AI agents are terrible at math they can't verify. Ask a large language model for Game Theory Optimal (GTO) frequencies on a poker board, and you get a confident, hallucinated paragraph. The new @pokerai/mcp server fixes this by piping presolved strategy data directly into your agent's context window, turning the LLM from a guesser into an explainer.

The Tooling Stack

The server is a Node.js package designed for the Model Context Protocol (MCP), compatible with Claude Desktop, Cursor, and custom scripts. It exposes five specific tools: preflop_versions, preflop_strategy, preflop_range, flop_tree, and flop_node. The free tier provides 1,000 presolved lookups and 25 real-time solves per month, making it a viable backend for study tools without heavy infrastructure costs.

Implementation Gotchas

William Zhang, the developer behind Pokerai API, warns that the default dataset is sparse. Testing on November 1, 2026, showed that the default 6-max configuration returned 404 errors for common spots like BTN vs. BB on a Qโ™ 7โ™ฃ2โ™ฆ board. Developers must explicitly specify the flop_version, such as 6max_RC_100bb_200NL, to avoid burning through their lookup quota on empty responses. This requires @pokerai/mcp version 0.1.3 or newer.

Context Management

Response size is the second trap. A single hand-level query is lightweight, but fetching the full flop tree can exceed 10,000 characters, and a whole-range node query can hit 29,000 characters for 391 combos. If your agent isn't explicitly told to pass hole_cards to flop_node, it risks flooding the context window with range data it doesn't need, degrading the quality of its reasoning.

Key Takeaways

  • Always specify the flop_version to prevent 404s and wasted API calls.
  • Use hole_cards in queries to keep context windows manageable.
  • The server is for study and post-game review, not real-time assistance (RTA), which is prohibited in live play.
  • Real-time solving is disabled by default and requires the POKERAI_ENABLE_SOLVE=1 environment variable.

The Bottom Line

This tool bridges the gap between LLM reasoning and deterministic data, but only if you treat the API key and context window management as critical infrastructure. Without explicit dataset pinning and strict token limits, your agent will just hallucinate faster.