A game loop asks the same question every tick: what do I do now? A decision model answers with a typed choice, such as a move, a lane or a target, so the game can apply it directly. Because the choice is typed, there is no text to parse and no invalid move to repair. The examples in this hub run on Jev and other decision models, and Laya appears in several. Examples include anxkhn/JevPlaysPokemon, which plays Generation 3 Pokemon through Showdown, jev-gomoku by mizchi, and jev-doom-agent by lukaske. Others are benchmarks or toys, like ably-labs/jev-pong, where the ball moves one step per model decision. Before you copy one, check what state the model sees, how legal moves are listed, how quickly a decision has to come back, and whether game code or the model owns the world state. Games make a good test bed because the rules are clear and a wrong move is cheap, so you can try a decision model without real stakes.
79 entries
Showing 1 to 48 of 79
A Minecraft agent that pairs an Astra planner with Jev as the real-time controller, with native run recording, tested routes and run verification.
536 starsFastest run 8m 43s, -40%Decisions used 131 Jev calls
Agent that plays Super Mario Bros. by feeding structured emulator state to Jev and acting on the returned controller action.
Laya, running locally on Apple Silicon via MLX, and Jev, over TypeSafe's hosted API, play Chrome's T-Rex game side by side on a shared obstacle course, with response times, survival streaks and crash replays recorded.
Laya, running locally, and Jev, over TypeSafe's hosted API, face off in a snake race and a Mortal-Kombat-style fight, both answering the same typed questions on the same input; from a Prompt Engineer 48 video.
Harness that lets Jev control the original StarCraft shareware through synthetic keyboard and mouse input, recording action probabilities for each decision.
Snake auto-played by Jev: one Choice per game tick, with legal moves and board facts computed in code before the call.
Two AI models play real-time Tetris against each other: Jev faces Claude Haiku, Gemini Flash, or an open-weight typed-decision model, with the same piece sequence and clock for both sides.
A Three.js first-person flight game where a glowing fly chases goblins in a garden; Jev chooses its heading and dive about five times per second.
Three runnable experiments: support-message triage producing six independent judgments, a driving simulation where Jev picks lane and target speed against traffic and signs, and an alternative Fable implementation. Shows questions and probability distributions for each call.
Browser-native Doom agent using a Chocolate Doom WebAssembly runtime, structured spatial state, composable AI controls and live decision telemetry.
Browser stealth game where Jev makes typed guard judgments while deterministic game code owns all world state.
Unmodified Laya clears Super Mario Bros. 3 World 1-1 on an NVIDIA DGX Spark through Stable Retro NES emulation, with 16.79ms median inference and every frame's decision recorded for replay.
Jev plays Generation 3 Pokemon through Showdown and a real FireRed ROM, with no screenshots involved. Each turn the host reads the battle out of RAM and hands Jev the legal moves and switches; the author points a deliberately underlevelled party at the Elite Four.
Autonomous Pokemon Red agent on a PyBoy emulator. Jev reads a compact text snapshot of the game state, answers parallel yes/no questions each turn, and deterministic code turns the answers into button presses. Jev has no vision and no history, so the harness holds the short-term memory.
Toy archipelago where Jev makes every in-flight call: divert or continue, declare an emergency, give way, hold, and who lands first when two aircraft want the same runway. Each aircraft knows only itself and nearby traffic, asks four batched questions about every 1.6 s, and gets answers in roughly 150 ms.
A browser pong game whose left paddle is decided by one Laya choice question per frame, 18.7ms p50 on Metal, while the right paddle plays fixed arithmetic for comparison.
A desktop Pygame Snake game with HUMAN, Laya AI and greedy-baseline modes, showing a live decision dashboard with real SDK probability bars, entropy confidence and per-move inference time.
Jev plays a browser table-tennis game in real time from structured telemetry, driving ordinary Chrome inputs and logging auditable decision evidence.
Experimental full-game harness where Jev plays the original Civilization II in a browser, displaying live action probabilities.
Jev plays Minesweeper, picking each move as a typed decision instead of generated text.
Pong where the ball moves one segment per model decision, so a lane's speed is that model's latency and nothing else. Four lanes replay a recorded run with Jev, Gemini 3.8 Flash, Claude Haiku 4.5 and GPT-5.6 Sol on the right paddle, every player and agent on an Ably channel.
Desktop Snake where Jev assesses the board and Python combines the answers into a move, one API call per game tick. One recorded run reached 29 foods in 461 ticks before trapping itself; the agent spots immediate hazards but does not plan a route.
Minecraft Java player controlled by Jev, with live decision display, a Canadian flag building task, and a side-by-side dashboard.
An interactive Chicago traffic simulator comparing three signal policies on the same scenario: fixed timing, adaptive queue-based signals, and a citywide policy from Jev.
Browser fighting game with 21 satirical world-leader fighters, where the AI opponent's moves run on the Laya decision model client-side via WebGPU.
An open-weight Laya model plays Breakout live on Windows, choosing each paddle move as a single choice question at about 70 decisions per second on an RTX 5080, with no planner or safety shield.
Adaptive model predictive control guides and vetoes a Laya decision model piloting a lunar lander. Laya alone lands 0/90; with MPC's veto it lands 90/90 through a 60% engine power loss.
Screenshot-to-action game agent for Hordes.io: Qwen plans from screenshots, Laya makes the typed real-time decision, and Apple Vision OCR reads the screen, all running locally on Apple Silicon.
A 322M-parameter Laya checkpoint plays a Pac-Man-inspired game on Windows with CUDA, using typed decisions and live probability output.
A guide-directed leveling agent for a private World of Warcraft: The Burning Crusade 2.4.3 server, built so it gets cheaper to run the longer it plays.
chalkychalk42Openchalkychalk42/jev on github.com
Atari-style lunar lander simulation with a CUDA-trained Laya checkpoint that lands 30/30 held-out flights locally on Apple MLX, plus a browser flight recorder.
A browser maze game where every move is a typed choice answered by a locally run Laya model. Correct moves are marked green and overruled ones red, testing whether Laya's confidence score predicts whether its own move was right.
kspviswaOpenChakravyuha on github.comA penalty shootout game where Jev decides the goalkeeper's dive, built on Render Workflows and Render Postgres, with example implementations in Python and TypeScript.
Benchmarks Jev against Luna, Haiku and Gemini at choosing the next action in games with explicit states, legal actions and measurable outcomes. Each game is a state machine where code owns the rules and the model only picks among the options it is given, with every model reached through Vercel AI Gateway.
hegargarciaOpenhegargarcia/jev-playground on github.comMoonBit client for the Jev API plus a Jev-vs-Jev gomoku match with per-move timing logs. Write-up at zenn.dev/mizchi/articles/jev-plays-gomoku.
Minecraft mod in which Jev attempts to finish the game from scratch with no scripted route.
Hardel-DWOpenjev.mods on github.com
The Chromium T-Rex Runner game with Jev at the controls: it picks jump, duck or keep running for each new obstacle and chooses a short or full jump profile. The browser keeps speed-aware frame timing, collision geometry and input execution, and manual play still works without a key.
Hackathon project: 200 people versus one agent behind a typed System One action gate that learns from each breach, evaluated with W&B Weave and hardened by a defender loop. Playable at vnmoorthy.github.io/siege.
Wikiracing solver that uses Jev to rank outgoing links by fewest expected remaining hops, scoring up to 128 candidate links per request with up to five concurrent requests, shown live in the terminal.
A browser chess coach on Cloudflare where Stockfish owns the truth about the position and Jev makes the coaching judgment after every move.
A local Gomoku game for Apple Silicon where Laya-MLX plays white, choosing among up to six rule-filtered candidate moves with a real choice inference call, and the browser shows the model's per-candidate probabilities.
smile-magicOpenYizhao (Laya-MLX Gomoku) on github.com
A video demo rebuilds a Tesla Full Self-Driving-style simulator controlled by Jev, built in under an hour.
A video demo of Jev playing Super Mario Bros. in real time, using its fast inference and structured output for moment-to-moment decisions.
A video demo of Jev V13 playing 5+0 blitz chess against Fable 5.1 and GPT-6 Astra, one move per API call, answering in about 2.6 seconds per move. Fable was ahead on material but lost on time; Astra mated Jev in 18 moves.
AI/ML API (@aimlapi)OpenBlitz chess against frontier models (X) on x.com559k viewsMove latency ~2.6 svs Astra mated in 18 moves
A demo of Jev playing Subway Surfers at superhuman speed, including 50 games running at once, for under a cent total.
A video demo of Jev controlling all 4 characters in a Smash Bros match against itself, picking each next move within a fraction of a second; the match used over 22 million tokens for a few cents.
Mau Baron (@maubaron)OpenJev playing all 4 characters in a Smash Bros match (X) on x.com311k viewsTokens used 22M+
A video demo of a Rubik's Cube solving itself with Jev: a standard beginner method is coded in, and at each step Jev picks which case the cube is in while code checks the pick. It took 94 moves rather than the 22-move optimal solution, using about 4 seconds of model time.
Paolo Rosson (@redp314)OpenA Rubik's Cube that solves itself (X) on x.com207k viewsMoves used 94 vs 22Model time ~4s total
A browser fighting game where you play against Jev, TypeSafe's decision model, from the keyboard on desktop or mobile.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.
The loop is: read the game state, list the legal actions, ask a choice question, apply the move. anxkhn/JevPlaysPokemon reads the battle out of memory and hands Jev the options each turn, with no screenshots. komikat's jev-bfs uses scores instead, ranking Wikipedia links by expected remaining hops. Search-style games work with scores because the model can rank many options in one request instead of making a move at a time.
Keeping game code in charge works well. heist-one by AbdelStark has Jev make typed guard judgments while deterministic game code owns all world state. A Rubik's Cube solver on X uses a coded beginner method and asks Jev only which case the cube is in. This split keeps the rules in code, where you can test them. The model supplies only the judgment, such as how risky a move is.
Latency is the limit. Every decision is a network call unless the model runs locally, so real-time games either slow the game to match, as jev-pong does, or batch decisions. Check whether a project runs a hosted model or a local one before you plan for frame rate. Batching helps when many agents share a model, as one entry shows with many agents running in parallel.
Yes. Entries here have Jev playing Pokemon battles, chess, gomoku, Minesweeper, Pong and Doom. Each game gives the model the current state and the legal options, and the model returns one choice.
Some entries do, but each decision is a model call, so speed depends on the model and the network. Several projects run a turn-based or slowed-down game for that reason. Local models like Laya remove the network step.
A typed choice cannot be an invalid move or malformed text, which suits a game loop. An LLM can plan further ahead. Some entries here benchmark the two against each other.
