Omo Space — The AI Team You Walk Through

The Big Idea

Right now your AI agents are invisible. You can't watch them think. You can't stop one halfway and say "no, do this instead." The "AI workforce" you were promised is really just a chat box bolted to a black box.

So we turned your AI team into a place you walk through.

Omo Space is the spatial layer that makes this real. If the world is your team, then building it is something you can finally watch happen. You hire by walking up and asking — a room rises, an agent walks in, a screen lights up. You don't read about your AI working: you stand next to it while it does.

One command drops your HQ onto a plain Minecraft world. Your founding crew — a Chief of Staff and two specialists — are already at their desks. The dark screens behind them are their thoughts, streaming live as they reason with Gemini, coordinated through Google's ADK, and reaching for tools over MCP.

Built net-new for the Google AI Agents Challenge, Omo is a working prototype — not a demo. Clone the repo, drop in a free Gemini key, run one command, and walk into your AI company in about ten minutes.

How It Works — The Three Ores

Three components power Omo, each doing real work. Together they make an AI team that feels like a place rather than a process.

Ore One

◆ Gemini — The Brain

Every agent thinks with Gemini: the Chief of Staff, the specialists, and the new helpers it hires. Same brain, many jobs. Live reasoning streams appear on screens behind each agent as they think.

Ore Two

◆ ADK — The Company

Google's Agent Development Kit makes a real team — a boss that hands work to specialists, with the hand-offs visible as they happen. The Chief of Staff delegates to the right agent for each job.

Ore Three

◆ MCP — The Hands

Every tool plugs in over MCP — Gmail, Meta Ads, notes, world_build. The trick: the company grows itself by calling MCP tools too. Rooms gate tool sets by name prefix.

Why It Feels Like Magic

◿ It's all live. Three streams at once — the agent thinking, the world rising, and your data updating — all in front of you, no reload.

✦ Zero setup. You say what you want. The company builds the team, the rooms, and every tool connection itself.

✏ You stay the boss. Before it touches the outside world, it asks. One tap to say yes. Walk in, talk, change your mind, stop it — any time.

Architecture

Three processes talk over one WebSocket bridge at :8765. The plugin is the only code that touches the Minecraft world. Full wire protocol in the repo docs.

Minecraft Layer

Paper Server + Plugin (Java)

The only code that touches the world. Builds the HQ island, raises rooms block-by-block, spawns villagers, drives in-world screens. Vanilla Minecraft 1.21.4 — no mods required.

Paper 1.21.4 Java 21 World Builder

Runtime Layer

Node.js Runtime (TypeScript)

Single WebSocket server. Runs the agent loop, maps rooms to tool sets, and hosts the Gemini world-architect that designs a unique building per request.

WebSocket :8765 Agent Loop World Architect

Brain Layer

ADK Crew (Python, Gemini)

Chief of Staff + specialists running on Gemini through Google's ADK. Delegates work and calls tools over MCP. Sensitive actions always wait for in-game approval.

Gemini ADK MCP Tools

Room Gating

Rooms gate tool sets by name prefix: mail-* grants Gmail access, ads-* grants Meta Ads, workshop/code rooms get coding tools. Each agent's reach is defined by the room it occupies.

mail-* → Gmail ads-* → Meta Ads Code Rooms

Approval Gate

Sensitive actions (gmail_send, meta_ads_pause, meta_ads_update_budget) always wait for in-game approval. A tap-to-approve gate hovers over the agent — one tap to say yes.

/omo approve /omo deny Always-On Guard

Minecraft Java Client Paper Plugin Node Runtime Python ADK Crew WebSocket :8765

The Campus

Your company isn't a tab you check. It's a place you stroll through.

HQ Island

Drop It Anywhere

One command — /omo hq — places your headquarters on any flat Minecraft terrain. A glowing island with a bridge to the main path.

Self-Building Rooms

Rise Live

Ask the Chief of Staff for a new teammate and a room builds itself block by block, right in front of you. No loading screen. No "please wait."

Agents at Desks

Villagers With Jobs

Each agent is a Minecraft villager at a desk. Walk up, talk to it, watch it think. The dark screen behind it streams live reasoning from Gemini.

Live Screens

Reasoning + Dashboards

In-world screens show real data — ad spend, click rates, email summaries — redrawn as fresh data lands. No abstraction, no separate dashboard.

World Architect

Unique Per Request

Gemini designs each new room uniquely — a classroom, an ad analysis wing, a DJ booth. No two rooms look the same. The architect builds what the job needs.

Many Rooms at Once

One Company

A teacher. An ad analyst. An email helper. A DJ. Every building is a job you handed over, all working at the same time. Stroll between them.

Key Properties

Live

◿ Everything Real-Time

Agent reasoning, world building, and live data update simultaneously. No reload, no polling. You watch it happen.

Zero Setup

✦ One Command

./agentcraft starts everything — ADK crew, runtime, Paper server. Clone, add your Gemini key, and walk in. About 10 minutes on macOS.

You Stay the Boss

✏ Tap to Approve

Before it sends email or changes a budget, it asks. One tap in-game to say yes. Walk in, talk, change your mind, stop it — any time.

Spatial

□ Rooms Define Reach

An agent's tools, memory, and personality are tied to its physical room. Walk into the mail room to read email. Walk into the ad wing for campaign data.

Multi-Agent

◆ Concurrent Crew

Chief of Staff + specialists + ad-hoc hires all coexist. The ADK org delegates work. Agents hand off tasks visibly.

Vanilla

☰ No Mods Required

Runs on vanilla Minecraft Java 1.21.4. The plugin handles everything server-side. You play with the standard client.

In-Game Commands

Every interaction happens through Minecraft chat. No external dashboard needed.

        /omo hq                         drop your HQ island and founding crew
/omo spawn <id> <role...>       hire a Gemini agent at your position
/omo say <id> <text...>         talk to an agent directly (or just chat near it)
/omo approve <id>                 approve a pending sensitive action
/omo deny <id>                    deny a pending sensitive action
/omo list                          list all agents
/omo despawn <id>                 remove an agent
/omo room define <name>           mark your current region as a named room
/omo capture                       capture a real app onto an in-world screen
/omo cinema <url>                 stream a webpage onto an in-world screen
      

Run It Locally

Clone the repo, drop in a free Gemini key, run one command, and walk into Omo in Minecraft. About 10 minutes on macOS (Apple Silicon).

Prerequisites

brew install openjdk@21 maven node python jq

You'll also need Minecraft: Java Edition 1.21.4 (the real game client) and a free Gemini API key from Google AI Studio.

Quick Start

        # 1) Clone the repo
git clone https://github.com/harrythentrepreneur/omo-minecraft.git
cd omo-minecraft

# 2) First-time setup — fetches Paper, builds the plugin, installs deps
./agentcraft setup

# 3) Add your free Gemini key
echo 'GOOGLE_API_KEY=YOUR_KEY' >> runtime/.env
mkdir -p omo-agent/omo && echo 'GOOGLE_API_KEY=YOUR_KEY' >> omo-agent/omo/.env

# 4) Install the ADK crew
python3 -m venv omo-agent/.venv
omo-agent/.venv/bin/pip install google-adk

# 5) Start everything
./agentcraft

# 6) Open Minecraft Java 1.21.4 → Multiplayer → Direct Connect → localhost
#    Then in chat, drop your HQ:
/omo hq
      

Management Commands

        ./agentcraft           start / reload everything (run after ANY change)
./agentcraft stop      stop everything
./agentcraft logs      tail all logs
./agentcraft restart   force a full clean restart
./agentcraft status    show what's running
      

Papers & Reports

Technical reports related to Omo Space and spatial agent habitats.

Omo Space: The AI Workforce You Build by Walking Through Minecraft

Technical Report · 2025

Read ↗

Towards Real-Time Spatial Agent Communication

Preprint · 2025

PDF ↗

AgentJam: AI Mentors That Play Minecraft With You

Technical Report · 2025

PDF ↗