The rules
You own the six pits on your side of the board and the big store to your right. On your turn you scoop up every seed from one of your own pits and drop them one at a time into each following pit, going anticlockwise. You drop one into your store as you pass it, and you always skip your opponent's. Then two things can happen, and they are the whole game:
Land your last seed in your own store and you go again. Stringing several of these together in a single turn is how a lead gets built, and it is why the first move matters more than it looks.
Land your last seed in an empty pit on your own side and you capture — that seed plus everything sitting in the pit directly opposite goes to your store. The board can swing a dozen seeds in one move.
When one player's six pits are all empty the game ends, and the other player banks whatever is still on their side. Most seeds wins. Both of those rules are switchable on the start screen, because Mancala is a family of hundreds of local games and the version most sites hard-code is only one of them.
Every difficulty is free, and the ladder is measured
The loudest complaint in this category is not about the rules — it is that "only one level is open to users who didn't pay" and that "the computer doesn't really get harder". All three levels here are free and there is no paid tier at all.
More to the point, the ladder is checked rather than claimed. On every build the test suite plays the levels against each other: level two has to beat level one and level three has to beat level two, or the build fails. At the time of writing that is 88% and 100%. A difficulty label that nobody measures is just a word.
Hard is not guessing, and it is not cheating either
Mancala has no hidden information. Every seed is on the table for both players, so there is nothing for a bot to peek at — the accusation that gets thrown at Battleship and dice games genuinely cannot apply here.
What our Hard bot has is a solver. Once few enough seeds remain it stops estimating and searches the position exactly, so it plays the provably best move rather than a good-looking one. That is arithmetic, not a trick, and it is why it gets sharper as the board empties — the opposite of a bot that fades.
How we know the rules are right
Kalah is a solved game. Irving, Donkers and Uiterwijk proved in 2000 that with perfect play the first player wins the standard six-pit, four-seed game by exactly ten. That published number is the target our engine is aimed at, and the smaller variants it can solve outright come out exact — Kalah with two seeds a pit is a first-player win by ten, in under three thousand positions.
Our solver cannot reach the full four-seed game: it is still unfinished after three hundred million positions, because the original solve leaned on endgame databases we have not built. We would rather say that than quietly test something easier and imply otherwise. What the engine is checked against on every build: an independent unpruned search that agrees with the fast one on every position tried, the two rules everyone gets wrong tested one at a time, and forty-eight seeds counted at every single ply of three hundred random games — because a sowing bug shows up as a seed quietly appearing or vanishing long before it shows up as a bad move.