Is Jev generating free-form flight commands in this demo?

No. Jev answers six typed questions per request: one Choice per stick axis plus two Booleans for landing-commit and motor-cut. Code builds the situation report and maps probability distributions to deflections.

How is this different from prompting a general model to fly?

A general model would emit prose you must parse. Here Jev returns calibrated probabilities in about 200 ms, so thresholds, deflections, and safety limits stay in deterministic code.

What setup is needed before Jev can take the sticks?

Install dependencies, set TYPESAFE_AI_API_KEY, run the dev server, open the printed local address, and press J to hand over control. Press R to reset the mission.

Can I replay an interesting flight and see what it cost?

Yes. Every refresh builds a new city, the HUD shows the map seed, appending the seed to the URL replays it, and the landing banner reports total tokens and decisions used.

What are the controls, and can I take over mid-flight?

J toggles autopilot and R resets the mission. W and S are throttle, A and D yaw, the up and down arrows pitch forward and back, the left and right arrows roll, and a gamepad works too. Touching any stick while Jev is flying hands control straight back to you.

What does the mission consist of, and what does the HUD report?

Take off from pad A, cross a randomly generated city and land on pad B. The city regenerates on every refresh, the HUD shows the map seed, and the landing banner totals the tokens and decisions Jev used for the flight. Nothing in the demo replays a fixed city unless you pass the seed in the URL.

Which SDK does it call through, and what is the license?

It is TypeScript on Three.js, and the model call goes through the AI SDK's @ai-sdk/typesafe-ai provider rather than a hand-rolled client. Setup is npm install, write TYPESAFE_AI_API_KEY into a .env file, npm run dev, then open the printed local address.

CommunityJev Demos & Experiments

Jev Autopilot

Three.js drone simulator where TypeSafe Jev flies a generated city from pad A to pad B through the AI SDK provider, with code turning telemetry into a situation report and probabilities into stick deflections.

See highlights

Open github.com

Stars: 13
Language: TypeScript
Published by: Community
Author: arielweinberger
Last verified: Oct 2026

arielweinberger/jev-autopilot

User repository on GitHub

View on GitHub

This demo uses Jev from TypeSafe AI to autonomously fly a drone in a random city from point A to point B, avoiding obstacles along the way. A trip costs $0.01.

Forks: 3
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A 3D drone simulator built with Three.js, flown autonomously by TypeSafe's Jev model through the AI SDK's @ai-sdk/typesafe-ai provider. The mission: take off from pad A, cross a randomly generated city, land on pad B. Open and press J to hand the sticks to Jev. A gamepad works too. Touching any stick while Jev is flying gives you control back. Every refresh generates a new city. The HUD shows the map seed, and /?seed= replays one.

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: TypeScript

Highlights

Press J to hand control to Jev; touching any stick returns manual control immediately.
Each decision is one request with four Choice stick axes plus landing-commit and motor-cut Booleans.
Cities regenerate on every refresh, with the HUD seed and URL seed parameter replaying a layout.
Landing banner reports the total tokens and decision count used for the flight.

Watch out

Flight quality depends on live model answers and the safety limits kept in code; manual sticks override Jev at any time. 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-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.