OptiChess
A chess engine written in C++, with its own analysis interface and a bot that plays on Lichess.
The first engine, in Python, dates from 2019 and turned out too slow for what it was meant to do. It was rewritten in C++ for speed, and the project has not really stopped since.
The engine and its interface are written from scratch in C++ with raylib: the search, the evaluation, the rendering, down to the pieces, drawn by hand. No piece of Stockfish inside, no file from anywhere else.
The search is the interesting choice of the project: the main algorithm, GrogrosZero, is a Monte Carlo tree search (MCTS) guided by UCT, with a quiescence at the leaves, rather than the more common alpha-beta. Alpha-beta is still there in a supporting role: it bounds that quiescence, and a classical alpha-beta search exists separately, as an option. An MCTS-Solver, which propagates proven wins and losses up the tree, is under test. It is a harder path in chess, where tactical depth forgives little; the search is bounded with no allocation during exploration, so that memory does not become the limiting factor.
The evaluation remains hand-written, material, pawn structure, king safety, activity, and it is rendered as result probabilities, win, draw, loss, rather than a single score. The interface shows these components separately, the only convenient way to see which term is wrong.
The most worked-on term is king safety, the hardest to estimate without calculation: either it is mate or it is not. It is deliberately aggressive, even if that means slightly overvaluing the attack. So the engine happily sacrifices material to go after the enemy king, in the spirit of romantic chess. Against Stockfish, everything gets refuted; against a human, these attacks are harder to defend. Each evaluation function formalises what a player judges by instinct, so the engine plays very much like its author, only more precise tactically.
A threshold was crossed the day its judgement in analysis mode became more reliable than its author’s: the rare times doubt remained, Stockfish or Leela Chess Zero proved it right. The full story, and why being surpassed by your own creation is a pleasure: Create.
The engine runs as a bot on Lichess under the name Grogros_Zero, which forces it to hold up against real opponents, at real time controls and with a real protocol.
To track its progress without depending on the opponents of the day, it also runs a bench of 5,000 puzzles drawn from the public Lichess database. Measured on 15 October 2026: 4,049 solved (81.0%) with 500 search iterations on one thread; 72.7% with 100 ms per puzzle on one thread, and 71.9% with 100 ms on twelve threads, where parallelism does not yet bring a gain.
This search choice deliberately departs from the classic approach. It draws on AlphaZero but aims at computing the way a human player does rather than the way a network does: explore what looks promising first, and be able to say why. With no ready-made documentation for that approach, the interface unrolls the reasoning move by move: without it, there is no way to see what the search believes.
Four years on, the coexistence of transposition tables and repetition detection remains an open problem. An identical position does not have the same value depending on the path leading to it, and nothing guarantees that the current design allows for that. It is the one place in the project where it is unclear whether the problem is a bug or a decision taken too early.
It is the oldest project on this site still alive. Started in 2022, preceded by a first engine in Python, it predates the AI tools used today: that one is written by hand.
