User repository on GitHub
TypeSafe's Jev model plays NetHack 5.0: code lists the legal moves, Jev picks one, no LLM in the loop
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Demos & Experiments
Nethack Jev
TypeSafe Jev plays NetHack 5.0 with no LLM in the loop: each turn the code lists the legal moves, Jev picks one as a choice question, and a small motor translates the answer into keystrokes, with a live dashboard and Hardfought support.
- Stars
- 3
- Language
- Python
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- statico
- Last verified
- Oct 2026
Used inJev for Gaming
Highlights
- 570 games through 2026-10-02, all dwarven Valkyries: deepest Dlvl 11, longest game 34,423 turns, highest score 18,708 points.
- The NoModel baseline, which always takes the first option without any Jev call, also reached Dlvl 11, matching the deepest Jev game.
- The dashboard at 127.0.0.1:8770 shows the live screen, each option with Jev's probability, decision history and money spent, and lets you pause or step turns.
- Default spend limit is $5, changeable with JEV_BUDGET_USD; every game writes its decisions to runs/<id>/decisions.jsonl.
- The structure follows olivier-motium's Jev Plays Doom: code lists options, Jev picks one, a motor does the keystrokes.
Quickstart
bash
scripts/build-nethack.sh
/opt/homebrew/bin/python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
(cd web && npm install && npm run build)
echo 'JEV_API_KEY=...' > .env
.venv/bin/python -m jev.serverWatch out
Grounded in the truncated repository README and GitHub metadata reviewed on 2026-10-15; the README notes all code was written with Claude Code and Claude Opus 5.5. Game status numbers are the author's own runs through 2026-10-02, and the README was cut off in the watcher section.
Questions about this entry
Short answers drawn from the linked project or page.
The code reads the game screen, builds the list of legal options for the turn, sends it to Jev as a choice question, and a small motor performs the keystrokes for the chosen option. No LLM is in the loop.
Through 2026-10-02: 570 games, no ascension, deepest level Dlvl 11 with twelve games reaching Dlvl 10, longest game 34,423 turns, and a highest score of 18,708 points before dying in Sokoban while praying.
The README's NoModel test always takes the first option and never calls Jev; two games still reached Dlvl 11, equal to the deepest Jev game, so the model did not beat that baseline on depth.
You need Python 3, Node.js and a Jev API key. Build NetHack with scripts/build-nethack.sh, install the Python requirements and the web dashboard, put JEV_API_KEY in .env, then run python -m jev.server and open 127.0.0.1:8770.
Yes. The dashboard shows the live screen, each option with its probability, the decision history and money spent, and lets you pause, step one turn, change the speed, start a new game or give Jev a standing order.
The default spend limit is $5, configurable via JEV_BUDGET_USD in .env. Each game writes its decisions to runs/<id>/decisions.jsonl, and the dashboard tracks the money spent as the game runs.
Yes, on Hardfought: add your username and password to .env, paste jev/nethackrc into your NetHack 5.0 options file once, then run python -m jev.server --hardfought. The bridge uses Hardfought's web terminal because some networks block port 22; SSH works with HARDFOUGHT_SSH=1.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Jev picks which Apple HIG guides go into the agent's context. NatashaTheRobot built a small AskHIG CLI and…
X: Three decision models, one game of Snake
X: Diogo Almeida with a16z on Jev
X: Luna plans, Jev plays poker
More Jev resources like this
An experiment where Jev plays Zork I: every turn Jericho supplies the valid actions and Jev answers one Choice of which command to type, with confidence bars, an anti-loop memory of the last 8 turns, a JSONL log and a French replay dashboard.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
resadan-dev/jev-zorkPython
resadan-dev/jev-zorkPythonA 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-jevPythonAn autonomous ViZDoom agent that decouples Doom's simulation from Jev SystemOne inference: the game runs at 35 ticks/sec while Jev is queried asynchronously at about 10 Hz for macro goals, targets, movement and firing decisions.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
amoghcreator/doom-jevPython
amoghcreator/doom-jevPythonTypeSafe's Jev plays the real Steam game Vampire Survivors - it picks the character, stage, every level-up and the walking direction four times a second, with a live decision dashboard. A BepInEx plugin and Python brain; native Linux only.
Jev Demos & Experiments#community#catalog-import#logicrw
A demo that walks a real Pokémon FireRed ROM: the code reads the overworld out of RAM, works out every reachable destination, and Jev picks where to go — and how to fight wild encounters — through typed decisions, not screenshots.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
daniel4x/jevemonHTML
daniel4x/jevemonHTMLTwo TypeSafe System One game loops in one Python repo: Dino-Jev drives Chrome's chrome://dino/ dinosaur and Krunker-Jev runs a local Krunker-style arena, both feeding structured state to Jev and logging decisions.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
mittal-parth/jev-experimentsPython
mittal-parth/jev-experimentsPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.