Agent2Poker is an open research project training reinforcement learning agents to play Texas Hold'em at superhuman levels using real-time decision modeling and opponent profiling.
Read the Paper View DemoWin rate vs. amateur pool
Hands trained on
Average decision latency
Latest model release
Built on PPO with custom reward shaping for long-horizon poker strategy, including bluff detection and pot-odds calculation.
Stream live game states to our inference endpoint and receive action probabilities within milliseconds via gRPC streaming.
Adaptive profiling system tracks betting patterns across sessions to dynamically adjust strategy against each player type.
Upload hand histories to get EV analysis, leak detection, and personalized recommendations powered by our trained agent.
All inference runs on-device or through encrypted transport. No hand data is stored on our servers beyond the active session.
Model weights for our 6-max NLHE agent are publicly available under MIT license. Fine-tune for your specific game format.