What is mario-decisions-agent?

A Super Mario Bros agent that plays by looking at the screen one screenshot at a time. Its entire brain is a remote vision model on the other end of an API, with no memory reading and no hidden game state.

How does mario-decisions-agent use Jev?

The README lists TypeSafe's Jev (jev-preview) as an alternative to the OpenAI Decisions API (gpt-6-luna), which is what the recorded runs used. Reproducing the runs needs your own API key.

How does the agent decide what to press?

It sends a frozen screenshot and asks a single question about it: here's a picture of the screen, per the README. The model returns the button, and pixels in, button out is the whole loop.

Does the agent cheat by reading game memory?

No. The README stresses that most game bots read memory to find Mario's position and speed, while this one cannot: it sees only what you would see in a screenshot.

What platform does it run on?

The badges list Python 3.12 and macOS plus Linux, and the status is marked experimental, per the README. The digest excerpt does not include the setup or run commands.

What do the demo runs show?

A GIF of the model playing World 1-1 from raw screenshots, slowed down for viewing. The README notes the real model is remote, so reproducing the runs needs your own API key.

Who built mario-decisions-agent and under what license?

The GitHub user parjanyaacoder publishes it under the MIT license. The repository was created 2026-10-14 and last pushed 2026-10-16, with 4 stars and no forks when the sweep ran.

Jev directory

CommunityJev Demos & Experiments

Mario Decisions Agent

A Super Mario Bros agent whose only brain is a remote vision model on the other side of an API: screenshots in, button out, and no memory peeking. It runs on the OpenAI Decisions API by default and supports TypeSafe's Jev (jev-preview) as an alternative.

See highlights Open github.com

Stars: 4
Language: Python
License: MIT
Reading time: 2 min
Published by: Community
Author: parjanyaacoder
Last verified: Oct 2026

parjanyaacoder/mario-decisions-agent

User repository on GitHub

View on GitHub

A Super Mario Bros agent whose only brain is the OpenAI Decisions API (POST /v1/decisions, gpt-6-luna). Screenshots in, button out — no RAM peeking.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity github mario gaming

Works forJev for Realtime control

Used inJev for Gaming

Languages: Python
Platforms: macOS
Capabilities: Open sourceRequires API key

Highlights

No memory hacks and no hidden game state: pixels in, button out, per the README.
Its brain is a remote vision model: the OpenAI Decisions API (gpt-6-luna) or TypeSafe's Jev (jev-preview).
The only question the model is ever asked is about a picture of the screen, according to the README.
Experimental status, Python 3.12, macOS and Linux, MIT licensed, last pushed 2026-10-16.
Reproducing the shown runs needs your own API key, per the README.

Watch out

README reviewed 2026-10-17; the digest excerpt ends at the question the model receives, so setup, emulator requirements, decision cadence and any cost or accuracy figures for the Jev path are not documented.

Questions about this entry

Short answers drawn from the linked project or page.

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

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

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 AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai