The rules, in the order they bite
Four players, thirteen cards each, and the whole pack is dealt. Before a card is played you choose three to pass — left on the first hand, right on the second, across on the third, and on the fourth nobody passes anything. Then it repeats. That fourth hand is the one most online versions forget, and it matters more than it sounds: it is the only hand where you cannot get rid of the queen of spades, so a hand you would happily break up on any other deal has to be played as it is.
The two of clubs leads the first trick, and it is whoever holds it after the pass. It is not a choice — it is the only card that player may put down. On that first trick nobody may throw a heart or the queen: Bicycle's official rules say that "if a player has no clubs when the first trick is led, a heart or the queen of spades cannot be discarded". The single exception, which Pagat records and which matters because without it the game can deadlock, is a player holding nothing but penalty cards.
You must follow the suit that was led if you can. If you cannot, you may throw anything — and that is the moment the queen of spades leaves your hand and lands on somebody else's trick. Hearts may not be LED until they are broken, which means a heart has already been played to an earlier trick. Again there is an exception: a hand of nothing but hearts may lead one.
The scoring is one number: twenty-six. Thirteen hearts at a point each plus the queen at thirteen, and every one of those cards ends up in somebody's pile — so the four scores for a hand always add to 26. Bicycle states the same fact from the other side: "the aggregate total of all scores for each hand must be a multiple of 26." Our engine asserts it on every completed hand of thousands of test games, because that single identity catches a miscounted heart, a queen scored as one point, a trick given to the wrong seat and a card that silently vanished from a hand — all of them, with one line.
Shooting the moon, and the fact that the rulebooks disagree
Take all twenty-six and you have shot the moon. What happens next is a real fork in the published rules rather than a house preference, so it is a setting here instead of a decision we made for you:
- Bicycle: "that player scores zero and each of his opponents score an additional 26 points." This is the default.
- Pagat: "your score is reduced by 26 points, or you may choose instead to have all other players' scores increased by 26 points."
The shooter's-choice version is the interesting one, and it is worth knowing why the choice matters, because it is not what it looks like. Taking −26 yourself and giving +26 to the other three move every gap between players by exactly the same 26, so the standings after the hand are identical either way. What differs is when the game ends: adding 26 to everyone else drags all three of them toward 100, and subtracting 26 from yourself does not. So the real question is whether you want the game over sooner — which you do, exactly when you are the one in front. Our engine works that out rather than guessing.
"The bot always gets the queen" — the accusation, and the answer
Every online hearts is accused of dealing the computer the queen of spades, and the accusation is unfalsifiable from the player's side, which is exactly what makes it corrosive: nobody can win the argument, so nobody believes the answer. There are two separate claims hiding in it and they need two separate answers.
Does the bot look at your cards? No, and that is structural rather than promised. The bot is handed a view of the table containing only what a person in that chair could see — its own thirteen cards, every card already played, who has failed to follow which suit, what each seat has taken — and there is no path from that view to anybody's hand. Our test suite proves it the only way it can be proved: it stops mid-hand, asks the bot for a card, shuffles every hidden hand at the table underneath it, and asks again. A peeking bot cannot survive that; a blind one cannot notice. Nine hundred paired decisions across all three levels, every one identical.
Was the deal itself fixed? Before a single card moves we publish the SHA-256 of a secret seed. You supply a seed of your own — type anything you like — and every shuffle in the game is a pure function of those two values and the hand number. Nothing else: not the score, not who is winning, not whether you are about to beat your best result. When a hand is scored we hand you that hand's seed, and when the game ends we hand you the master, so you can hash it, confirm it matches the commitment you were shown at the start, and re-deal the whole game yourself. The same protocol on our dice, with a SHA-256 self-test against the published FIPS vectors.
🔴 And the honest limits, because a hash that reassures without proving anything is worse than no hash. It proves the deal was fixed before the cards were dealt and that it matches the seed. It does not prove there is an honest server somewhere — against the bots there is no server at all, the whole thing runs in your tab. In a live game one seat derives the deck, so that browser holds the seed while the hand is played; both seats commit before either reveals, so neither side can pick the deal, and that part is proved. Hiding your thirteen cards from a modified opponent client is a different problem (mental poker) and we have not built it.
How strong the bots actually are
Numbers rather than adjectives, and they came out of
tools/hearts_test.mjs on the run that shipped this page — 1,500 full games per
pairing, two seats each so the seating cannot favour a level:
- Medium beats Easy 96.7% of games (average final score 29.5 against 89.9). That gap is enormous because Easy is close to random.
- Hard beats Medium 53.3% of games (74.6 against 77.6). That is a real edge — 2.6 standard deviations over 1,500 games — but it is a small one, and we would rather say so than call it "expert". Hard is a slightly better hearts player than Medium; it is not a strong hearts player.
⚠️ And Hard was worse than Medium until the measurement said so. It shipped with two "smart" behaviours that every hearts strategy guide recommends — pass for a void, and defend against the moon early — and both of them lost games. The void pass kept throwing away three small diamonds while holding the ace and king of spades that were about to be handed the queen. The early moon defence made Hard win tricks on purpose to stop a shoot nobody was attempting, and it ended up accidentally shooting the moon 62 times in 4,511 hands while eating points every hand it did not. With both in, Hard lost to Medium 174–226. The version that wins is the one with less cleverness in it, and the whole record is in the engine's comments.
There is no published optimal play for hearts to check any of this against — no solve, no perft, no reference table — so nothing here says these bots are good in absolute terms. It says they are ordered, and by how much.
What the published numbers say, and where one of them is wrong
A hearts deal is arithmetically a bridge deal — four hands of thirteen from one pack — so bridge's published combinatorics are hearts'. Wikipedia's Contract bridge probabilities gives 635,013,559,600 distinct thirteen-card hands and 53,644,737,765,488,792,839,237,440,000 distinct deals. Our suite computes both from scratch in arbitrary-precision integers, by two independent routes each, and asserts them against those digits; it then measures our own dealer against the exact distribution they imply, cell by cell, over 240,000 dealt hands.
🔴 One widely-surfaced figure about this game is simply wrong. hearts.co's "Decoding Hearts — A Statistical Dive into the Popular Card Game" states that "the probability of receiving the Queen of Spades in the initial deal is 1 in 17 (approximately 5.88%)". It is 1 in 4. One card, four hands, and the deal does not know which card it is — 13/52, exactly, and our suite proves it three separate ways. The published figure is wrong by a factor of 4.25 and it is one of the first answers you find if you ask the question.
The same page gives the chance of somebody shooting the moon as "approximately 1 in 525 (or 0.19%)" with no methodology, no sample size and no source. We cannot check that against anything, and neither can anyone else — the moon rate is a property of the players, not of the game, so a table of beginners and a table of experts do not share a number. What we can do is measure ours and label it: across 3,977 bot hands the moon was shot in 1 hand in 67. That is our bots, not hearts, and the two figures do not check each other.
Live play seats two people and two bots, and here is why
Press Private table and you get a four-letter code and a link. Send it to whoever you want to play and they are sitting opposite you, with a house bot on each side. Find a table drops you into the oldest hearts table that is waiting; if nobody is waiting it opens one for you and seats the next person who goes looking. Live games move a rating; bot games do not, because a browser can lie about its own score and we would rather label that than launder it.
🔴 Two humans rather than four is a real constraint, not a corner cut. A room here is one broadcast channel and every seat in it receives every message. With two people that is enough for a hidden-hand game — the seat running the rules sends the other seat its own thirteen cards and there is nobody else listening. With four it is not: a snapshot carrying four hands is readable by all four clients, so a modified client could read the table. Doing it properly needs per-seat encryption — mental poker — and we have not built it. Two humans plus two bots is a full four-handed game of hearts with no such hole, so that is what ships, and this paragraph exists instead of a "4 players!" badge.
A dropped connection does not end the game. Refresh, close the laptop, lose signal in a tunnel — the same link puts you back in your seat with the hand, the scores and the history where you left them. While you are away the house bot covers your turns so the table is not frozen, and it hands the seat straight back when you return.
What this version does not do
Stated rather than left to be discovered. Four players only — the three, five and six-handed versions need cards removed from the pack and are not implemented. No variants: no Jack of Diamonds (−10), no Spot Hearts, no Black Maria, no Omnibus, no Cancellation Hearts. No clocks. The Learn coach explains the bot's reasoning and is available against bots only — refused in game logic, not merely hidden, because engine help against a person is cheating whatever the UI does. And the bot that covers a disconnected player's seat plays for them, never for you.
Everything here is free, and that is the whole difference
All three bot levels, live play, every hand, no account, no coins, no chips, no daily limit and no adverts between you and the table. Nothing is locked and nothing is sold. If you would rather play something else for nothing, gin rummy, cribbage and chess are here on the same terms.