User repository on GitHub
Visual Jev lab for multiple games and emulator platforms
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev Games
A visual lab where Jev plays Mario, Kung Fu and Doom, with a plugin split so new games and emulators do not require rewriting the loop. Code owns buttons, hold duration and Mario landing interrupts; Step and Scripted keep the API bill down.
- Stars
- 2
- Language
- Python
- Published by
- Community
- Author
- shantanugoel
- Last verified
- Oct 2026
About this project
A visual lab where Jev plays multiple games. This is the next step after mario-jev: the same Jev controller style, a Zero-shot Lab UI, and a plugin split so new games and emulator platforms can be added without rewriting the loop. The model stays Jev. Code owns buttons, hold duration, and (for Mario) landing interrupts. Python 3.13 via uv, frontend via bun.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Three titles on separate platforms: Mario through gym-super-mario-bros, Kung Fu via nes-py, Doom via ViZDoom.
- Adding a game means implementing create_env, observe, questions, compose, scripted and advance, then registering it.
- Tape replays recorded buttons on the emulator, and the Scripted policy avoids spending calls while you debug.
- Headless without the UI: uv run jev-games play --game doom --policy scripted --decisions 20.
Watch out
Doom is not guaranteed to be Doom: with roms/doom1.wad or doom.wad present you get the original E1M1 Hangar, and otherwise ViZDoom falls back to Freedoom, which the README states is a different first map - so a run on the fallback is not comparable. Kung Fu requires a legally obtained NES ROM in roms/ or KUNGFU_NES set, since Atari is no longer used for it, and Mario needs no ROM at all. The README says TYPESAFE_API_KEY is already present in .env in that checkout, which means you should check what is tracked before copying that setup and adding your own key. Toolchain is specific: Python 3.13 via uv plus bun for the frontend, two terminals, and the UI on 127.0.0.1:5173 with the API serving frontend/dist on 127.0.0.1:8000 or JEV_GAMES_PORT after a build. Nothing about model behaviour is reported - no question sets, no thresholds, no confidence figures, no token cost per decision and no scores or outcomes for any of the three titles, so you cannot tell from this page how well it plays or what a run costs. The Game interface signature is the documented extension point; the questions themselves live in the per-title modules and are not shown. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
For Kung Fu, yes - put a legally obtained roms/kungfu.nes in place or set KUNGFU_NES, and the RAM map comes from the NES original. Mario needs nothing, it runs through gym-super-mario-bros. Doom uses the original E1M1 Hangar if roms/doom1.wad (shareware) or doom.wad is present and otherwise falls back to ViZDoom's Freedoom Phase 1, which the README warns is a different first map - so check which one you actually got before comparing runs.
Use the Scripted policy instead of Jev, and single-step rather than run. The README is blunt that every Jev decision is a paid TypeSafe request and recommends Step or a short Run, preferring Scripted until the picture and buttons look right. Step/F10 takes exactly one decision, and the Tape feature replays recorded buttons on the emulator, so inspecting a completed run costs nothing further.
Not if a platform already exists. You add src/jev_games/games/<name>/game.py implementing Game with create_env, observe, questions, compose, scripted and advance, then register it in src/jev_games/catalog.py. For another NES title that is make_nes_rom_env with the ROM path and action button lists, reading memory off env.unwrapped.ram and writing questions the way Mario does. A new console such as SNES or Game Boy is a new platform module with create_env, frame and memory access, and games should not import emulator libraries directly if the platform can own that.
Python 3.13 via uv and bun for the frontend, then uv sync and bun install plus bun run build. The lab needs two terminals: uv run jev-games lab and bun run dev in frontend. Open http://127.0.0.1:5173, and after a frontend build the API also serves frontend/dist at http://127.0.0.1:8000 or whatever JEV_GAMES_PORT is set to. Checks are uv run pytest and uv run ruff check src tests.
Use Python 3.13 via uv and bun for the frontend. Run uv sync, then cd frontend and run bun install and bun run build, and start two terminals: uv run jev-games lab and cd frontend and bun run dev. Open http://127.0.0.1:5173; after the build the API also serves frontend/dist at http://127.0.0.1:8000 or JEV_GAMES_PORT.
Run or F5 starts Jev or the scripted baseline, Step or F10 takes one decision, and Reset or R restarts the episode. Cartridges in the header switch between Mario, Kung-Fu, and Doom, History or Tape lists attempts, and Replay re-runs recorded buttons on the emulator with Esc returning to live.
The repo runs uv run pytest for the suite and uv run ruff check src tests for lint. The layout separates engine.py (the run, step, and reset worker with WebSocket snapshots), platforms/ for emulator adapters, games/ for per-title observations and Jev policies, and frontend/ for the visual lab UI.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPython42GitHub stars
Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
JoshuaSP/open-jevPython
JoshuaSP/open-jevPythonFrom 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.