Is the model actually playing, or just labelling what I press?

Actually playing, within a narrow contract. Jev picks jump, duck, or keep_running for each new obstacle and a short or full jump profile when jumping. Frame timing, collision geometry, duck duration and input execution stay in the browser, so a late answer costs accuracy but cannot corrupt the simulation.

Do I need a TypeSafe key just to play it manually?

No. The API key is only needed for AI mode. Without one the manual game remains fully usable and the UI says AI mode is unavailable, which makes it a reasonable way to try the extraction before wiring up any credentials.

Is there any evidence the Jev controller scores better than a human?

No numbers are published. The README documents the maneuver contract and shows a demo recording, but no score, distance, or win-rate comparison against manual play or any other controller, so treat the AI mode as a demonstration of the decision shape rather than a benchmarked player.

How much work is it to drop this into a React or Next.js app?

Six documented steps: copy the resources folder, copy the images folder into public, import trex.css, add the offline-resources div from index.html, then write a script that imports the Runner class from offline.js and constructs it with your container element. You may need to adjust container styling, and the CSS import is not optional.

How do I run the Jev T-Rex demo locally?

Node.js 20 or newer; npm install, copy .env.example to .env and set TYPESAFE_API_KEY, then node --env-file=.env server.mjs and open 127.0.0.1:3000; without a key manual mode still works and the UI reports AI mode is unavailable.

What is Slow Speed Mode in this T-Rex build?

It is an accessibility setting that reduces acceleration from 0.001 to 0.0005, lowers the max speed from 13 to 9, widens obstacle gaps and adds audio cues; enable it by clicking the T-Rex before starting the game.

CommunityJev Demos & Experiments

Jev T Rex Runner

Q: What controls does Jev have over the T-Rex game?

For each new obstacle it chooses one semantic maneuver, jump, duck or keep_running, and picks a short or full jump profile when jumping; the browser keeps speed-aware frame timing, collision geometry, adaptive duck duration and input execution.

The Chromium offline T-Rex Runner extracted into a standalone ES6 Runner class with no circular dependencies, plus an optional Jev controller that picks a maneuver and jump profile per obstacle.

See highlights

Open github.com

Stars: 7
Language: JavaScript
License: BSD-3-CLAUSE
Published by: Community
Author: joshlarsen
Last verified: Oct 2026

joshlarsen/jev-t-rex-runner

User repository on GitHub

View on GitHub

Chrome dino game played by Typesafe AI Jev model

Forks: 4
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

This copy can be played manually or controlled by TypeSafe's Jev model. Jev chooses one semantic maneuver (jump, duck, or keeprunning) for each new obstacle and selects a short or full jump profile when jumping. The browser keeps ownership of speed-aware frame timing, collision geometry, adaptive duck duration, and input execution. Requirements: Node.js 20 or newer and a TypeSafe API key.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Jev chooses one semantic maneuver per new obstacle: jump, duck, or keep_running, and picks a short or full jump profile when jumping.
The browser owns frame timing, collision geometry, adaptive duck duration and input execution; the model never touches pixels or events.
Refactored from the migrated Chromium tree into an ES6 class with the singleton pattern and circular dependencies removed, tagged with change-reason comments.
Node.js 20+ and a TypeSafe key are needed for AI mode only; without one, manual play stays fully usable and the UI reports AI mode is unavailable.

Watch out

The Jev part is a controller, not an agent driving physics: it returns a maneuver and a jump profile, and the game decides timing and geometry, so there is no published score showing it plays better than a person. The key is read only by the local Node server launched with node --env-file=.env and never reaches browser code. The extracted source is brittle in two documented ways: sprite edits must keep the existing layout and coordinates, and at least one stylesheet must be imported or the game logic crashes when it injects CSS into the first one. The Contributing section links point at a different repository. 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.

More Jev resources like this

laya-vs-jev-arena

Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.