Human move prediction API Private beta
A neural network that doesn't guess the best move, but the one a human will actually play in a given position. Trained on human games. The live demo below is free — no signup. An API key (private beta) is only needed to call it from your own code via the HTTP API.
Pick a piece, then click squares to place it. ✕ erases. Click a square again to overwrite.
How it works
The position is encoded into 17 planes of ×ばつ8 (pieces, side to move, castling rights) and passed through a small residual convolutional network that returns a probability distribution over every from → to move. The probabilities are then restricted to legal moves and ranked.
The model is trained on human games, so it learns human habits and mistakes — not perfect engine play. This is a small proof-of-concept.
How accurate is it?
Measured on held-out human games the model has never seen — the move a player actually makes is:
...versus ~6% for random legal guessing. Predicting a human's exact move is inherently uncertain — even the best research systems reach roughly 50% for the single top pick — so we report how often the real move lands in our top suggestions.
Use it via API
The same engine is available as a hosted HTTP API. Send a position as FEN, get the most likely human moves with probabilities. Access is in private beta and granted per request.
curl -X POST https://bruncsoft.com/api/chess/predict \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"fen":"rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"}'
# → {"turn":"w","moves":[{"san":"Nf3","uci":"g1f3","p":0.29}, ...]}
Request API access
Enter your email and we'll send a verification link. Once you confirm, we review the request and email you an API key. You get access to the API — not the model or training code.
We email a magic link (valid 30 min) to verify your address. No password. Your email is only used for access and key delivery.
Data & privacy
Training data comes from the public Lichess open database (games under CC0). Inference runs on Bruncsoft's server (ONNX Runtime) — the model weights and training code are not distributed; the API returns predictions only. Positions you send are used to compute the response and aren't stored.