Meow Chess

Why it exists

A grade on a move does not tell a beginner what went wrong. Meow Chess is built for a player rated around 500 on Chess.com. It downloads their recent games, grades every move, and says in short sentences what the move did, what it allowed, and what was better. The pieces are cats drawn in code, with moods and meow sounds, because the review should be something a beginner wants to open.

The engine

The engine lives in core/ and depends only on serde and serde_json. It has bitboards and attack tables, legal move generation, SAN and PGN parsing, a tapered evaluation with piece-square tables, and an alpha-beta search with iterative deepening, a transposition table, quiescence, null-move pruning, late move reductions and killer and history ordering.

Move generation is checked with perft against five standard positions, including Kiwipete to depth 3 (97,862 nodes) and the start position to depth 4 (197,281). All 21 engine tests pass. The SwiftUI app calls the engine through a small C interface that exchanges JSON, and any error comes back as JSON rather than a panic, since a panic in a release build would crash the app.

Explaining moves

The review layer finds hanging pieces, forks, pins, missed mates and missed free material, and a separate pass describes what a move is trying to do: attack, defend, develop, castle, push a passed pawn, or answer a threat found with a null-move search. A plain pawn move gets a reason such as “opens a diagonal for the bishop” instead of “pushes a pawn”.

On the player’s own moves the app can ask “What was your idea?” and then say whether the move really did that, and which move would have. Mistakes become practice puzzles. The coach tab estimates playing strength from average centipawn loss. The first formula rated real ~400 games at about 1,600, so I recalibrated it on those games.

Optional AI

Everything above works offline. For chatting about a move or an AI training plan, the user can point the app at any OpenAI-compatible server, OpenRouter by default or a local model such as Ollama, with their own key stored in the iPhone Keychain. Jev, a decision model served through OpenRouter, can read what the player typed in the idea quiz and pick which idea they meant. Jev requests carry no usernames.

What leaves the phone

I run no server, and the app has no accounts, analytics or ads. Games come from Chess.com’s public API. Reviews, quiz answers and settings stay on the phone. The AI features are off until a key is entered, and the privacy policy lists exactly what each one sends.

Where it stands

The v2 task list is complete, and anyone with an iPhone can join the beta on TestFlight. Getting there took an icon without an alpha channel, the export compliance key set, and a privacy policy written for testers. Two end-to-end UI tests drive the app against the real Chess.com API and a mock AI server. Next on the list are optional Stockfish checks and recalibrating the rating formula with games at other ratings.