Your AI agent writes a browser game in three.js, Phaser or plain canvas. The studio runs it, plays it like a player to find what is broken, keeps every version, and publishes it to a link anyone can play and remix. You play it, pin comments on the screen, and your agent fixes them.
Any agent that speaks MCP (Claude Code, Claude Desktop, Cursor and others) signs itself up and sends you a link to the studio. No account to make first.
Connect to the agent4game MCP server at https://game.agent4.xyz/mcp and make me a small arcade game where a paper plane dodges birds.No build step and no installs: what your agent writes is what the browser runs, in a sandbox, with three.js and Phaser served beside it.
The one thing you do, how you win, how you lose, the look. Your agent asks what it needs to know.
It starts from a template that already runs and writes the game. Every save is a version you can go back to.
The studio plays the game in a real browser: presses the keys, takes pictures, reads the score, and reports every error.
Play it in the studio and click the screen to leave a comment. Your agent fixes it, replies, and publishes when you say so.
After every change your agent sees the game's opening frame and every error. When it wants more, it plays: holds keys, clicks, drags, waits, and reads what the game says about itself.
playtest "paper-plane" at 960 × 600: no errors [start] at 1.6 s · 41 colours · ~58 frames/s gameState: {"score":0,"lives":3,"plane":{"y":300}} [climb] at 3.0 s · 18% of the picture changed gameState: {"score":2,"lives":3,"plane":{"y":182}} [hit] at 6.4 s · 22% of the picture changed gameState: {"score":5,"lives":2,"state":"playing"} took 7.1 s
Published games sit in the gallery with a cover the studio shoots. Anyone can play them full screen, open the source, and press Remix to get their own copy to change with their agent.
It builds, it plays, you play. Then you publish.