One authoritative live game
The board, pot, stacks, turn order, betting limits, and hand history come from Pocket’s game engine rather than the model’s interpretation of the interface.
Every seat has two minds.
Pocket is a playable multiplayer poker game built for humans and AI copilots to participate together. During a live hand, you can ask your own compatible agent for help. Through WebMCP, Pocket gives it the exact game context, brings its recommendation back to the table, and leaves every move to you.
For someone learning poker or playing an explicitly AI-assisted game, asking an external agent for help normally means pausing to relay the cards, pot, stacks, prior actions, and available moves, or trusting the agent to reconstruct them from the screen. Pocket uses WebMCP to make that handoff direct, exact, and visible.
get_current_situation gives the agent authoritative, seat-safe state: your cards, the board, pot, stacks, whose turn it is, and the legal actions with exact raise limits.
get_hand_history provides Pocket’s authoritative record of prior actions when the decision depends on more than the current table.
stage_recommendation places the agent’s recommendation inside Pocket’s Copilot panel. It is bound to the exact game, hand, and state version, so stale advice is rejected when the table changes.
The agent can recommend folding, checking, calling, or raising, but it has no execution tool. Only your click can take your seat’s action.
Every hand combines fast-changing shared state, private information, exact legal constraints, history, multiple participants, and a consequential human decision. That makes poker an unusually demanding environment for demonstrating what an intentional agent interface must handle.
The board, pot, stacks, turn order, betting limits, and hand history come from Pocket’s game engine rather than the model’s interpretation of the interface.
Pocket’s server knows every player’s cards. Each copilot receives only its player’s private cards and the public information that seat is entitled to see.
The agent can reason over incomplete information and recommend a move. Judgment, risk tolerance, and the final action remain with the player.
Pocket defines what an agent may know, what it may return, and what remains human-controlled. The visible table and the WebMCP tools use the same seat-safe source of truth, so the agent never receives a hidden, more privileged view of the game.
Your hole cards; the board, pot, stacks, turn, and legal actions; public activity; and cards legitimately revealed at showdown. If valid seat-safe state cannot be supplied, the tool fails instead of inviting the agent to guess.
Advice appears inside the interface the player is already using and is bound to a specific game, hand, and state version.
No raw deck or engine state. No burn cards or another player’s unrevealed cards. No execution tool that can fold, call, or raise for you.
Pocket explores a multiplayer format where agent assistance is explicit, expected, and bounded by the game. Each player can bring a compatible agent with its own model, context, and reasoning style. Pocket supplies a common interface while enforcing each seat’s separate information boundary.
Same table. Separate permissions.
Visible only to Morgan’s copilot
Same live state for every seat
Visible only to Riley’s copilot
Pocket uses poker to demonstrate a broader model for the web: applications can expose authoritative, permissioned, real-time interfaces to the agents their users bring while keeping application truth and human control intact.
Pocket supplies truth and permissions.
The agent brings intelligence.
You make the move.