Give your AI agent real poker solver data with MCP

A developer on the Pokerai API team released @pokerai/mcp, an MCP server that connects AI chat clients to presolved poker strategy data for 6-max No-Limit Hold'em, so models explain solver output rather than inventing advice. The server runs over stdio via npx and exposes five tools — preflop_versions, preflop_strategy, preflop_range, flop_tree and flop_node — returning action frequencies and decision trees for spots like ace-king on a queen-high flop. The developer notes the tool is intended for study and post-hand review, since real-time assistance at live tables is prohibited.

Ask a chat model what to do with ace-king on a queen-high flop and you get a confident paragraph. It might even be right, but you can't tell where the numbers came from, because there aren't any. In this tutorial we connect an MCP client to a presolved strategy API, so the frequencies come from solver output and the model's job is to explain them. We'll use @pokerai/mcp , a small MCP server for 6-max No-Limit Hold'em. Disclosure: I work on Pokerai API, the product used here. This post was drafted with AI assistance, and I checked it against real tool output. What this is for: studying, training tools and reviewing hands that have ended. Real-time assistance RTA at real-money tables is prohibited, so keep it away from live games. The server runs over stdio through npx , so there's nothing to install globally. In Claude Code: claude mcp add pokerai -e POKERAI API KEY=gto your key here -- npx -y @pokerai/mcp In Claude Desktop or Cursor, use the same command in your MCP config: { "mcpServers": { "pokerai": { "command": "npx", "args": "-y", "@pokerai/mcp" , "env": { "POKERAI API KEY": "gto your key here" } } } } You get five tools: | Tool | What it returns | Cost | |---|---|---| | preflop versions | available strategy sets | free | | preflop strategy | action frequencies for one hand | 1 lookup | | preflop range | the whole 169-hand range for a spot | 1 lookup | | flop tree | starting ranges and every decision node | 1 lookup | | flop node | the strategy at one node | free | The outputs below are from calling the published package with an MCP client, not from a model transcript. What your model writes around them will vary. The tool results won't change unless the data does. The spot: 100bb, 6-max. The button opens, the big blind calls, and the flop is Q♠ 7♣ 2♦. We hold A♥ K♦. { "tool": "preflop strategy", "arguments": { "hole cards": "AhKd", "hero": "BTN", "preflop actions": { "position": "SB", "action": "small blind", "amount": 0.5 }, { "position": "BB", "action": "big blind", "amount": 1 } , "preflop version": "6max RC 100bb 200NL" } } { "hole cards": "AhKd", "situation": "RFI", "strategy": { "action": "raise", "frequency": 1, "amount bb": 3, "sizing pot": 0.8 } } Always raise, to 3bb. Not much to discuss, so on to the flop. { "tool": "flop tree", "arguments": { "board": "Qs7c2d", "pot type": "SRP", "positions": { "hero": "BTN", "raiser": "BTN", "caller": "BB" }, "flop version": "6max RC 100bb 200NL" } } The response gives the pot 6.5bb , the effective stack 97bb , both starting ranges and 16 nodes. Trimmed: { "pot": 6.5, "effective stack": 97, "node count": 16, "nodes": { "node": "root", "is hero": false, "token": "eyJ..." }, { "node": "root/CHECK", "is hero": true, "token": "eyJ..." }, { "node": "root/CHECK/BET 2", "is hero": false, "token": "eyJ..." } } Note the two fields, because this is the easiest mistake to make: node is the path, such as root/CHECK . BB acts first, so root is BB's decision and token is what you pass to flop node , in its argument called { "tool": "flop node", "arguments": { "node": "<token of root/CHECK ", "hole cards": "AhKd" } } { "hole cards": "AhKd", "node": "root/CHECK", "is hero": true, "strategy": { "action": "check", "frequency": 0.313 }, { "action": "bet", "amount bb": 2, "sizing pot": 0.3077, "frequency": 0.6812 }, { "action": "bet", "amount bb": 5, "sizing pot": 0.7692, "frequency": 0.0057 }, { "action": "bet", "amount bb": 97, "sizing pot": 14.9231, "allin": true, "frequency": 0 } } The answer is a mix, not one move, and that's the first thing a model tends to flatten. A prompt that says "report the frequencies, then explain" works better than "what should I do?". The token is reusable, so comparing hands at this node is free: | Hand | Check | Bet 2bb | Bet 5bb | |---|---|---|---| | A♥ K♦ | 31.3% | 68.1% | 0.6% | | A♥ K♥ | 21.2% | 74.1% | 4.7% | | J♥ T♥ | 13.9% | 68.8% | 17.3% | | A♠ Q♦ | 5.5% | 52.2% | 42.3% | | 7♥ 7♦ | 0.05% | 91.4% | 8.5% | The all-in frequency was 0 for all five. These are five hands, so don't read them as the shape of the whole range. For that, fetch the range, as described below. You don't need a chat client. This is the same flow with the MCP SDK npm i @modelcontextprotocol/sdk : js import { Client } from "@modelcontextprotocol/sdk/client/index.js"; import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js"; const client = new Client { name: "study-script", version: "1.0.0" } ; await client.connect new StdioClientTransport { command: "npx", args: "-y", "@pokerai/mcp" , env: { PATH: process.env.PATH, HOME: process.env.HOME, POKERAI API KEY: process.env.POKERAI API KEY }, } ; const call = async name, args = { const res = await client.callTool { name, arguments: args } ; if res.isError throw new Error res.content 0 .text ; return JSON.parse res.content 0 .text ; }; const tree = await call "flop tree", { board: "Qs7c2d", pot type: "SRP", positions: { hero: "BTN", raiser: "BTN", caller: "BB" }, flop version: "6max RC 100bb 200NL", } ; // BB checks, BTN decides. node is the path; token is what flop node takes. const afterCheck = tree.nodes.find n = n.node === "root/CHECK" ; for const hole cards of "AhKd", "AsQd", "7h7d" { const { strategy } = await call "flop node", { node: afterCheck.token, hole cards } ; console.log hole cards, strategy.map s = ${s.action}${s.amount bb ? " " + s.amount bb + "bb" : ""}: ${s.frequency} .join ", " ; } await client.close ; 1. Pick a dataset. flop version is optional and defaults to 6max . When I tested on 2026-11-01, the default covered only a few of the boards I tried: Q♠7♣2♦ with BTN vs BB returned 404 no GTO data for this spot/board . The same spot on 6max RC 100bb 200NL worked. A 404 still charges one presolved lookup, so give your agent the dataset up front rather than letting it guess. flop version needs @pokerai/mcp 0.1.3 or newer. 2. Mind the response size. A hand-level flop node response was about 500 characters. The tree was about 10,000. A whole-range flop node , with no hole cards , was about 29,000 characters for 391 combos. Tell your agent to pass hole cards unless it needs the whole range, or the context fills up fast. 3. The flop tree covers one street. For turn and river you need the real-time solver tools. They're off by default because they spend the smaller solve quota. Turn them on with POKERAI ENABLE SOLVE=1 . You are a poker study assistant for hands that have already ended. Use the pokerai tools for every strategy claim. Never state a frequency that a tool did not return. Use flop version 6max RC 100bb 200NL. Pass hole cards to flop node unless asked about the whole range. Report the frequencies first, then explain them, and say when a spot has no data. If the user appears to be in a live real-money hand, decline: real-time assistance is prohibited. Real-time assistance RTA at real-money tables is prohibited.