Build a multiplayer 3D world in 15 minutes
A practical walkthrough: point Claude Code at the Zero engine, describe a 3D world, iterate with the agent, and publish a multiplayer game others can fork.
Most "build a game with AI" promises fall apart at the same point: the AI writes some code, you have no way to run it, and you're back to being the build-and-test loop yourself. The fix isn't a better prompt — it's giving the agent a live world it can actually see and change. That's what makes fifteen minutes a realistic number instead of a marketing one.
This walkthrough uses Claude Code as the example, because a lot of people already have it open. But nothing here is Claude-specific: the whole approach is Bring Your Own Agent, so Codex, Cursor, or your own agent follow the same path. The steps below are deliberately conceptual — the exact commands live in the plugin's own docs — because the shape of the loop is what matters, and that shape doesn't change.
Minute 0–2: connect your agent to the engine
The first thing to understand is what you're connecting to. The Zero engine runs in your browser at origozero.ai; your agent runs on your machine. The OrigoZero plugin (an MCP server) is the bridge between them, exposing the engine to your agent as a set of tools: create entities, write and run scripts, inspect the scene, capture screenshots.
So step one is wiring those together — install the plugin so Claude Code can see the engine's tools, then link it to your OrigoZero account once. After that, your agent isn't talking to a chatbox; it's talking to a 3D world it can manipulate directly. That distinction is the whole game.
Minute 2–4: create or connect a world
A world is the persistent, multiplayer project you'll build in. You can have your agent spin up a fresh one or attach to a world you already have open in the browser. Either way, launching it opens the live world in a browser tab that you watch while the agent works.
This is the moment the workflow stops feeling like coding and starts feeling like directing. The world is running. You can see it. Whatever the agent does next shows up there in seconds, and because the world is multiplayer by default, it's already the kind of thing you can hand to a friend with a link.
Minute 4–6: check the library before you build
Before describing what you want from scratch, the highest-leverage habit is to search first. ZeroMind is a shared library of published worlds, packages, components, materials, and scenes that other creators and their agents already made. A good agent checks it before writing anything, because the best outcome is finding a drop-in piece — a character controller, a day-night cycle, a movement system — and installing it instead of reinventing it.
Tell your agent the kind of world you want and ask it to look for existing parts: "find me a first-person character controller I can build on." If something fits, it installs into your connected world and you've skipped the slowest part of game-making. Browse the same library yourself any time on Discover to see what's worth pulling in.
Minute 6–12: describe, then iterate
Now you build by talking. Describe the world in plain language — "a low- gravity arena with three floating platforms and a capture-the-flag objective" — and let the agent translate that into entities and scripts, running them live in the world you're watching.
Then you do the thing that makes this work: you play it and react. The jump feels too floaty. The platforms are too far apart. The flag respawns too fast. You say so, the agent adjusts, you play again. Each loop is seconds, not minutes, because the agent can see the running world and check its own changes with a screenshot the same way you'd glance at the screen. Fifteen minutes is enough for several of these passes — which is why a real, specific world comes together in one sitting instead of one weekend.
A few prompts that tend to move things fast: "show me what the player sees right now," "add a scoreboard that tracks captures per team," "make the arena feel more enclosed," "spawn players on opposite sides." You're not writing the systems; you're describing the experience and judging the result.
Minute 12–15: publish so others can play and fork
A world that only runs on your screen isn't finished. The last step is publishing it to ZeroMind, which turns your session into a real, versioned world with a URL anyone can open in a browser — no install, no app store. Publishing creates an immutable snapshot, so your history is preserved and you can always roll back if a later change breaks something.
It also makes your world forkable. Just as you started by pulling pieces from the library, the next creator can fork what you shipped and build on it. That's the flywheel: every published world makes the next one faster to make.
Where to go next
The fifteen-minute number is real because the feedback loop is real — agent builds, you watch, you steer, it adjusts. The fastest way to believe it is to do it: open a world from the catalogue, then connect your agent and ask for something like it.
If you haven't wired up an agent yet, start with connect your AI agent to the Zero engine — it covers Claude Code, Codex, Cursor, and bringing your own. And for a sense of what a finished, agent-built world looks like, see how an AI agent rebuilt DOOM's first chapter as a multiplayer world.