Build1 publisher3 min readPublished
Genex wraps Claude Code and Codex in an MIT-licensed harness for Three.js games
Genex released a free, MIT-licensed desktop app that points existing coding agents at Three.js browser games. Each game stays an ordinary codebase on the developer's machine, so dropping the harness later still leaves a working project.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Users hand separate models the planning, building and judging jobs, and visual judges inspect what the builder agents produce.
- Genex connects to Claude through Claude Code, to ChatGPT through Codex, and to local models through Ollama or Bonsai, according to its documentation.
- Only macOS and Linux downloads are offered for now, and the product page still lists Windows as coming soon.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Teams with Unity or Unreal projects cannot use this release, since Genex lists those integrations as planned and currently drives only Three.js browser code.
- cost Model spend lands on the user's own Claude or ChatGPT plan or local hardware, so subscription usage limits cap how much a multi-agent session can do.
- decision Picking Genex over GDevelop or Rosebud AI means running coding agents and a code repository yourself, which suits developers already comfortable working in code.
A session starts with a plain description of a game or a feature. Genex then splits the work: one model plans, other agents build, and visual judges inspect the result [3]. Users can pick a different model for each of those roles [9]. While the run is underway, the desktop app shows the workers changing code, checking builds and taking feedback from the human [4].
The model plumbing is what I would check first. Genex says the app runs on existing Claude or ChatGPT subscriptions and does not need API keys [10]. That makes setup cheap. It also moves the ceiling to the plan's usage limits. Runtimewire notes that those limits, or the cost of local compute, stay with the user [16].
The project format is the strongest choice Genex made. Genex calls the software a customizable harness for game development and says it is not a new engine [2]. Its public GitHub repository bears that out: each project is a normal Three.js codebase written by the coding agent, with Genex adding tools for assets, multiplayer and publishing [12]. Files stay on the user's computer [5]. A finished game exports as a static web bundle or publishes through Genex [6]. Walk away from the harness and the repo is still a Three.js project a developer can open and keep working on [22]. GDevelop, by comparison, pairs its own engine with an AI agent that edits projects inside GDevelop's editor, according to Runtimewire [17]. Rosebud AI takes a plain-language prompt to a playable game in the browser [17].
The self-editing mode is experimental and off unless the user enables it [14]. I would not ship an agent that rewrites its own instructions with any other default. When it is on, prompts, skills and tools live in a Git repository the app can edit [13]. The app takes a snapshot before a change, and a blind judge compares the result [13]. A version that fails to start is rolled back [13]. Keeping that history in Git means a person can read what the agent changed as a diff afterward [23]. Startup is the only automatic rollback condition in Genex's description. A revised prompt that boots cleanly but makes the builder worse has to be caught by the blind judge's comparison [21].
Assets come from Blender for local 3D modeling and Meshy for generated models and characters, and the app accepts custom plugins and skills [11]. Per Genex's product page, the desktop app is free under the MIT license, and the paid component is optional asset generation through a Genex plugin [15].
Scope is narrow for now. The app targets Three.js games that run in a browser [6]. Unity and Unreal integrations are planned but not in this release [8]. The product page still lists Windows as coming soon [7]. In the launch thread, the makers said Codex is already part of how they build games, and that they wanted a harness of their own for combining models, tools and workflows [18]. Runtimewire frames the open test as whether the multi-agent loop can dependably build and verify useful Three.js games while the developer keeps control of the project [19].
What to watch
- Whether Genex ships the planned Unity and Unreal integrations, which would test the harness against engines that keep projects in their own editors and formats.
- The Windows build, still marked coming soon on the product page.
- Any published success rate for plan-build-judge runs, or user reports on how fast multi-agent sessions exhaust Claude or ChatGPT subscription limits.