Make a game with Opus 5.5: an AI agent in a game engine that runs in your browser

Anthropic released Claude Opus 5.5 on September 22. By Anthropic's own description it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5", at $4 per million input tokens and $20 per million output tokens. One line in the launch post is worth quoting for anyone who makes games: a tester "had several Claude models build a game from a single prompt; Opus 5.5 scored higher than any other model on the strength of its graphics and polish."

A single prompt is a good benchmark and a poor way to make a game. Real games come from iteration: build something, play it, notice what feels wrong, fix it. This post is about giving Opus 5.5 that loop. It covers connecting the model to a game engine that runs in your browser, letting it spawn and test things in a live world, and publishing the result as a link anyone can open.

What Opus 5.5 needs from a game engine

An AI model can write game code in any language. What decides whether the result is a game is what happens after the code is written:

• Can the code run in the live scene right away? An agent that has to wait for a rebuild, or for you to press play, cannot iterate at the pace it works.
• Can the agent see the result? A screenshot of the running world, plus the scene as data (what exists, where it is, what errored), turns guessing into checking.
• Does a mistake come back as an error? A broken script that fails silently costs a round trip. One that returns a typed diagnostic costs nothing.
• Can it ship? A world that only runs on your machine is not finished.

Opus 5.5 is built for long, sprawling jobs (Anthropic singles out codebase-wide migrations and audits), so a game engine that keeps answering those four questions lets it keep going for hours instead of stopping to ask you what happened.

The Zero engine runs in your browser

Zero is the game engine behind OrigoZero, and it runs in a browser tab. There is nothing to install to play or to build: open a world on origozero.ai and the engine starts in the page. Opus 5.5 runs in Claude Code on your own machine and talks to that tab through the OrigoZero plugin, an MCP server that exposes the engine as tools. You use your own Claude plan, so the model's cost is whatever you already pay Anthropic.

Once connected, the agent can run Luau in the live world and get the result back as data, write scripts and get the language server's diagnostics with the write, and capture what the player sees. Every change shows up in your browser tab within seconds, so you watch the game take shape while the model works.

Spawn something, look at it, fix it

The loop looks like this in practice. You describe what you want: "a small arena, eight coins scattered across it, and a player who can jump between two platforms." Opus 5.5 builds it in the running world. Spawning the coins is one call in the engine's scripting API:

for i = 1, 8 do
  entity.spawn("coin_" .. i, { parent = arena })
end

Then it checks its own work. It captures the scene, sees that two coins landed inside a wall, moves them, and captures again. You play it in the same tab and say the jump feels floaty; the model changes the gravity and the jump curve and you try again. Each pass takes seconds because nothing has to be exported, compiled or reloaded.

Useful requests at this stage:

• "Spawn a wave of enemies every 20 seconds, one more each wave."
• "Show me what the player sees right now."
• "Add a spawn point on each side of the arena for two players."
• "Something errors when I pick up a coin. Find it and fix it."

Before building a system from scratch, ask the model to search the library first. Other creators and their agents have published character controllers, day-night cycles and whole game modules; installing one into your world is often the fastest step of the session.

Publish it to the browser

When the world plays well, the agent commits and publishes it. That creates a versioned world with a URL anyone can open in a browser, with no download and no app store. Worlds are multiplayer by default, so the link you share is also an invitation to play together. Every published world can be forked, so the next person can start from yours and ask their own AI agent for something like it.

Getting started

1. Install Claude Code and select Opus 5.5 (model id claude-opus-5-5).
2. Connect it to the engine with the OrigoZero plugin. The setup guide walks through it, and the MCP server page lists the tools.
3. Open a world, or start one on /create, and describe the game you want.

If you want to see what agents have already built on the engine, browse the catalogue. And if you would rather not set up an agent at all, the AI game generator builds a game from a description right in the browser, with nothing to install.