A chess opponent that plays like a person

Chess vs Computer is a chess site where you play against an engine that you can tune from beginner to very strong, and that tries not to feel like a machine. It is a Python (FastAPI) back end with Stockfish behind it, in a Docker container behind the same nginx as everything else on this server.

Playing

You pick standard chess or Chess960 (a random one of the 960 starting positions each game), a strength from 400 Elo up to a configurable maximum, and whether the game is timed. When it finishes you can export the moves as PGN or save a screenshot of the board. On a phone the layout collapses to the board, the status line and a few buttons, with settings in a popup.

Making the engine weaker, on purpose

Below 2500 Elo the app asks the engine for its top candidate moves and picks among them using a centipawn-loss budget: a weaker setting allows a bigger average drop in move quality. The tables for this live in one place in app/main.py, so making a level easier or harder is a matter of editing numbers.

For something closer to human play there is an optional second opponent, Maia, a neural network trained on human games. It needs lc0 and the network files, which the Dockerfile installs on a best-effort basis: if that fails, the site still works and the Maia choice is greyed out.

Feeling human

A lot of small things add up:

• Style: balanced, aggressive, solid, trader, or a random pick. Each has a name and a preference for different kinds of moves.
• Time: in timed games the reply time follows the engine’s clock, with the odd long think or instant reply, and in some games it becomes “panic-prone” and plays faster and sloppier as time runs low.
• Mood: it can tilt after a bad run and get careless when far ahead.
• Resigning: when it is clearly lost for a couple of moves it may resign or offer a draw, and how readily depends on a random temperament per game.
• Chat: an occasional remark, and a goodbye at the end.

Learning from a game

While you play, the server rates each move (best, good, inaccuracy, miss, mistake, blunder) by asking Stockfish about the position. A “miss” means failing to punish your opponent’s mistake. Afterwards the analysis screen gives:

• a short written summary built from those ratings
• accuracy broken into opening, tactics, strategy and endgame
• a rough “possible Elo” for each side, from the average centipawn loss mapped back through the same tables the engine plays by (only a ballpark figure, and it needs at least six moves per side)

There is also a daily puzzle. The server fetches Lichess’s free daily puzzle once a day and caches it, falling back to a small built-in set if Lichess is unreachable. Your streak is stored in your browser, so it does not follow you between devices.

Running it

It is a Progressive Web App, so it can be installed from the browser menu on a phone. The container publishes its port on localhost only, and the number of simultaneous Stockfish processes is capped in the compose file so a busy moment cannot eat the whole machine. Think time, strength range and the process cap are all environment settings.