Your training for today. Items appear only when there is something to do.
Checking…
The whole app is a single file that works offline. Save a copy to open it from your Files app or a desktop; progress in that copy stays in the browser that opens it.
Set-up position. Saved, but not counted in your record.
Saved. The moves are on your Today page.
The numbers are targets, not ratings. No level has been played against rated humans, so a level is where a setting is aimed rather than a strength anyone measured. The dials shown are what actually changes.
Self-play, six games a pair with colours alternated: level 800 beat level 0 six times out of six, and level 1600 scored 5½/6 against 800, so the scale does climb. Neighbouring levels were not separated — a hundred-point gap is about a 64% expected score, and no affordable number of games tells that from a coin — so the step between, say, 900 and 1000 is a design choice. If a level feels wrong, move up or down until games are close and ignore the number.
The strength estimate on an analysed game comes from a second measurement (5 September 2026): eight levels from 0 to 2100 each played two games against themselves, and every game was analysed at each of the three depths — 96 analyses. Average loss per move falls with the level closely enough to fit a line through the logarithm (R² 0.75 to 0.81) up to about 1200; above that every level looks the same to the analyser, so the estimate stops there and says "or above".
Changing side starts a new game.
The engine's assessment. Positive means it considers you better.
Answers are written from what the engine found in this position — the pieces, the evaluation, the top lines, and any move you name, searched on its own. The coach may not tell you anything the engine did not find, and it may say there is no clear reason.
The coach needs a connection to Claude, which this view does not have. Everything else here works offline: the engine, the analysis and the training.
Hand this position to the computer, at the level set on Play. You take the side to move.
Your games are compared with the lines below. The opening named is the one the game followed furthest, and transposing into another line you know does not count as leaving. The cost is two searches of equal depth, one after your move and one after the book's — and it is allowed to be nothing.
Most opening study fails the same way: you memorise ten moves, your opponent deviates on move three, and you are on your own with no idea what the position wanted. The fix is fewer moves and more reasons.
Every line and branch has been checked move by move against this app's engine, and none leaves you worse than about half a point.
Tap a move to see its purpose.
Your public games, read directly from Chess.com. No password is needed.
Chess.com publishes public games through an open archive, newest month first. The request goes from this browser to Chess.com and back; this app has no server. Importing brings the moves, ratings and clocks, which is enough for the repertoire report and the rating chart. Accuracy and a mistake list come from reviewing each game, below.
Oldest first, saved as it goes. Stop at any time and continue later. Mistakes worth practising are added to Drills.
It analyses every game that has not been analysed — and re-measures any that were, back when the reviewer was given a fixed number of seconds rather than a fixed number of positions. Under a time budget the figures depended on how fast the device was and what else it was doing, so the same game reviewed twice could come out up to 220 rating points apart. Re-measuring puts every game on Progress on the same footing. Games already measured by positions are left alone: running those again would return the same answer.
It also re-judges any game judged by a different judge from the one picked here. A stronger judge finds more wrong with the same game, so the accuracy of a game judged by Stockfish cannot be set beside one judged by this app's engine; judging them all again puts Progress back on one judge.
Paste a game from Chess.com, Lichess or anywhere else.
Chess.com: open the game, then Share, then the PGN tab, then Copy. In the app it is the share icon under the board.
Lichess: open the game, scroll to the bottom of the move list, and use the download or clipboard icon under FEN & PGN.
Paste the whole thing, headers and all. Just the moves also works.
Moments your opponent went wrong and one move punished it.
No puzzles yet. Play a game, or import or paste one, and the engine will search it for tactics.
A second, deeper search confirms that a single move is clearly best; anything with two equal answers is dropped rather than asked. Not the same as Drills: a drill is a position where you went wrong. A puzzle is a chance that was on the board for you.
Three minutes. One position after another.
Everything else here is untimed, because a drill you rush is a drill you did not do. Recognition is the exception: you know what a fork looks like; this trains finding it in four seconds. A wrong move costs ten seconds rather than ending the run.
Essential techniques, against a defence that gives nothing away.
King and pawn against king is not estimated, it is solved — all 393,216 positions worked out once and looked up after every move. That is how the app can name the exact move a win became a draw. The search cannot: shown the drawn opposition, it is sure White is winning. Where the search referees instead, the panel says so.
A game a move at a time, with the engine's view of each.
Stockfish, the strongest engine there is, against Reckless, one of this app's levels, or itself. Every move is explained: what it does, what it threatens, the line the engine expects to follow, and why it was chosen over the next best.
Two opponents drawn at random, with an opening neither of them chose. Every game is different.
With Stockfish playing, every move is explained from two things that were measured: Stockfish's own analysis either side of the move — its best move, the line it expects, and its next two choices — and the board itself: what the move takes, attacks and threatens, and any fork, pin or skewer that replaying the line proves. The line is Stockfish's real reason for a move. Where a strong move's point lies deeper than anything the line shows, you are given the line and the number rather than a story nothing computed. Stockfish looks at a quarter of a million positions a move, a fixed amount of work, so a position gets the same explanation on any device.
Against Reckless, the other engine's moves are explained the same way from its own analysis — the line Reckless expects is its reason — and Stockfish says whether it agrees, and by how much. Neither side is called a mistake for disagreeing: they each look at the same number of positions, and the rest of the game shows who was right. Reckless is free software (AGPL); the build and the one change made to it are in the app's source.
For the other games there are three sources. While the game is still in one of the app's openings, that opening's own note for the move just played. On a famous game, at the moves that have one, a sentence written about that exact move. And after every move, a search deeper than the one that played it — if it preferred something else, the gap is given in points. What you will not get is what the player was thinking; nothing here can know that.
A recorded checkmate has to be checkmate on the board or the build fails. A resignation is a fact about a room, not a position, and the screen says so rather than pretending the board proves it.
How moves are written, read from the board in front of you.
—
Tap one to read it back.
Squares come first: files are lettered a to h from White's left, ranks numbered 1 to 8 from White's side. Every square has one name, and it does not change when the board is turned — e4 is e4 from either side. The Vision screen drills those until they are automatic.
Moves come in pairs, numbered once: 1. e4 e5 is White's first move and Black's reply. A Black move quoted alone gets three dots — 1… e5. A game ends with its result: 1-0 White won, 0-1 Black won, ½-½ a draw.
Annotators attach judgements to moves: ! good, !! brilliant, ? a mistake, ?? a blunder, !? interesting, ?! dubious. Those are opinions, not notation. Game Review uses its own marks and says what each one measured.
Know every square without counting from a1.
The openings, the review and the coach all name squares. Thirty seconds a round, until the names are automatic.
Press Start.
Positions where you went wrong. Chances your opponent gave you are under Puzzles.
No drills yet. Review a game and its mistakes come back after a day, a week, then longer — for as long as you keep getting them right.
How you are playing, across every game you have reviewed.
With a mouse: right-click for red, Shift for green, Alt for blue, Ctrl for yellow. Right-drag between squares for an arrow. Tap an empty square to clear.
How hard the analysis board thinks, counted in positions rather than seconds so that a setting means the same thing on a phone as on a laptop. Opponents on Play have their own budget, fixed by level, and game review has its own setting.