A browser-based multiplayer Texas Hold’em platform — no registration, type a nickname and sit down. Up to nine seats per table, any of them human or AI, with reconnection handling so a dropped connection doesn’t fold your hand.
The part I care most about is the probability layer: a Monte-Carlo simulation engine computes your live win equity as the board develops, a card tracker keeps the deck state visible, and the decision hints are derived from the numbers — not vibes. The AI opponents come in three difficulty levels and play the same probabilistic game you do.
Details that make it feel finished: adaptive background music that follows the state of the hand, spectator mode for busted players, responsive layout down to phones.
Auditing the engine (October 2026)
A poker game that looks right can still pay out wrong. So the engine got the same treatment as any other system I depend on: a test harness that tries to break it, and checks on the harness itself. The audit was run by the engineering agent of my agent fleet, with me reviewing the results.
- A rules fuzzer. Random tables (2–9 seats, random stacks, bots plus scripted “humans”) play through the same functions the server calls, and every hand is checked against 16 rule categories. On the original engine, 100,000 hands turned up 9 rule bugs. The worst: no side pots (a short all-in could win the whole pot), tied hands not split, blinds always posted by the same two seats, and acting out of turn.
- Each bug got a reproducing test and its own fix. The 15 new tests fail on the original engine and pass on the fixed one. After the fixes, the same 100,000 hands produce zero violations.
- The fuzzer was tested too. Five bugs were planted on purpose in the fixed engine; each one was reported in the right category.
- The hand evaluator was cross-checked against the independent
treyslibrary: 1,981,857 of 1,981,857 hand categories and 1,982,033 of 1,982,033 showdown winners agree. Two planted evaluator bugs were caught, so the agreement is not a dead check. - A new bot, measured rather than asserted. An equity-vs-pot-odds bot played 20,000 duplicate heads-up hands (same cards dealt twice, seats swapped) against each existing level. It beats the beginner bot by +76.8 bb/100 (95% CI 70.1 to 83.5) and the advanced bot by +41.0 (25.5 to 56.5). Against the intermediate bot the interval includes zero, and so did a retest of a tuned variant on fresh cards. That is reported as a null result, and the default settings stayed as they were.
The fixes sit on a review branch and are not merged into the public repository yet.