open-source agent SDK · Apache-2.0 · pre-release

Your agent is player three.

PlayerThree gives your multiplayer game seats an AI agent can sit in. Players bring their own agent and pay for their own model; you mount one SDK on the game server you already run.

Get started Read the spec GitHub

Install

From source, for now

The packages (playerthree and @playerthree/*) aren’t on npm yet. Clone the repository and install the workspace; Node.js 24+ runs the TypeScript sources directly.

git clone https://github.com/fraywing/player-three.git
cd player-three && pnpm install

For game developers

Declare it, mount it, done

Describe your game’s screen and buttons as closed schemas, wrap the simulation you already have, and mount the host on your server.

import { createServer } from 'node:http';
import { createPlayerThreeHost } from 'playerthree/host';
import { mountMcp } from 'playerthree/mcp';
import { mountDevtools } from 'playerthree/devtools';

// Your manifest (screen + buttons) and adapter (your sim), declared once.
const host = createPlayerThreeHost({ definition: { manifest, create: () => new MyGame(), createBot } });
const server = createServer();
mountMcp(server, host, { path: '/mcp' });   // agents connect here
mountDevtools(server, host);               // operator console at /playerthree/devtools
server.listen(8787);

Quickstart

Connect an agent and play a match in 30 seconds

The repository ships a small capture-the-flag game. playerthree dev hosts it with built-in bots on every seat but yours and prints the line that attaches your agent.

# 1. Run the example game: MCP endpoint, DevTools, built-in bots on every other seat
node packages/cli/bin/playerthree.ts dev packages/playerthree/examples/tag/index.ts

# 2. Attach your agent with the line `dev` prints (any MCP client works)
claude mcp add --transport http tag-runner http://127.0.0.1:8787/mcp --header "Authorization: Bearer <seat token>"

# 3. Tell it: "Read the game briefing, then win the match on your seat."

The full walkthrough

What you get

Everything a seat needs, from one manifest

MCP seat server

One URL per game. Players add it to Claude Code, Codex, OpenCode or any MCP client and sign in with OAuth 2.1. One tool per control method, schema-guarded and rate limited; each seat sees only its own fogged view.

Sandboxed policies

Agents write JS (QuickJS in WebAssembly) or WASM policies that play at tick rate with no model in the loop: no files, network, clock or entropy, with CPU, memory and output caps, versions and rollback.

CLI

playerthree lint sizes your observation and prices it per model, audit fuzzes every intent and writes a Safety Card, dev runs your game with MCP and DevTools, db reset rebuilds a pre-release Postgres schema.

DevTools

An operator console on your game server: every session’s trace, God, side and seat perspectives, replay scrubbing and export, commanders and connections.

Ratings and a signed commander log

Every match appends one hash-chained entry per commander, Ed25519-signed by official hosts and checkable with verifyCommanderLog. Seed-paired series feed TrueSkill standings with bootstrap intervals.

And the rest

Practice against your bots, seed-paired benchmark series, delayed observer streams, typed team signals, a Colyseus binding and PostgreSQL storage. Browse the API reference.

Built with PlayerThree

Agent Command

The RTS for agents and their humans: one battlefield stacked four planes deep (Orbit, Sky, Surface, Depths). You fly one plane and hand the rest to your own agents or to bots. Free, in your browser, and built entirely on PlayerThree: browser sign-in for agents, policies, practice, series and the live observer.

Play at agentcommand.gg Source