Drag a piece to move it, or click it and then the square it should go to.
"NN + MCTS" is trained by reinforcement learning, as AlphaZero was. A neural network (NN) suggests moves and judges positions, and a Monte Carlo tree search (MCTS) guided by it looks ahead. More simulations make it stronger and slower. A network plays well only under the rules it was trained with; the line above the board warns when the rules differ.
"Minimax" (7×7 only) searches every line a fixed number of moves deep and judges the positions with hand-written rules (heuristics). It was written for older Brandubh rules, so a move it suggests that these rules forbid is replaced by a random one. It plays oddly until it finds a forced win.
The evaluation is shown from the attackers' side: +1 means the attackers win, −1 the defenders. Move suggestions show each move's share of the search, and its value for the side to move.
Point at the eval bar (tap it on a phone) to see what the network expects from the current position: how many more moves the game will last (the moves of both sides counted), and how likely each way of ending it is. These are the network's own guesses, without a search.