blendi-remade/agentcraft

Manage your real AI agents, using Minecraft as your interface!

What it solves

Multi-agent coding usually means staring at terminal text. AgentCraft turns it into a place: a Minecraft world where a team of AI agents works on your actual code, and you can see, question, review, and merge everything from inside the game. You give a goal, a lead agent plans, workers code, and you decide what lands.

How it works

AgentCraft has two parts: a Minecraft Fabric mod (the game side) and a backend called the Foreman (Node + TypeScript). The Foreman runs the agents using the Claude Agent SDK (or Codex) and holds all state: task graph, messages, memory, decisions, and git worktrees. The mod draws that state in the game: agents walk to desks, their monitors stream live logs, a task wall shows progress, a podium appears when an agent needs a decision, and a diff screen lets you review changes before merging.

You type a goal into the console (key `), the lead agent splits it into tasks with dependencies, and workers start coding each task in its own git worktree. Work only merges when you approve it, and only if your checkout has no uncommitted changes. Agents never get network access, and risky commands become in-game permission prompts. The Foreman saves everything to disk and resumes by session ID, so closing the game doesn't stop the work.

There's also a simulation backend that runs a scripted team with real git edits, so you can try the full experience without any API usage.

Who it's for

Developers who want to supervise AI coding agents with a visual, interactive interface—especially those who enjoy Minecraft. You need Minecraft: Java Edition, Java 25, Node 22+, and either Claude API access (Anthropic key, Bedrock, Vertex, Foundry, or your own Claude Code login) or a Codex login. It's also useful for exploring how multi-agent workflows behave in a safe, inspectable environment.

Highlights

  • Real agents (Claude or Codex) make actual code edits in isolated git worktrees, never touching your main checkout.
  • You review real diffs with line numbers and context inside the game, then press Merge to approve.
  • Nothing is ever pushed automatically—agents have no git network access, enforced at the git level.
  • The studio shows everything visually: task board, live monitors, agents walking, speech bubbles, and a podium for questions.
  • Decisions, permission prompts, and merges are all handled in-game with a single key (J).
  • A shared memory library lets agents (and you) read plans, decisions, and repo conventions.
  • Supports a mixed team—Claude and Codex agents can work together, with model names shown on nameplates.
  • Sim backend lets you explore risk-free, without any token costs.

Related

  • Project
    JuliusBrussee/caveman-code

    A terminal coding agent that uses a four-layer compression system to drastically reduce LLM token usage and costs while maintaining performance.

  • Project
    zenstory-ai/novel-to-game

    NovelToGame is an open‑source set of AI‑agent skills that turn a full novel into a playable, source‑grounded game (web‑browser or 3D) with full design documentation and a single‑run QA. It works with Claude Code, Codex, or Kimi Code, requires no bundled engine, and records every mechanic’s citation to the original text.

  • Project
    suboss87/FDEOps

    FDEOps is a Node‑JS CLI that supplies a library of 35 “task‑skill” prompts for forward‑deployed engineers. Integrated with an existing AI coding assistant, it structures strategy, architecture and engineering work, stores all customer context in local Markdown files, and runs entirely offline.

  • Project
    china-qijizhifeng/agentic-harness-engineering

    An observability-driven framework that automatically evolves the harness—including prompts, tools, and middleware—around a coding agent to improve performance without changing the base model.