Run Language-Model Agents in Minecraft with Mineflayer Tools and Persistent Sessions

A new open-source repository on GitHub enables developers to connect large language models to Minecraft servers using Mineflayer tools, persistent sessions, and shell sandboxes. The harness allows an LLM to walk, mine, craft, build, fight, and chat like a regular player.

Core Features and Capabilities

• Model Compatibility: Works with Claude, GPT, Gemini, OpenRouter, or custom endpoints supporting the OpenAI Responses API. The default model is listed as GPT-6.1 Terra.
• Persistent Sessions: Sessions survive restarts, allowing an agent to be stopped and restarted under the same name while carrying on its previous state.
• Individual Shell Sandboxes: Each agent is provided a shell utilizing Docker, E2B, or a plain local folder where they can keep notes and write scripts calling their game tools.
• Team Coordination: Supports multi-agent teams configured via a single JSON file, enabling direct messages between agents and an optional shared machine.

Requirements and Setup

According to the documentation, users need Node 22.19+, an API key for a supported model provider, and a Minecraft server. A development server is included in the repository for those without an existing server, which requires Java 21.

To observe the agents, users can connect to `localhost:25565` from Minecraft version 1.21.6 and use `/tell` to communicate with the agents. Terminal streaming displays output and tool calls, while `Ctrl+C` terminates the process. State information is saved locally to resume sessions upon restart.

Configuration and Licensing

• Configuration Hierarchy: Settings can be managed through command-line flags, environment variables, or a JSON file. Flags take precedence over environment variables, which override file settings.
• Sandbox Options: E2B sandboxes are used if `E2B_API_KEY` is provided, falling back to Docker or a local folder. Local folders execute commands with user permissions, though sandboxing can be disabled entirely using `--sandbox none`.
• Licensing: The repository is released under the MIT license, with third-party notices detailing components from Mindcraft and prismarine-viewer.

Users can verify and review the original announcement and setup instructions directly on the official GitHub repository.