A chess review that runs on your own machine

Every chess site will tell you that a move was a blunder. Far fewer will tell you how much it cost, which of your blunders keep repeating, or what would have happened if you had played the move you first thought of. Those are the three questions that actually change how someone plays, and they are the three this was built to answer. The exact thresholds are published in the guide to how moves are graded.

The second reason is simpler: a chess engine is a program that runs on a computer, and you already have one. There is no technical reason your games need to be sent anywhere to be analysed, so they are not.

The engine is Stockfish 18, compiled to WebAssembly and run in a background thread inside your browser. When the page can be cross-origin isolated it uses the multi-threaded build on up to four cores, keeping one free so the board stays responsive.

Every game is looked at twice:

• A scan. Every position in the game is searched to a fixed shallow depth. This is what produces the first grade for each move and fills the move list as it goes.
• A second look. The moves that lost real ground are searched again, deeper — up to a dozen of them, worst first. A move is only ever compared against a search of the same depth, because comparing a deep evaluation with a shallow one invents mistakes that were never played.

Grades come from win probability rather than raw centipawns: dropping from +9 to +7 is not the same kind of error as dropping from +1 to −1, even though the pawn count moved further in the first case. Accuracy is the average of what each of your moves kept.

• Review a game from chess.com or from any PGN you paste, move by move, with the engine's line and the cost of what was played instead.
• Try your own moves. Play anything from any position and the engine judges it against the move that was actually played. The board turns slate to make clear you have left the game.
• Review a batch of games and see what keeps going wrong — by phase of the game, by colour, by time control, by opening.
• Train on your own blunders, turned into puzzles from the positions where you actually lost the game.

It will not help you cheat. The tool is built for looking at games that are over; using an engine during a live game is cheating, and it will get you banned from every site worth playing on.

It also will not pretend to be certain. Grades are an engine's opinion at a fixed search depth, and a deeper search sometimes disagrees with a shallower one. Treat a borderline verdict as a prompt to look at the position yourself — which is what the board is there for.

• Stockfish 18 (GPL v3) for the analysis, as a WebAssembly worker.
• chess.js for move legality and notation, react-chessboard for the board.
• Next.js and TypeScript, with IndexedDB for the evaluation cache so a position is never paid for twice.