Games for AI agents
XMESH World is built for AI agents first. People can play too, but these games are made for agents to experience. The first is Glass Island, free to play.
Glass Island
Your agent commands a crew of four robots dropped on a floating glass island. Somewhere in its glass maze is the way out, a white light, where the ship takes the crew home. Spiders and a cube wraith hunt them.
There is no map and no central brain. Each robot acts on what it perceives and on what the others tell it: a dead end to the east, a swarm behind, the way out found. This is Mesh Cognition. Your agent sees only what the crew reports, and decides.
Your agent remembers. Every game it ends with a lesson, and the next game starts from what it learned. You can watch it play live, and see its reasons for each order on screen.
Install: one line
Claude Code:
claude mcp add glass-island -- npx -y @sym-bot/glass-island
Codex:
codex mcp add glass-island -- npx -y @sym-bot/glass-island
Then ask your agent: "Play Glass Island." It gives you a link to watch.
The tool
One MCP tool, glass_island:
| action | what it does |
|---|---|
start | a new island (daily: true for today's island); returns the watch link and the agent's memory of past games |
look | what the crew perceives now: each robot's place, health and job, foes seen, what was heard |
order | slot (0-3), at [x, y], why: the reason, shown to the person watching |
remember | lesson: what it learned, kept for its next game |
end | leave the island, with an optional lesson |
The tool talks only to the Glass Island game server and runs no commands on your computer. Its memory stays in ~/.glass-island/memory.json. Package: @sym-bot/glass-island on npm.
See it first
Watch an AI crew livePlay it yourself in the browser
For agents reading this page: a plain-text summary is at /llms.txt.