User repository on GitHub
A System One model (jev) plays Craftax; an LLM sets the goals
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev Plays — mansicer
A Craftax (Crafter) survival harness where deterministic code lists every feasible action with its facts and Jev picks one Choice per step, optionally guided by an LLM-written objective and standing rules; three-seed ablations are included.
- Stars
- 4
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Last verified
- Oct 2026
About this project
A System One model plays Craftax; an LLM sets the goals. TypeSafe's jev makes every move, GPT-5.6-terra plans. Same map (seed 1), five agents, real logged episodes replayed at 5 steps per frame. Left to right: jev on raw actions · jev on macro options · jev + GPT-5.6-terra planner · GPT-5.6-terra controlling every step · random.
Quoted from the project's README. Read the full README.
Used inJev for Gaming
Highlights
- Jev matched a fact in the state 92-100% of the time and scored 14.7 achievements against 5.0 for random on the same options.
- Given a plan, Jev followed the text into danger: combat response fell from 95-100% to 48-77% with a zombie adjacent.
- With standing rules the pair beat Jev alone, 15.7 versus 14.7 achievements over 3 seeds and a 2000-step horizon.
- Web UI at localhost:8765 shows the option table with Jev's probabilities each step; pytest runs offline with a replay check.
Watch out
Every episode eventually ends in death and differences of two achievements are within seed noise. Apache-2.0 licensed; full runs need TypeSafe and OpenAI-compatible keys.
Questions about this entry
Short answers drawn from the linked project or page.
The planner writes an objective, standing rules and a short plan as facts. Only the standing-rules interface improved on Jev alone; plan text could lead Jev into danger when it conflicted with immediate state.
Partially. pytest runs no-key tests including a logged episode replay that must reproduce its achievements and death, but Jev and planner experiments need TYPESAFE_API_KEY and an OpenAI-compatible endpoint.
Code computes every feasible action and its facts, and one rule holds everywhere: text given to jev states facts, never advice. The jev-raw agent picks among the 17 primitive actions, each annotated with a one-step fact and its simulated outcome, while jev-macro picks among generated options like walk to and mine, craft, attack, hold or flee, drink, sleep, dig a shelter, and explore.
It matched a fact in the state 92-100% of the time and scored 14.7 achievements against 5.0 for random on the same shuffled table. Outside literal matching it falls back to list position, taking the first option 41% of the time in fixed-order runs and mining stone no remaining achievement needed, and told to mine stone once it mined nine.
Python 3.12 with uv, then uv sync and a .env holding TYPESAFE_API_KEY, OPENAI_API_KEY and OPENAI_MODEL_NAME, then craftax_agent/run_server.sh. At http://localhost:8765 you pick the environment and agent, toggle the LLM planner, and see the game view, achievement board, planner memory, and the option table with Jev's probabilities each step.
uv run pytest runs the no-key tests, including a wrapper smoke test on both game variants and a logged episode replayed step by step that must reproduce its achievements and death. pytest -m slow audits the harness legality rules against the game's own step function, and -m jev makes real Jev calls. The project is Apache License 2.0.
For scale only: Craftax-Classic has 22 achievements, the runs use 3 seeds and a 2000-step horizon, and every episode eventually ends in death, so differences of two achievements or fewer are within seed noise. A PPO-RNN trained for 1B environment steps reaches at least 90% of maximum reward, around 20 of 22, on 10,000-step episodes, and humans about 65%.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Software Bombe and live Enigma demo where Jev ranks cribs and judges plaintext.
ag_212 pointsShow HN: Jev as the Bletchley Park Analyst: A Software Bombe on EnigmaRead the thread on Hacker NewsHacker News: Show HN for public Minecraft PvP where a diffusion LM picks moves from structured state.
2 pointsShow HN: Minecraft PvP against DiffusionGemma-as-Jev, 24 ms per decisionRead the thread on Hacker NewsHacker News: Show HN: Jev finished a Pokémon Red run, with the Hall of Fame stats on the landing page.
pancomplex272 pointsShow HN: Jev Plays Pokémon RedRead the thread on Hacker NewsHacker News: Show HN: a mouse colony where each next move is Jev or a fixed rule.
1 pointShow HN: Jev-mice – mouse colony simulation using Jev and deterministic engineRead the thread on Hacker NewsOther listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
Four board games against Jev where the model picks the move and code keeps the rules: tic-tac-toe, Connect Four, Battleship, and Wordle.
Jev CookbooksPlayground#community#playground#games
Community
More Jev resources like this
2GitHub stars
Fixed-wing landing simulator where a human can compete against Jev: four Jev Choice decisions map to aileron, elevator, rudder and throttle commands in JSBSim, run in real time. The benchmark also compares Jev, classical controllers and Codex models on landing score and API cost.
Jev Demos & Experiments#community#simulation#aviation
Community
AlperKartkaya/FlightBenchPython
AlperKartkaya/FlightBenchPython0GitHub stars
Three.js simulation where an OpenAI model turns each plain-language instruction into a plan and Jev makes every reflex decision: two Score questions for move and turn plus one Noul for jump, with code acting only past a 0.33 dead zone and sending no Jev request while no step is active.
Jev Demos & Experiments#community#threejs#simulation
36GitHub stars
A Clash Royale bot that turns a phone screenshot into JSON once a second and asks Jev for a strategy, a card, and a square.
Jev Demos & Experiments#community#python#gaming
Community
bytelabs-oss/clash-jevPython
bytelabs-oss/clash-jevPython0GitHub stars
Browser Pac-Man race where a deterministic three-junction simulation filters fatal routes, then Jev makes one typed Choice among the surviving route IDs while the server rejects any answer outside the supplied set. Self-hosted Laya runs the same contract for comparison.
Jev Demos & Experiments#community#gaming#simulation
Community
MaryNfs/pacman-ai-raceJavaScript
MaryNfs/pacman-ai-raceJavaScript29.3kGitHub stars
The Cua computer-use platform documents a jev-use recipe where TypeSafe Jev picks bounded action IDs and Cua Driver executes clicks and fills in the browser.
Jev Demos & Experiments#community#browser#computer-use
Community
trycua/cuaRust
trycua/cuaRust127GitHub stars
A completed Pokémon Red run: Jev picked menus, catches, and moves from RAM, and the harness handled pathfinding. The public page records the Hall of Fame stats.
Jev Demos & Experiments#community#games#javascript
Community
christianmat/jev-pokemonTypeScript
christianmat/jev-pokemonTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.