sound, and turns them into computer input. MindOS is an open-source prototype, built at TreeHacks 2026, that does this with cheap parts: small MyoWare muscle sensors feed a microcontroller, a machine-learning model recognises a few mouth shapes, and an AI agent based on a large language model (LLM) uses the result to operate a web browser. This paper explains how the system works, from the sensors to the browser, and summarises the project team's tests on the public EMG-UKA silent-speech dataset, where a simple Random Forest beat every neural network tried (best result: 62.5% accuracy on five words). It then presents three new checks that need no new recordings. (1) Because MindOS spells words with only five mouth-shape groups, many words share the same pattern: even with perfect recognition, the first suggestion is right only 78.6% of the time, worse than 99% of random letter groupings and worse than T9 phone keypads (95.3%). (2) Small recognition mistakes add up across a word: at 96% accuracy per mouth shape, the right word appears first only 64.8% of the time, and a simple decoder that allows for mistakes raises top-five recovery from 81.8% to 93.3%. (3) The filter that actually runs in the code differs from the documented one and removes almost all of the slow signal the recommended sensors produce. We end with practical suggestions for improving such designs.