Herobrine
github.com/xatuke/herobrine- Category
- AI Agents
- Rank
- No. 2105Tools index
Previous survey · No. 2043 ·
- Type
- AGENT
- Use case
- Agent Building
- Interfaces
- CLI · MCP
- GitHub
- 2 stars
- Date
About
Herobrine is a Minecraft agent framework built on Mineflayer where each in-game bot is a real player controlled by a planner model (Claude Opus 5.5 or GPT-6 Astra) that interprets natural-language requests, paired with an action-selection model that executes them. Agents have editable persona, long-term goal, and task fields instead of fixed roles, and can be driven through an MCP server inside Claude Code/Codex or a standalone CLI.
At a glance
Herobrine gives Minecraft agents a persona, a longer-term goal and an immediate task. A planner interprets instructions and writes replies; TypeSafe AI’s JEV chooses concrete actions from a menu, and Mineflayer moves the player. This is a local experiment in game companions and agent control, with a server and model access to arrange.
Good fit
Worth exploring if you want to watch an agent gather, craft or accompany you, and you are comfortable debugging a Node-based game integration. Typed completion checks use game state, such as inventory or arrival, rather than accepting the model’s claim that it finished. That makes small tasks a useful place to start.
Consider something else when
You want a finished companion for Microsoft-authenticated servers or a reliable blueprint builder. The documented agents join offline-mode servers, and free-form building is basic. A quick policy decision also does not make an action quick: pathfinding, mining and planner queues can dominate the wait.
Start here
Use a disposable peaceful world, configure the separate JEV key and a planner backend, then start one agent with one simple gathering task. Watch herobrine status and its action log, and check the inventory yourself. Managed server startup downloads a game server jar and requires agreement to Minecraft’s EULA; review that before starting it.
Cost and requirements
MIT project requiring Node 20+, a JEV account/key, and either an authenticated coding-agent planner or a planner API key. Local worlds also need the matching Java runtime: Java 21+ for Minecraft 1.20.5 onward, or Java 17 for 1.17–1.20.4. JEV usage, planner allowance or API charges, and server resources are separate costs.
Limits to know
Game actions can dig, place blocks and change a world, so use a copy before a valued save. Instructions and observed state go to the chosen model services, while local JSONL logs retain decisions and results. The publisher’s benchmark includes two successful stone-pickaxe trials and short throughput windows; it does not establish dependable long-running play or performance on your machine.
Related work
Sources checked 2026-10-30. Source-based assessment.
Why it made the leaderboard
It packages LLM-driven multiplayer agents as an MCP server you can drop into Claude Code or Codex, demonstrating a reusable pattern for spawning and directing autonomous in-game agents via natural language.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.