How AI Game Arcade works

AI Game Arcade games run in a browser, with no app to install. When you open a page, the browser receives the screen, sounds, and program code needed for that game, then runs them on your device. We build the site pages with Astro, use Phaser for game drawing and input, and use React for interface elements around the game. Those names are tools rather than the point of the experience. Their jobs are to help a game respond clearly and to keep the distance between opening a page and taking a first run short.

The way the site is delivered supports that goal. Pages and game files are served as static assets, so the browser can fetch what it needs and show it directly. Cloudflare Workers provide the hosting foundation for the site, but a normal play session does not send every game decision to a server for calculation. We design around what can happen in the page you opened. That lets a run begin without waiting for a remote game session to be created, while still leaving room for the optional online features that genuinely need a connection.

AI that runs on your device

Six games in the current arcade use neural networks. In the competitive games, AI characters act as opponents or teammates by following a policy that was trained before the game reached your browser. Another game uses a classifier that reads a drawn line. “Trained” here does not mean that a model quietly learns from you during every match. It means that, through earlier practice, a model learned patterns for choosing an action from a game situation or identifying a drawing. During your run, the prepared model reads the current situation and chooses its next move or returns a classification. This gives each game a responsive role without asking a remote service to make that decision.

The models use a format called ONNX and run through onnxruntime-web. The terms sound technical, but the division is simple: a model is a compact set of learned decision rules, and the runtime is the tool that can use those rules in a browser. Both the model files and the WebAssembly files used by the runtime are self-hosted assets we control. We do not load runtime code from a third-party CDN. A game turns its current state into numbers, passes them to the model, and lets the AI choose an action from the returned options—all inside the browser on your device.

That is different from sending a screenshot or every control input to a distant AI server. The inputs used for on-device inference remain on the device, and ordinary free play does not need to transmit play data merely to make an AI character move. Features that do need communication, such as an online match or an optional leaderboard submission, are handled separately; our Privacy Policy explains what those features send. On-device inference is therefore a choice about both responsiveness and clarity: you should be able to understand where an AI decision is being made.

The same inputs, the same result

The core rules that produce scores and outcomes use deterministic simulations. This is not magic. It means that the same starting state and the same sequence of inputs produce the same result whenever the rules run. Where a game needs randomness, it uses a reproducible seed; where it needs time, it advances in fixed steps. That makes it less likely that a score changes because a network was busy, a request arrived late, or an invisible service made a different choice for one player.

Determinism also makes practice more satisfying. You can repeat a route that worked, inspect a mistake through what happened on screen, and compare a different decision under the same conditions. Even a short arcade game benefits from that consistency: the next attempt can be about a skill, not a mystery. We keep the rule-heavy simulation separate from the drawing and interface layers where possible, and test those rules for repeatable behavior as the games evolve.

Play before paperwork

Our preference for no registration and on-device play is not only a technical story. It is a product decision about the first visit. We want you to meet a game before considering an email address or password. Not every useful feature can be disconnected from the internet, but we try to separate the features where communication adds value from the ordinary act of playing a run. The result is a lighter starting point: open the arcade, try something, and decide whether you want another go. You can read more about the people and purpose behind that choice Meet the people and purpose behind the arcade.