IT OPENS THE APP, CLICKS THE BUTTON, TYPES THE TEXT BEFORE YOU FINISH THE SENTENCE Every voice assistant demo so far waits for you to stop talking, then processes the whole command as one block > jevme doesn’t wait - it acts on each completed clause while you’re still mid sentence It reads the front window through the macOS accessibility tree instead of screenshots - 20 to 250 milliseconds per read - so most steps never touch a vision model at all > That’s the actual unlock: a vision model call is slow and expensive specifically because it has to interpret pixels. Skip the pixels, skip that cost One line in the repo says exactly what this design is for: acting on each completed clause while you’re still talking Not faster transcription - a genuinely different pipeline where perception and decision happen in parallel with your own speech instead of after it 144 unit tests run on macOS in CI on every push, which is unusual rigor for a personal voice agent project - this wasn’t shipped as a demo and abandoned The pattern shows up everywhere Jev gets used for interface control, not just here: a closed-source macOS voice agent called Yappy reports 275-690ms per decision at $0.003 for five actions, on the exact same principle - read the accessibility tree, pick one operation, only wake a language model when actual text needs writing The design lesson underneath all of it: latency in a voice interface isn’t a nice to have, it’s the entire product - and it’s solved by removing a model from the loop, not by making the model faster full below