This one's for the people who came here actually curious, not looking for a download. There's a real history of open-source poker AI work, and it's worth knowing about. It just doesn't give you what you'd hope it gives you.
Libratus was the Carnegie Mellon heads-up no-limit hold'em AI that beat four top human pros in a 2017 challenge match. It used counterfactual regret minimization plus endgame solving and ran on a supercomputer. The code was never released, and the team has been clear about why — they didn't want to hand a real-money exploitation tool to the world. Papers, yes. Binaries, no.
Pluribus came next, also from CMU plus Facebook AI Research, and it beat humans at six-player no-limit hold'em — a much harder problem than heads-up. The Science paper was published. The training method was described. The code, again, was not released. People have written clones based on the paper, and some of them work on toy poker variants, but none of them is "Pluribus you can download and use." That's a deliberate choice by the authors, not an oversight.
Search GitHub for "poker bot" and you get hundreds of repos. Most fall into three groups. First, learning projects — somebody's CFR implementation that plays a simplified game against itself, useful for understanding the math, useless against humans. Second, bot frameworks for IRC poker or local-only games like the old "neuron poker" environments — fun for research, not connected to any real room. Third, abandoned 2014-era hopium — empty scaffolds where someone planned to build a real bot, got two commits in, and disappeared.
You'll occasionally see a repo claim to play on a real room. Check the commit history. It's always either a screenshot grabber that never had a decision engine, or a wrapper around someone else's commercial bot that requires a license key the README "forgot to include."
Yes, but it's useful for understanding poker AI, not for using one. The papers describe how modern bots think — abstraction, equilibrium approximation, subgame solving, blueprint strategies, depth-limited search. If you read three or four of them, you'll have a better mental model of poker than most regulars. That's a genuine free thing the open-source world gives you. Just not the thing you searched for.
Because a bot that plays heads-up against a research server isn't a bot that plays nine-handed at a real room. You'd need client integration — vision or memory reading. You'd need anti-detection — input timing, mouse curves, session patterns. You'd need maintenance — updating every time the room ships a patch. The academic work skips all of that on purpose. The people who fill the gap don't publish, because publishing kills the edge in a week.
That's the honest answer to "why isn't there a free open-source online poker bot." The math is public. The integration, the stealth, and the ongoing maintenance — those have always been private, and probably always will be.
Curious how the private side works? You can ask.