Everything runs on this device: the microphone signal goes through a pitch tracker (YIN, 16 kHz), the contour is scaled to your calibrated low and high notes, and simple shape rules decide: flat and high = tone 1, rising = 2, low dip = 3, sharp fall = 4. Nothing is uploaded.
Tested on the kit's native recordings: isolated syllables were accepted 15 of 16 times and the wrong tone was rejected 46 of 48 times. Tone pairs are rougher: 34 of 40 words accepted, and about 1 wrong tone in 8 slips through, mostly a rising 2nd tone passing as a 3rd. Treat the verdict as a mirror, not a teacher: when it says "again" three times in a row, listen to the native recording and compare the shapes yourself.
Syllables with "sh", "zh", "r" and creaky low 3rd tones are the hardest for the tracker. A quiet room and the phone about 20 cm from your mouth help most.