ClubGG Bot Developer FAQ

Both, with the proportion heavily weighted toward scam. A working club-app bot is an engineering project on the order of a year of focused work for a small team — solver pipeline, opponent modelling, mobile UI automation, behavioural shaping. The serious products do exist but are not sold openly on Telegram to retail buyers; they run inside specific clubs as embedded operations and are not portable to the buyer's environment. Almost everything sold under "ClubGG bot" on public channels is either a repackaged generic solver, a credential-stealing front end, or remote-access malware. The economically rational behaviour for someone who has built a working club-app bot is not to sell it.

More distributed, less centralised. On GGPoker the detection is a single network-layer stack that has to make all the calls. On ClubGG the detection is split between the NSUS network layer (same signal families, behavioural fingerprinting and statistical play-pattern analysis and account-graph models) and the per-club layer (agents and club owners with direct view of seat-level behaviour and operational tools to restrict play). The club layer is what catches most single bots at small and mid stakes; the network layer catches farming and statistical outliers. Neither layer is sufficient alone, and a bot that survives one can still be caught by the other.

Short-term yes, long-term no. The mechanics work — you control seat assignments, you decide who sits with whom, and the agent-layer enforcement is in your hands. The failure mode is settlement-side. Your club develops a reputation among the surrounding settlement partners, your softer players lose money faster than they expect to lose it and exit, and the broader ecosystem refuses to take your action. The network layer notices a club whose Diamond flows have an unusual pattern and applies its own scrutiny. The pattern that survives is small-volume embedded bot play inside legitimate clubs run by inattentive owners, not full-control farming.

Yes, in the technical sense. The platform-side anti-cheat is less mature than what GGPoker runs against high-volume regulars on the public site; the UI is mobile-only which simplifies the input-automation surface to a single Android target; and the population baseline is wider, which gives behaviourally shaped output more room to hide. The catch is the agent layer, which a GGPoker bot does not have to deal with. A technically simpler product gets caught faster in operations than a more sophisticated GGPoker bot does on the network layer.

Small-and-mid clubs at USD-equivalent $0.10/0.25 to $1/2 NLH 6-max are the sweet spot. The regulars are exploitable, the agent layer is inattentive, and the rake structure is more favourable than a public operator because clubs compete on rake to attract grinders. PLO at the same stakes works for teams that have the solver compute. High-stakes club games are essentially closed — attentive ownership, vetted players, and small samples make external bots structurally non-viable.

Two reasons. First, regulatory: public real-money poker is restricted in much of Asia (Korea, Japan, China) and Latam, while private club games operate in a grey zone that the offline-settle model fits cleanly. Second, the agent-and-club economy already exists in those regions — pre-existing networks of poker rooms, backers, and stake-sharing relationships translate naturally to a club-app model with offline settlement. The same product launched into Western Europe or the regulated US states would have no comparable existing channel; it would have to build the agent layer from scratch.

Probably yes, but no public confirmation. NSUS Group runs a single security organisation with shared infrastructure. The plausible internal architecture is a shared abuse graph that ingests device fingerprints, payment methods, KYC documents, and behavioural fingerprints from both products. There are anecdotal reports of cross-product enforcement when fingerprints overlap; the operational assumption from a bot operator's perspective should be that any signal generated on either product is visible to the other. The detection note walks through the structural argument.

Varies sharply by club. Owners of small clubs running on tight margins sometimes tolerate small-volume bot play because the bot is paying rake and is not eating into the soft-player population fast enough to matter. Owners of established mid and high-stakes clubs actively hunt bots — both to protect their player base and because settlement-side partners do not want to take action from clubs known to harbour them. The pattern that works for a bot operator is finding the inattentive owner; the pattern that does not work is assuming all clubs treat bots the same way.

Technically no per the terms of service, but enforcement is variable. The platform does not have first-party HUD support equivalent to GGPoker's Smart HUD. Third-party HUDs that scrape from the rendered client exist; some clubs treat them as routine and others ban them on sight. OCR-based HUDs that read from a screen recording rather than from the client process are harder for the platform to detect through client telemetry. The club-side enforcement is more consequential than the network-side rule.

Structurally similar — three-handed hyper-turbo with a random multiplier on the buy-in, so the math has the same shape as Spin and Gold. The differences are that the Diamond Race buy-in is denominated in Diamonds rather than cash, that the multiplier distribution is club-configurable in some setups, and that the field has more variance because the population is smaller. From a bot-architecture perspective the engine that handles Spin and Gold maps almost directly to Diamond Race; the only meaningful difference is the multiplier-aware ICM table needs to be regenerated for the local multiplier distribution.

Same architectural requirement as on desktop, harder to implement. The phone shares one CPU and one display, so the input-automation layer has to swap focus between tables, and any cross-table action-timing correlation is detectable. Production systems use a preemptive per-table decision queue: a tough decision on table A surrenders the CPU when easier decisions are pending on table B, and each table's behavioural-timing distribution is sampled independently. Three tables is comfortable on mid-range Android hardware; four is the practical ceiling because focus-switching latency starts to introduce its own fingerprint.

On a modern Android device, around 30-80ms for solver-table lookup, 5-15ms for the opponent-model update, 15-40ms for accessibility-tree inspection of the rendered client. Total compute budget around 100-150ms per action. The visible action latency is the behaviourally-shaped sample on top of that — typically 600ms for a snap-fold up to several seconds for a deliberation. Compute budget is comfortable; the binding constraint is power consumption over a long session and the accessibility-service stability across ClubGG client updates.

Roughly every two to three months, sometimes more often after a UI redesign. The break is almost always in the accessibility-tree layer or the rendered-pixel layer that the input automation depends on, not in the solver or opponent-model stack. Maintenance overhead per client update averages a few engineer-days. Teams that ship bots commercially treat this as a fixed ongoing cost and design their architecture to keep the screen-scrape layer isolated and small.

Against the weakest population segment, possibly, but not durably. Heuristic strategies (push-fold charts plus basic continuation-betting rules) generated positive expected value against play-money populations in 2018-2020. On ClubGG in 2026 the small-and-mid stakes regulars are aware of solver-derived theory and a pure-heuristic bot loses to them over thousands of hands. The minimum competitive engine is solver-anchored, even at small stakes. Heuristic-only software is the easiest tier to spot from the network-layer statistical analysis as well, because its play-pattern is distributionally far from population.

Much faster than on anonymous-table operators, because identity is stable across sessions. With per-session updates accumulated over a few hundred hands of joint play, you get VPIP, PFR, 3-bet by position, and fold-to-cbet by board texture to within reasonable confidence. The trade-off is that the same stable identity makes you visible to the agent layer over the same horizon. The bot that builds a confident model of every regular at its table is also the bot the club owner has had time to notice.

The baseline is the hard part. On a public operator with anonymous tables, repeat co-occurrence is itself a signal. On a club app where regulars see each other every session, the baseline expected co-occurrence is high, and the right reference is per-club joint behaviour rather than per-network. Showdown-correlation models and joint-aggression indices are still useful, but the population priors have to be per-club and constructed from small samples. The pseudocode sketch in the detection note walks through the structure.

Action-timing distributions are easy to spoof once you understand the distributional shape — a state-conditional log-normal sampler with right-tailed distraction noise produces output that matches the population shape. Touch-coordinate clustering is medium difficulty; you need a believable model of where humans actually tap on a button. Hardware-level signals — accelerometer drift, gyroscope readings, CPU clock variance, battery discharge curve — are the hardest because the device has to actually be in a believable physical situation, and the alternatives (sensor injection, emulator) leave their own fingerprints.

Same signal families with a different population baseline. The ClubGG population skews more Asia-mobile and more Android than the GGPoker public-site population, which means the bet-sizing distributions, the action-timing distributions, and the touch-coordinate clusters that count as "normal" for the classifier are different. A behaviourally-shaped bot tuned on GGPoker population data needs retuning for the ClubGG baseline. Transfer learning between the two populations is genuinely open as a research question.

Two paths. The network-layer path is quieter and longer: a quiet flag, a soft restriction (withdrawal or transfer limits drop, KYC re-verification requested), an account closure if the signal persists. The club-layer path is faster and more arbitrary: a club owner notices unusual behaviour, restricts the seat, refuses settlement on the held balance. The fastest enforcements tend to be club-side; the largest financial consequences tend to be network-side, but only when an account has been quietly running long enough to accumulate the signal.

The chat link at the bottom of every page is the right starting point. Implementation questions, observations on cross-NSUS enforcement you have seen, club-side detection patterns, references to academic work on adversarial classification or club-app detection — all of it is welcome. The chat is read by the Poker Bot AI team. Product-pitch messages are auto-archived.