BetOnline Poker Bot Developer FAQ

Yes, an established market exists, mostly serving Western and Latin American buyers. What ships ranges from rebadged generic CFR-output bots with a thin BetOnline UI layer, through credible solver-anchored engines with active opponent modelling, down to outright credential-stealing software dressed up as a bot. The credible end of the market clusters at a small number of teams and a couple of hundred to a few thousand dollars for a license. The rest is sales-copy with no engineering behind it. If the landing page promises hole-card peeks, RNG prediction, or guaranteed winnings, the product is in the bottom category by definition.

Three reasons compound. The Chico Network client is older and less aggressive on client-side behavioural fingerprinting than the GGPoker build. Screen names are stable and HUDs run unobstructed, so opponent modelling has a long-horizon prior rather than needing to converge online in 80–150 hands. And the security team is smaller, so the offline play-pattern analysis runs at a slower cadence and the human-review queue moves more slowly. None of these are operator mistakes — they are budget allocations consistent with the room's size — but they add up to a softer environment for solver-anchored play.

Empirically, the sweet spot sits in the NL10 to NL100 range for 6-max NLH cash, where rake economics are reasonable (5–7% gross, partly offset by rakeback through the room's loyalty program) and the population is soft enough that a competent solver-anchored bot can earn 3–6 BB/100 against pool average. PLO viability tracks similarly at PLO10–PLO100. MTT and Jackpot SNG viability is meaningful at low buy-ins (under $25) because the field skill ceiling is low, but EV-per-hour drops because tournaments take longer than cash sessions. Anything above NL200 hits a thinner pool with sharper regs and lower EV per hand.

Same logical architecture (behavioural fingerprinting, statistical play-pattern, collusion graph, human review) at a meaningfully smaller budget. PokerStars and GGPoker run nightly offline analysis with substantial reviewer capacity and proactive escalation. BetOnline runs slower-cadence offline batches with a smaller review queue, and most bot bans here are triggered by external events — large withdrawals, formal complaints, forum pressure — rather than by the system firing on its own. The detection probability inside a quiet stretch is meaningfully lower; the detection probability inside a public-pressure-driven sweep is meaningfully higher. The full architecture is on the detection page.

BetOnline is licensed in Panama and operates from outside US jurisdiction. US federal law on online poker is complicated — the Unlawful Internet Gambling Enforcement Act of 2006 targets payment processing, not players, and the post-Black-Friday environment leaves a grey zone in most unregulated US states. Players in regulated states (Nevada, New Jersey, Pennsylvania, Michigan, others) are expected to use state-licensed operators. Players in unregulated states play at BetOnline at their own risk — the operator accepts US-IP traffic, but the player is responsible for their own legal exposure. This is not legal advice. Consult a licensed attorney in your jurisdiction.

Yes, and most accounts on this side of the network do. BetOnline accepts Bitcoin, Ethereum, Litecoin, Bitcoin Cash, USDT, USDC and a few smaller assets. Deposits and withdrawals are fast and cheap compared to bank-wire and money-order alternatives. The operational caveat for bot farms is that shared wallets are a join key for the operator's collusion graph layer — depositing to multiple accounts from one address is the fastest path to a graph-driven bust. Wallet hygiene is non-trivial: one wallet per account, no on-chain links between accounts, and patience on the deposit cadence.

Holdem Manager 3 and PokerTracker 4 work against the BetOnline client without operator obstruction. Hand histories save to disk in standard formats. Screen names are stable, so the long-horizon HUD data accumulates and stays useful — different from GGPoker, where anonymous tables and rotating names collapse the long-horizon attack. For bot authors, this means the opponent-modelling prior is essentially free; the marginal cost of building a multi-thousand-hand profile on most opponents is a hand-history dump rather than online convergence.

The 2014 cleanup was small in scale — a single ring, a handful of accounts, limited refunds — but visible because it followed weeks of TwoPlusTwo forum threads documenting the suspicious play. The 2018 action was larger, involving multiple rings caught in a coordinated sweep over several weeks, with account closures, balance confiscations, and refunds issued to affected players. Both actions were batched human-review decisions triggered by external pressure, not realtime detection firing. The detection page covers the empirical pattern in more depth.

Less critical than at GGPoker. The combination of stable screen names and tolerated HUDs means most opponents arrive at a session with a multi-thousand-hand prior already populated. The online model contributes session-level adjustments (tilt detection, fatigue patterns, table-image dynamics) but does not need to converge from a zero-information starting point. The interesting open question is how much marginal EV the online model adds on top of a long-horizon HUD prior in a soft pool — empirically the answer looks like 1–2 BB/100 at NL25–NL100, but the confidence interval is wide.

The compute budget is comfortable on any modern desktop or mid-range Android device. Solver-table lookup runs in 30–80ms, opponent-model update in 5–15ms, UI inspection on the rendered client in 10–30ms — total compute around 100–150ms per action. The user-perceptible latency is the behaviourally-shaped sample on top of that, typically 600ms to several seconds depending on decision difficulty. The constraint is not compute; it is consistent power draw over long sessions and avoiding telemetry signals from the device that look unlike a human-held device.

All three picked up post-Black-Friday US liquidity and all three have softer pools than European regulated rooms. ACR (Americas Cardroom) on the Winning Poker Network has a slightly larger and slightly tougher MTT pool, more aggressive anti-bot rhetoric publicly, and a separately-developed anti-RTA stance. Ignition (Bovada in the US) runs anonymous tables and is harder on the opponent-modelling side, much like a smaller GGPoker in that respect. BetOnline sits between them: stable IDs and tolerated HUDs like ACR, smaller security team than ACR, softer pool than Ignition. Cross-room arbitrage is real but each room has its own UI layer.

Android is the production target. The Chico Android client is older and the sandbox is permissive enough for accessibility-service-driven input automation with stable behavioural shaping. iOS is significantly harder because the iOS sandbox blocks the input-automation surface; the available alternatives (jailbreak, MDM exploitation) are unstable across OS versions. Most production poker AI against BetOnline runs Android-only as a result, which concentrates detection-side telemetry on Android-typical patterns — a small but useful population shift.

Hardware-level entropy is the hardest. Sensor data on mobile (accelerometer drift, gyroscope wobble during play, ambient-light variation) and CPU clock-frequency jitter on both mobile and desktop produce a signal that is expensive to fake. A device sitting in a cradle for ten hours has measurably different sensor readings from a phone being held. Programmatic perturbation of sensor inputs defeats the lazy version of this check; whether it defeats side-channel analysis at a sophisticated operator is unclear. BetOnline does not currently appear to weight these signals as heavily as GGPoker does, but the cost of starting to is low.

Bimodal. Most accounts run for months to years uncaught — quiet detection cadence, slow review queue, no triggering event. A minority are caught in batched human-review waves, triggered by external pressure or a single account-level event (large withdrawal, formal complaint, media coverage). The cycle from triggering event to confiscation typically runs 2 to 12 weeks once the queue picks up. The 2014 and 2018 cleanups are the public reference points for this pattern.

Partially. The trigger events for past sweeps have all been observable in advance — forum threads building for weeks, media coverage with named accounts, regulatory pressure spikes. A risk model that monitors public discussion and pulls accounts during pressure periods reduces caught-in-sweep risk meaningfully. It does not eliminate it: the operator can act on internal signals at any time. The right account-lifetime model has a baseline detection probability plus an event-conditional spike, not a single stationary number.

Three streets of betting, a board that has revealed most of its information, and 2–4 opponents whose ranges have narrowed materially. The game tree under reasonable bet-sizing abstraction blows up faster than any other street. MonkerSolver's depth-limited subgame solving handles the math for specific tree shapes; producing a general-purpose multiway turn policy that compiles to a mobile inference budget is genuinely open. Production approach is opportunistic — solved for high-frequency shapes, heuristic-plus-opponent-model for the rest.

Equity distributions flatten because each player has six two-card combinations rather than one. Preflop equity differences shrink, blockers matter more, and turn decisions carry more EV than river decisions on average. Solver outputs are larger and harder to compress, opponent-modelling features differ (3-bet frequency by hand class is less informative; reraise frequency given board texture is more), and the UI layer needs to read four hole cards instead of two. The engineering scaffolding is the same; the strategic content is meaningfully different.

Two compression directions, applied together. First, state abstraction — bucketing game states into equivalence classes on board texture, stack-to-pot ratio, position, and action history, reducing the lookup key from a full state to an integer index. Second, action abstraction — discretising bet sizes into a finite set (33%, 66%, 100%, 150% pot etc.) and solving for that abstraction. The combined compression typically gets the raw CFR output down four to five orders of magnitude with bounded EV loss compared to the uncompressed strategy.

UI automation reliability. The interesting work is in solvers and opponent models; the actual failure mode is almost always at the screen-scrape and action-emit layers. The Chico client ships fewer updates per year than GGPoker — three to four major releases versus four to six — but each release has a non-trivial chance of breaking an input-layer assumption. The engineering effort to keep the boring layer healthy across releases is higher than the effort spent on the interesting math, every year.

Read the long-form notes — the hacks article, the detection note, and the open questions on the homepage. If you have data or work in progress on any of the open threads (long-horizon HUD as Bayesian prior, bursty-enforcement risk modelling, multiway compression, LLM-augmented analysis), the chat is the place. Volume is low and the team reads everything; "can I buy your bot" messages are auto-archived.