Jev Piano

The README lists the home note, scale, metre, tempo, the chord in every bar, the melody's shape, what the left hand plays, loudness, register, which note each bar rests on, each phrase's job, whether to restate the opening idea, whether to come to rest, and when the piece is finished. Code handles the instrument: equal temperament, octave placement and voicing.

Because Jev cannot emit a note: its API is {state, questions} to probabilities, with yes/no, pick-one and score questions only. The repo instead gives Jev a bank of chords described in words and asks a multiple-choice question per bar; code plays whichever chord Jev ranks highest.

Follow the README: npm ci && npm run build, then TYPESAFE_API_KEY=... npm start and open http://localhost:8787. Extra keys can be set as TYPESAFE_API_KEY_2, _3 and so on, or as TYPESAFE_API_KEYS, and GET /api/health reports how many were found.

The README publishes no cost figures. It says a phrase takes two Jev calls together lasting about 600 ms and yields roughly ten seconds of music, so cost scales with how long you listen; the TypeSafe API key is the only billed component named. Check current Jev pricing on typesafe.ai.

No. You describe a mood in words, like a slow, sad waltz, and Jev decides the theory: it picks the key, metre, tempo and harmony from chord descriptions given only as their notes, for example G B D F — a dominant seventh chord on G. The README notes nothing filters the chord bank to the key.

Add TYPESAFE_API_KEY_2, _3, _4 and so on, or a comma-separated TYPESAFE_API_KEYS list. Calls go round robin, a key that comes back rate limited stands aside for the next one, and GET /api/health reports how many keys were found.

It is an explicitly experimental X collaboration: @BaselAshraf81 tried to make Jev play, found it would not, and put five dollars on whoever could; @LamplighterPaul suggested music theory would let it pick the next chord. The README says the pair is building it together and it is live at jev-piano.zammitpaul.com.