Buyer's guide · code read 22 October 2026
AI poker bot: what the label covers, and what to ask a seller before you pay
"AI" on a poker bot listing can mean a trained neural network, a table of hand rankings, a few hundred random simulations or a chat-model request. I searched GitHub for "ai poker bot" on 22 October 2026 and read the code of 27 of the projects that came back. Seven train a model, seven simulate hands, one calls a language model, two follow fixed rules, and the remaining ten are either not bots or contain no code. The label tells you almost nothing about how the bot chooses an action.
One search, the first 30 results sorted by stars, 65 repositories in total. I cloned them without running anything and read the files. Three I skipped (a list of links, a profile repository and a one-page HTML tool), leaving 27. These are mostly coursework and contest entries, not products, so treat the counts as a picture of what the phrase attracts, not of the commercial market.
What the 27 projects do when you open the files
I sorted each by the code that picks fold, call or raise, ignoring README claims. The README of one project promised "advanced" play with a 65% win rate; the figure is against random opponents, the weakest benchmark there is.
| What picks the action | Projects | What I found in the code |
|---|---|---|
| A model trained by self-play or reinforcement learning | 7 | Methods named in the code: CFR and its Monte Carlo variant, deep CFR, NFSP, deep Q-learning and plain policy networks. Only three commit anything trained: a TensorFlow checkpoint, a set of pickled tables and a stored policy surface. One loads a model.pth that is not in the repository. |
| Monte Carlo equity against random hands, then thresholds | 7 | Deal two hundred to twelve hundred random runouts, compare the win rate with pot odds, then fold, call or raise by fixed cut-offs. One of them labels this tree search. |
| Equity table plus a call to a language model | 1 | A 749-line browser page. The action frequencies are hardcoded; a single API request turns the numbers into a three-sentence suggestion. |
| Fixed hand-ranking rules | 2 | Branches on the made-hand type and on a ranking of the 169 starting hands, with a random number as a tiebreaker in one of them. |
| Not a bot at all | 7 | Tournament engines, an API for bot contests, a spectator arena for agents, a hand-history parser, a scraper for a university leaderboard, a card game inside an editor plugin, a game-data generator. |
| Nothing in the repository | 3 | A README with one sentence, an empty repository, a file-less project from 2012. |
Four things sold as AI, and what each one can promise
Rules. A chart of starting hands and a few thresholds. It is cheap to build and easy to read, and it behaves the same way every time. Ask to see the chart. If the seller will not show it, it is probably short.
Simulation. The bot estimates its chances by dealing random cards, then compares them with the price of a call. It handles one opponent type well, the random one, and it treats a tight raise and a loose raise as the same event. Strength in this class comes from how the thresholds were tuned, which only a long sample can show.
A trained strategy. Self-play methods can produce a strong policy for a simplified game. The published research bots were built with heavy compute and played inside a fixed betting abstraction. A claim that a retail license ships something like that should come with a file size, the stack depths covered and the training cost. In my 27, one project is a large Rust implementation of a six-handed design from a research paper, which I read but did not run; none of the 27 is wired to a real-money room.
A language model. A chat model can explain a spot. As the only engine, it is slow and each decision is expensive. Its arithmetic is unreliable. The one project in my set that uses it keeps the real numbers in a conventional equity table and sends only a summary to the model. If a seller says a language model plays your hands, ask where the pot-odds calculation happens.
Eight questions to ask a seller
Put these in writing before paying. A seller who has built the thing can answer in a sentence each; one who resells a wrapper often cannot.
| If the listing says | Ask for | Evidence that counts | Weak answer |
|---|---|---|---|
| Neural network | Model file size, input features, training data source | A number in megabytes, a feature list, a statement of what was trained on what | "Proprietary deep learning" |
| Solver-based, GTO | Which stack depths and bet sizes the strategy covers | A named abstraction, such as one preflop depth and three postflop sizes | "Full GTO", with no tree description |
| Adapts to opponents | What it records about a villain, and after how many hands it acts on it | A stat list and a sample threshold | "Learns your opponents" |
| Uses an LLM | Which model, and whether it sits in the decision loop or only in the log | A call per decision with a latency figure, or a clear statement that it is advisory | No answer on latency |
| Plays like a human | What timing and sizing variation is built in, and what the room terms say | The room clause quoted back to you | "Looks natural" |
| Updates included | Update period in the contract, and whether old builds keep working | Dates in the license text | "Lifetime" |
| Works at my stakes | Hands-per-hour and CPU load on a machine like yours | Two measured runs | A screenshot of winnings |
| Tested | Sample size, opponent type, bb/100 with a confidence interval | Hands played and variance, not a graph | A rising line with no axis labels |
What none of the 27 shows
I searched the code for screen capture and input automation. Two projects use a browser automation library, and neither drives a poker client: one scrapes a university contest leaderboard, the other has it in a test file of a spectator arena. No project in the set reads a table in a real room and clicks. That gap is where a paid product earns its price or fails to, and it is invisible in any "AI" description. The questions that matter there are about table recognition, supported clients, update speed after a client patch, and what the room terms say about automated play. Most rooms prohibit it, and buying a license does not alter that. Our earlier pages cover that part: what rooms say and whether bots work, what a trial should ask of you, and where sellers appear.
Reading a demo
A demo video proves that the software runs. It does not prove what decides the action. Ask for a session in which you choose the hands: a board and stacks you name, played twice, with the same cards. A rules bot gives the same line twice. A simulation bot varies a little, since random draws differ. A trained policy can mix actions in the same spot, and the seller should be able to show the frequencies. If the seller refuses a repeat with your hands, you have only the video to judge.
The price page matters as much. The pricing on the home page lists what a license includes; set that list beside the answers above and mark every line without a date, a number or a file name.
What this read does not cover
These are public repositories, mostly written for a course or a contest, and a seller's product is closed code I cannot open. A commercial bot could use any of the methods above, or something else. Reading 27 public projects does not grade any vendor, ours included. When a listing says "AI", you have four things it might mean and a question to separate them.
The search itself is a limit. GitHub ranks by stars, then by recency, so older popular projects sit above new ones, and the phrase "ai poker bot" pulls in academic work. A search for a different phrase would produce a different mix. I did not search other code hosts, and I did not test whether any of these projects beats a human player.
Short answers
Is a neural network better than rules? Not by default. A well-tuned rules bot beats a badly trained network, and one trained on the wrong game can lose to a call-everything opponent.
Is open-source code a shortcut? It is readable, which is rare in this market. None of the 27 I read was a ready program for a real-money room.
Can I check the "AI" claim myself? Partly. Model size, repeatability on a fixed spot, and timing per decision are things you can measure in a trial. Win rate over a few thousand hands is not.
Will the counts stay the same? No. GitHub results move daily. Run the search again and count your own.