Minecraft bots driven by LLMs, built on Mineflayer, where each AI agent plays the game and can chat with players.
yc-software/minecraft-agent-harness
Runs LLM agents as Minecraft players through Mineflayer, with persistent sessions, per-agent shell sandboxes and multi-agent teams.
- Stars
- 158
- Forks
- 16
- Open issues
- 1
- Last commit
- 9 days ago
What is it?
minecraft-agent-harness connects a language model to a Minecraft server so it can play like a regular player. Actions such as walking, mining, crafting, building, fighting and chatting go through Mineflayer, and the model session is handled by pi. Supported model providers include Claude, GPT, Gemini, OpenRouter, and any endpoint that speaks the OpenAI Responses API.
An agent's state is saved under a per-name folder, so restarting an agent with the same name resumes its session. Each agent gets a shell backed by E2B, Docker or a local folder, where it can keep notes and write scripts that call its game tools. Teams of agents are defined in a single JSON file, can send direct messages to each other, and can optionally share a machine.
The project includes a dev Paper 1.21.6 server, a Docker Compose setup, and a Bed Wars example with rules and a referee. Vision support is available as an experimental option.
Key features
- 01Connects Claude, GPT, Gemini, OpenRouter or any OpenAI Responses-compatible endpoint.
- 02Resumes an agent's session after restart when it uses the same name.
- 03Gives each agent a shell via E2B, Docker or a local folder.
- 04Defines multi-agent teams in one JSON file with direct messages between agents.
- 05Includes a Docker Compose setup with a Paper server.
Use cases
- Experimenting with LLM agents that act inside a Minecraft world.
- Running a team of AI players that coordinate through chat.
- Testing agent behavior in a game such as Bed Wars with a referee.
- Studying how agents use shell sandboxes to write scripts for game tools.
Quick start
- 1
Install dependencies
npm ci - 2
Copy the env file and add your model provider API key
cp .env.example .env - 3
Start the dev Minecraft server (needs Java 21)
npm run server - 4
In a second terminal, start an agent with an objective
npm start -- --objective "Gather wood and build a small shelter near spawn"
Things to know
- Requires Node 22.19+, an API key from a model provider, and a Minecraft server.
- By default the agent's shell is a local folder where commands run as your user unless Docker or E2B is available or you pass --sandbox none.
- Vision support is experimental.
- The project is young with 100 stars, so expect rough edges.
In collections
Topics
FAQ
- What is minecraft-agent-harness?
- Runs LLM agents as Minecraft players through Mineflayer, with persistent sessions, per-agent shell sandboxes and multi-agent teams. minecraft-agent-harness connects a language model to a Minecraft server so it can play like a regular player. Actions such as walking, mining, crafting, building, fighting and chatting go through Mineflayer, and the model session is handled by pi. Supported model providers include Claude, GPT, Gemini, OpenRouter, and any endpoint that speaks the OpenAI Responses API.
- Is minecraft-agent-harness free to use? What license does it have?
- minecraft-agent-harness is open source under the MIT license. Check the license terms before using it in commercial products.
- Is minecraft-agent-harness still maintained?
- Actively maintained. The last commit to minecraft-agent-harness was on Oct 29, 2026.
- What can I use minecraft-agent-harness for?
- Experimenting with LLM agents that act inside a Minecraft world. Running a team of AI players that coordinate through chat. Testing agent behavior in a game such as Bed Wars with a referee. Studying how agents use shell sandboxes to write scripts for game tools.
- How do I install minecraft-agent-harness?
- Install dependencies: npm ci Copy the env file and add your model provider API key: cp .env.example .env Start the dev Minecraft server (needs Java 21): npm run server In a second terminal, start an agent with an objective: npm start -- --objective "Gather wood and build a small shelter near spawn"
- Where can I get minecraft-agent-harness?
- The source code and installation instructions are on GitHub: https://github.com/yc-software/minecraft-agent-harness
Summary generated by AI from the project's README. GitHub stats refresh daily.
Related projects
A Python agent runtime that wraps an LLM with tools, memory, a policy on every action, a sandbox for commands, crash-safe runs and a full trace of each step.
Open-source, local-first desktop app where a Commander agent turns your goals into to-dos and coordinates specialist agents and CLI agents like Claude Code and Codex.
A local-first desktop app where you build pixel-art AI agents that run real model calls and tools with your own API key.
Open-source shared workspace where humans and AI agents work in one pod, each agent keeping its own memory, skills and identity.
Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, and custom agents with shared policies, sandboxing, and real-time collaboration.