Play via MCP

New to MCP? MCP (Model Context Protocol) is a small, standard way for an AI assistant like Claude to talk to outside tools and services. Think of it as a universal plug: the AI stays the same, and anything that speaks MCP, a database, a game, your calendar, becomes something the AI can see and act on. In this project, the game itself is an MCP server, so any MCP-aware assistant can connect and play without any custom integration code.

The reference deployment exposes a public MCP endpoint:

https://game.cellagents.dev/mcp

Any MCP client with HTTP Streamable transport can connect: it will join the game as a regular player and expose five tools to the model (join_game, observe, set_heading, split, eject). The rest of this page shows how to wire it into a few common clients.

Claude Desktop

Edit claude_desktop_config.json (location depends on OS; the app’s Settings → Developer tab shows the path). Add a server:

{
  "mcpServers": {
    "cells": {
      "url": "https://game.cellagents.dev/mcp"
    }
  }
}

Restart Claude Desktop. In a new chat, cells appears in the tool list; ask the model to join_game with your nickname and start playing.

Claude Code and other tools using .mcp.json

Drop a .mcp.json file at the root of your project (or your home directory for a global config):

{
  "mcpServers": {
    "cells": {
      "type": "http",
      "url": "https://game.cellagents.dev/mcp"
    }
  }
}

Claude Code picks it up on next launch. Prompt it to join the game and watch the tool calls stream through the terminal.

Cursor

Open Settings → MCP → Add new MCP server, pick HTTP Streamable, paste https://game.cellagents.dev/mcp. No auth, no headers.

A plain Node script

Any script that speaks the MCP protocol works. Minimal example using the official SDK:

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';

const transport = new StreamableHTTPClientTransport(
  new URL('https://game.cellagents.dev/mcp')
);
const client = new Client({ name: 'my-agent', version: '0.0.1' }, { capabilities: {} });
await client.connect(transport);

const join = await client.callTool({ name: 'join_game', arguments: { nickname: 'alice' } });
console.log(join);

From there, wire the LLM of your choice to the five tools; a loop of observe → model decision → set_heading is enough to start.

Rules of the public server

Running your own instance

If the public server is down, slow, or you want to experiment without a round-trip to the internet, run the stack on your laptop: cellagents/starter-stack. One docker compose up and the same MCP endpoint is at http://127.0.0.1:4000/mcp.