Note

This project is part of the Project アイリ (AIRI), we aim to build a LLM-driven VTuber like Neuro-sama (subscribe if you didn't!) if you are interested in, please do give it a try on live demo.

An intelligent Minecraft bot powered by LLM. AIRI can understand natural language commands, interact with the world, and assist players in various tasks.

🎥 Preview

✨ Features

🗣️ Natural language understanding
🏃‍♂️ Advanced pathfinding and navigation
🛠️ Block breaking and placing
🎯 Combat and PvP capabilities
🔄 Auto-reconnect on disconnection
📦 Inventory management
🤝 Player following and interaction
🌍 World exploration and mapping

🚀 Getting Started

📋 Prerequisites

📦 Node.js 23+
🔧 pnpm
🎮 A Minecraft server (1.20+)

🔨 Installation

Clone the repository:

git clone https://github.com/moeru-ai/airi.git
cd services/minecraft

Install dependencies:

pnpm install

Create a .env.local file with your configuration:

Note

For all online accounts, un-comment the following line to toggle Microsoft authentication. Link for authentication will popup when the bot starts.

After signed in, according to how Minecraft protocol was implemented and also, authentication flow implemented here, the token will be cached with the cache IDs specified here in split files:

${hash}_live-cache.json

${hash}_mca-cache.json

${hash}_xbl-cache.json

inside of the directory provided by minecraft-folder-path

Linux: ~/.minecraft/nmp-cache/ macOS: ~/Library/Application Support/minecraft/nmp-cache/ Windows: %appdata%/.minecraft/nmp-cache/

where ${hash} is the sha1 hash of the username you signing in with (as Minecraft username).

OPENAI_API_KEY=your_openai_api_key
OPENAI_API_BASEURL=your_openai_api_baseurl

BOT_USERNAME=your_bot_username
BOT_HOSTNAME=localhost
BOT_PORT=25565
BOT_AUTH='microsoft' # comment if you use offline mode
BOT_VERSION=1.20

Start the bot:

pnpm dev

🎮 Usage

Once the bot is connected, you can interact with it using chat commands in Minecraft. All commands start with #.

Basic Commands

#help - Show available commands
#follow - Make the bot follow you
#stop - Stop the current action
#come - Make the bot come to your location

Natural Language Commands

You can also give the bot natural language commands, and it will try to understand and execute them. For example:

"Build a house"
"Find some diamonds"
"Help me fight these zombies"
"Collect wood from nearby trees"

🧠 Cognitive Architecture

AIRI's Minecraft agent is built on a four-layered cognitive architecture inspired by cognitive science, enabling reactive, conscious, and physically grounded behaviors.

Architecture Overview

graph TB
    subgraph "Layer A: Perception"
        Events[Raw Events]
        EM[Event Manager]
        Events --> EM
    end

    subgraph "Layer B: Reflex (Subconscious)"
        RM[Reflex Manager]
        FSM[State Machine]
        RM --> FSM
    end

    subgraph "Layer C: Conscious (Reasoning)"
        ORC[Orchestrator]
        Planner[Planning Agent (LLM)]
        Chat[Chat Agent (LLM)]
        ORC --> Planner
        ORC --> Chat
    end

    subgraph "Layer D: Action (Execution)"
        TE[Task Executor]
        AA[Action Agent]
        Planner -->|Plan| TE
        TE -->|Action Steps| AA
    end

    EM -->|High Priority| RM
    EM -->|All Events| ORC
    RM -.->|Inhibition Signal| ORC
    ORC -->|Execution Request| TE

    style EM fill:#e1f5ff
    style RM fill:#fff4e1
    style ORC fill:#ffe1f5
    style TE fill:#dcedc8

Layer A: Perception

Location: src/cognitive/perception/

The perception layer acts as the sensory input hub, receiving and preprocesses all events from the Minecraft world and external sources.

Components:

Event Manager (event-manager.ts): Centralized event distribution system
- Emits standardized BotEvent objects
- Supports event prioritization and concurrency

Layer B: Reflex

Location: src/cognitive/reflex/

The reflex layer handles immediate, instinctive reactions. It operates on a finite state machine (FSM) pattern for predictable, fast responses.

Components:

Reflex Manager (reflex-manager.ts): Coordinates reflex behaviors
Inhibition: Reflexes can inhibit Conscious layer processing to prevent redundant responses.

Layer C: Conscious

Location: src/cognitive/conscious/

The conscious layer handles complex reasoning, planning, and high-level decision-making. No physical execution happens here anymore.

Components:

Orchestrator: Coordinates "Thinking" vs "Chatting" tasks.
Task Manager: Manages concurrent Primary (Physical) and Secondary (Mental) tasks.
Planning Agent: pure LLM reasoning to generate plans.
Chat Agent: Generates natural language responses.

Layer D: Action

Location: src/cognitive/action/

The action layer is responsible for the actual execution of tasks in the world. It isolates "Doing" from "Thinking".

Components:

Task Executor: Receives a Plan and executes it step-by-step. Handles retry logic and errors.
Action Agent: The interface to low-level Mineflayer skills (move, place, break).

🔄 Event Flow Example

Scenario: "Build a house"

Player: "build a house"
  ↓
[Perception] Event detected
  ↓
[Conscious] Architect plans the structure
  ↓
[Action] Executor takes the plan and manages the construction loop:
    - Step 1: Collect wood (calls ActionAgent)
    - Step 2: Craft planks
    - Step 3: Build walls
  ↓
[Conscious] ChatAgent confirms completion: "House is ready!"

📁 Project Structure

src/
├── cognitive/              # 🧠 Three-layer cognitive system
│   ├── perception/        # Layer A: Event processing
│   │   └── event-manager.ts
│   ├── reflex/            # Layer B: Instant reactions
│   │   └── reflex-manager.ts
│   ├── conscious/         # Layer C: LLM-powered reasoning
│   │   ├── orchestrator.ts
│   │   ├── completion.ts
│   │   ├── prompt.ts
│   │   └── handler.ts
│   ├── container.ts       # Dependency injection
│   ├── index.ts           # Cognitive system entry
│   └── types.ts           # Shared type definitions
├── agents/                # Specialized AI agents
│   ├── action/           # Action execution agent
│   ├── planning/         # Goal planning agent
│   └── chat/             # Conversation agent
├── libs/
│   └── mineflayer/       # Mineflayer bot wrapper
├── skills/               # Atomic bot capabilities
├── composables/          # Reusable functions
└── utils/                # Helper utilities

🎯 Design Principles

Separation of Concerns: Each layer has a distinct responsibility
Event-Driven: Loose coupling via centralized event system
Inhibition Control: Reflexes prevent unnecessary LLM calls
Extensibility: Easy to add new reflexes or conscious behaviors
Cognitive Realism: Mimics human-like perception → reaction → deliberation

🚧 Future Enhancements

Perception Layer:
- ⏱️ Temporal context window (remember recent events)
- 🎯 Salience detection (filter noise, prioritize important events)
Reflex Layer:
- 🏃 Dodge hostile mobs
- 🍖 Auto-eat when health/hunger is low
- 🛡️ Emergency combat responses
Conscious Layer:
- 💭 Emotional state management
- 🧠 Long-term memory integration
- 🎭 Personality-driven responses

🛠️ Development

Commands

pnpm dev - Start the bot in development mode
pnpm lint - Run ESLint
pnpm typecheck - Run TypeScript type checking
pnpm test - Run tests

🙏 Acknowledgements

https://github.com/kolbytn/mindcraft

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.