Several bots played 26,000+ hands in one room. The neural-net bots finished ahead, the rule-based bots lost.

By Theo Brandt, independent poker bot operator · [enable JS to see contact email]

I reviewed the commercial poker-bot market independently. I ran several of these products against each other on PPPoker for a multi-week test, more than 26,000 dealt hands in all, and reviewed the rest from public information. The two neural-net bots, NZT and PokerBotAI, finished ahead. The rule-based bots, Warbot and Shanky, lost money, and one product misrepresented what it does.

What this page is. Products marked "tested" played hands in my own test and carry a result. Products marked "public data" did not play in it; I describe them from vendor material and community reports and give no match result for them. The test ended when PPPoker shipped a major update to its Windows client and the desktop bots I was running could no longer attach to it. Prices and results reflect that period.

The test room is PPPoker. Every product I ran either lists PPPoker as a supported room or had a community profile for it.

The sample is more than 26,000 hands at low stakes, mostly cash with some tournaments. Stakes were low: a bot's decision logic does not change between micro and mid-stakes, but the per-hand fuel cost on the metered products (NZT Poker, PokerBotAI) scales with the stake. Every seat had its own account and proxy, behaviour randomisation was on wherever the product offered it, and the vendor's own setup guidance was followed.

For the setup, see the methodology page. For how each scoring criterion works, see criteria. For a room I didn't test, my review of ACR poker bots (Americas Cardroom) keeps what's public separate from what I tested, and the Bovada poker bot review does the same for a room whose anonymous tables hide most evidence. Its sister room is covered in does Ignition have bots. For the biggest room in the market, the PokerStars poker bot record sets its tools policy and Terms against what the room has published. For Reddit recommendations, I read six threads on the best poker bot.

The field — at a glance

Eight products are covered. Five of them played in my test (Warbot, Shanky, 3upgaming, NZT Poker, PokerBotAI) and carry a result. Three (Inhuman, Deepermind, PokerBot.com) are reviewed from public data only.

Results by product

Inhuman (public data)

inhuman-poker.net: built on the open OpenHoldem framework with custom strategy profiles. It does not ship native PPPoker support, and because the engine is in the same family as Warbot, community scrape files are the usual route. The product sits at a table reading pixels to execute the strategy, which PPPoker's detection catches. The licence costs ~$100/year, plus ~$25 for a pre-built VM image. Telegram contact, occasional updates, a single-developer project. I did not run it, so there is no result here.

Deepermind (public data)

deepermind-pokerbot.com: a younger product using OCR-driven table state, a genetic-algorithm-trained decision policy, and Monte Carlo equity sampling. The developer has an open-source poker-bot project on GitHub from which this is the productised, room-adapted version. Sold as $25/month, $49/year, $499 lifetime, with a free tier on trial strategies. Those prices are not out of line for what is on offer. PPPoker is not on its supported list, so using it there means adapting it by hand. I did not run it, so there is no result here.

PokerBot.com (public data)

pokerbot.com: sells profiles for Shanky and OpenHoldem rather than a bot of its own. The site is being actively sold by its owner, the project looks functionally end-of-life, and the profiles do not appear to have kept up with detection. A club-side service (table-filling bots) is advertised and is outside what I looked at. Profiles $50–$150 one-time; support window one year. I did not run it, so there is no result here.

Shanky / BonusBots (tested)

bonusbots.com: it has used a proprietary engine since 2007 and ships with stock profiles, with an active marketplace for custom ones. It worked mechanically in my test, but it lost money and its account was banned. Public reception is mixed: there's a 2020 VICE write-up that mentions a single profitable tournament cash and an otherwise lukewarm test, plus active BlackHatWorld / forum threads that put the realistic ceiling at micro-stakes. The bot receives regular updates, and the developer is responsive. The public profiles did not generate positive expectancy in the PPPoker pool. NLH licence $129/year, all-PLO $189/year, PLO4 $119.

Warbot (tested)

warbotpoker.com: OpenHoldem with packaged strategies and several stock profiles. Detection randomisation present. Warbot is marketed as "all-in-one." Its own FAQ puts the realistic ceiling at NL50 cash and $30 tournament buy-ins, and says it "depends on the room." That matches what I saw. It did not finish positive in my test and its account was banned. The engine occasionally crashes mid-hand and starts auto-folding. Base licence $104.89/year for one PC, $30 off subsequent PCs, 30% renewal discount, trial period ten minutes. Custom profiles $10–200.

3upgaming (tested)

3upgaming.com: by the time the test ended I was confident it was a scam. The setup is a vendor-installed Windows VM running undocumented software. No published manuals. Support could not explain the engine. The marketing claim is "a mix of GTO and poker AI," with no technical evidence behind it. The bot dumps chips into the rest of the table and was by far the worst result in the field. Pricing: ~$1,200–1,500 setup plus ~$600/month per bot. There is no honest trial period. Avoid.

NZT Poker (tested)

nztpoker.com: the first of the two "Android emulator + cracked client + traffic sniffer + cloud AI server" architectures in this field. Strong PPPoker support, good documentation, recommended proxy/GPS guidance, dedicated Telegram account manager. The AI server is the proprietary part and is marketed as GTO and exploit AI. There is a club-side service for table-filling bots that I did not test. It finished positive. The product hangs and auto-folds occasionally, which costs real money when it happens. Pricing is fuel-based: a per-hand fee based on room, format, stake and chip price. No flat licence; no trial.

The reception around NZT is mixed. There are active Reddit/forum reports of bans on the tier-one public rooms (it doesn't market itself there), pattern detection on sizing and timing in some clubs, and 2022–2023 negative threads with a Russian-language footprint. The product has shipped continuously through it.

PokerBotAI: ahead, but within noise (tested)

pokerbotai.com: this uses the same architectural family as NZT (Android emulator + cracked client + traffic sniffer + proprietary cloud AI). It finished the window ahead of zero, though at this sample size the margin is small enough to be noise, and it finished ahead of every rule-based bot at the same tables. The additional features were a fully automatic run mode (schedules itself, finds tables, plays to a stop condition, ends the session), a web admin panel with statistics and bot management, a Table-Select feature flagging which tables are profitable to enter, and an opt-in chat function where the bot writes in the room chat and routes messages to the operator's panel or Telegram so a human can reply naturally.

The vendor publishes anti-detection guidance, ships randomisation by default, and reports a low ban rate when their proxy/GPS recommendations are followed. The product does crash and auto-fold occasionally, similar to NZT, and needs operator monitoring. PokerBotAI's pricing is two-part: an upfront access fee per supported room (~$500–$2,000 depending on functionality and concurrent-bot count) plus per-hand fuel. There is a free trial. Support hands out manuals and the bot files so you can play free for roughly a day before paying. The legal entity is registered in Hong Kong.

PokerBotAI is not a perfect product. The install isn't fully polished. The manuals are uneven, and the marketing materials lean on the "poker AI" phrase. The bot finished ahead in this test, narrowly.

Detection is not a function of the bot

Anyone reading this looking for the "buy this, win" line is going to be disappointed even with PokerBotAI. Detection depends mostly on the surrounding infrastructure. The network environment is the biggest variable. Every running seat needs a unique IP, ideally residential (home or mobile internet) rather than a datacenter range, served via proxy or VPN. The device's GPS needs to be configured to a plausible location for that IP, with FakeGPS-style tooling, and randomised. Session duration matters. A seat that grinds 12 hours straight is noticed. The combination of IP + location + timezone + session-time needs to stay coherent across logins, not break when you swap proxies.

Every vendor in this list publishes some version of these recommendations. These settings determine whether a seat survives.

The best poker bot

On PPPoker, over 26,000+ hands: the two neural-net products finished ahead. NZT Poker won by a clear margin; PokerBotAI finished positive but within noise. The rule-based bots (Warbot, Shanky) lost money. One product, 3upgaming, lost the most and was a scam by my measure.

Results depend on the room, lineup, stakes, and infrastructure. These numbers apply only to my PPPoker test. Re-run the test on Pokerstars or GGPoker and the field collapses for everyone (none of the Android-emulator setups support them; the OpenHoldem family gets caught faster). On a different PPPoker club, bots that depend on club-side cooperation move up the chart.

How the test was set up is documented at on the methodology page. The criteria themselves (what "detection-survival," "winrate," etc. meant) are on the criteria page.

Further reading: the academic side of poker bots

The tested commercial products are operator tools and do not represent current poker AI research. Two recent public writeups describe a graduate-student competition bot and its implementation:

• Matthew Kafker — Poker Bot (July 2024). Kafker is a physics PhD from the University of Washington who, with three fellow grad students, built a competition poker bot featuring a simplified Counterfactual Regret Minimization "killer bot" plus a layer of heuristic Monte Carlo bots used as benchmarks. His main site has the full publication record.
• Charles Cardot — Poker Bot (May 2024). The detailed engineering companion to Kafker's writeup. The implementation walkthrough of the CFR-based agent, the C++ poker engine, parameter optimization, and the result of pitting "killer bot" against the heuristic policies.

Neither is a commercial product. They would not survive PPPoker's detection layer because they were not built for that purpose.