
AI Assistants Guess at Poker Math. We Gave Them Our Engine Instead.
Language models answer poker equity questions from intuition, and intuition is wrong postflop. So we exposed our exact enumeration engine as a public MCP server.
Aug 21, 2026 · 6 min read · PokerInk Team
A commenter on one of our YouTube hand breakdowns told us to check our math. Pocket aces against jack-ten suited, he said, is an 80/20 favorite, and we had built a whole video on a false premise. He was quoting a real number: preflop, aces beat jack-ten suited about four times in five. But this all-in happened on a 9♥ 8♥ 2♣ flop, where jack-ten of hearts has a flush draw, an open-ended straight draw, and two overcards. Count every card that completes something: the true number is 56.3% against the aces, exactly, across all 990 possible runouts, the turn and river combinations still to come. The intuitive answer is not just off, it is on the wrong side of even money.
Here is the part that matters if you use ChatGPT or Claude to talk through hands: the AI makes the same mistake, for the same reason. Ask an assistant a postflop equity question and it answers the way our commenter did, by pattern-matching to famous preflop numbers it has seen before. Equity is not a fact anyone can remember; it is a count over the exact cards, and it changes completely with every board card. Your instincts improve with practice. An AI's do not, so the fix is to stop letting it guess and hand it a calculator.
So that is what we did. PokerInk's poker engine is now available to any AI assistant, free, no account, no key. Connect it once and every hand debate you have with your assistant gets settled with exact math instead of confident guesses.
What "exact" means, and why you should care
Most equity calculators simulate: they deal a few hundred thousand random runouts, count the wins, and report the average. Close, but two tools will disagree on the decimals, and the argument reopens. Our engine does it the other way: it checks every single runout that can still come. For the flop above, that is all 990 turn and river combinations. For a preflop all-in, it is 1,712,304 full boards, checked in under half a second. The number you get is not an estimate of the answer; it is the answer, and the response tells you how many runouts it checked.
The engine is not new. It is the same one that grounds the AI coach inside PokerInk, where it exists so the coach can never misread your gutshot as a flush draw, and the same one behind our free odds calculator. Before it earned production, it had to agree with a slower, obviously-correct evaluator on thousands of randomly dealt showdowns, every single time.
Connect it to your assistant
Claude.ai and the Claude desktop app: Settings, Connectors, add custom connector, paste https://pokerink.com/api/mcp. ChatGPT (paid plans): Settings, Connectors, turn on Developer mode under Advanced settings, then add a custom connector with the same URL. On the command line, Claude Code and OpenAI Codex each take one line:
claude mcp add --transport http pokerink https://pokerink.com/api/mcp
codex mcp add pokerink --url https://pokerink.com/api/mcp
Anything else that speaks MCP, Cursor and friends included, takes the same URL. Or skip the assistant entirely, straight from a browser or terminal:
curl "https://pokerink.com/api/poker/equity?p1=JhTh&p2=AsAd&board=9h8h2c"
Then ask the question that started this: "was jack-ten of hearts really ahead of aces on a 9h 8h 2c flop?" Your assistant will stop guessing, call the tool, and give you the number we got: 56.3. If you are learning the game, this is also the fastest way to build equity intuition: guess the number first, then ask, and pay attention whenever the two disagree.
What it can answer
Two tools, both built for the questions players actually argue about. The first computes exact win, tie, and equity percentages for any two to four specific hands, on any board or preflop: all-in disputes, "was I ahead when the money went in," whether that cooler was actually a cooler. The second reads a single hand on a flop, turn, or river: what you actually have, plus every draw with its outs counted honestly, no double-counting the card that completes two draws at once. Every answer includes a link that opens the same spot in the odds calculator, so anyone in the conversation can see the computation for themselves. If you want the shortcuts for doing this at the table, where no calculator is allowed, that is the Rule of 2 and 4.
For the technically curious: the protocol details, request limits, and API specification live on our developers page.
Why give the engine away
Because poker arguments deserve correct answers wherever they happen. The community settles hands in comment threads, group chats, home game debates, and now in AI conversations, and the math that settles them should be exact and free to check in every one of those places. Free calculators have existed for twenty years; what we are adding is the exact kind, in the places arguments actually happen now. If a player never installs PokerInk and still walks away with the right number and a link to verify it, the endpoint has done its job.
We built PokerInk on the rule that the AI never gets to assert a poker fact; a deterministic engine computes the facts and the AI explains them. Giving the engine to every other AI is that same rule, applied to the whole table.
The same engine coaches your own hands
PokerInk is free to start on iPhone, Android, and the web. See pricing.
Frequently asked questions
Is the PokerInk MCP server free?
Yes. No account, no API key, no usage tiers. Compute is bounded by input caps and admission control rather than billing.
Why exact enumeration instead of Monte Carlo simulation?
Monte Carlo answers wobble on the decimals between runs and between tools, which reopens the exact arguments a calculator exists to settle. Enumeration checks every runout once; the answer is a fact, and the response tells you how many runouts it checked.
What card notation do the tools accept?
Rank then suit letter: "AhKs", "10d 7c" or "Td7c", with spaces and commas optional. Tens accept both "10" and "T".
Can the endpoint reach any user data?
No. The tools module is pure math over cards passed in the request, with a hard boundary against importing anything stateful; there are no accounts on the endpoint at all.
What are the limits?
2 to 4 hands, 0 to 5 board cards, requests over 16KB rejected, and a compute budget that returns 429 with a Retry-After header when the server is saturated.



