AgentCraft

A team of Claude agents doing real work on your code, inside a Minecraft studio you can walk around in.

Powered by Claude

Multi-agent coding usually means a wall of terminal text. AgentCraft turns it into a place.

You type a goal. A lead agent reads your repo, writes a plan and pins tasks to a wall. Workers walk to their desks, sit down and start coding in their own git worktrees while their monitors stream every file they read and every line they change. When a call is genuinely yours, an agent walks over to you with a question. When work is ready, you review the real diff and press Merge. Nothing touches your branch without that click, and nothing is ever pushed.

Close the game and the agents keep working. Open it again and the studio catches up.

How a goal plays out

1. You give a goal. Press </kbd> and type it. @juniper` messages a specific agent, with Tab completion.	2. The lead plans. Marlow splits the goal into tasks with dependencies. They land on the Task Wall, and the plan goes into the shared library.
3. The team builds. Each worker codes in its own worktree. Monitors stream the live log: tool calls, test runs, red and green diffs.	4. They come to you. A clay ! appears, the bell rings, the agent walks to the podium. Press J to answer.
5. You review and merge. A real code review screen: file list, line numbers, collapsed context, the worker’s summary and the reviewer’s notes.	6. Memory stays shared. The lead’s plan, decisions and repo conventions live in a library every agent reads, and you can too.

Meet the team

Six hand-pixelled characters, each with their own silhouette and colour. Marlow leads: he plans, splits work and reviews. Juniper, Kit, Wren, Rowan and Tove build. They walk the studio with real pathfinding, sit at their desks while they type, show what they are doing with small particles and nameplates, talk in speech bubbles, and come find you when they need a decision.

The studio

The Goal Atrium. Progress ring, task counts, and how many decisions need you.	The studio floor. Desks, status lamps, the library and the lounge.

Everything is information you can read at a glance. Far away, the lamps and the cupola beacon tell you who is working, who is stuck and who is waiting on you. Closer, nameplates and cards tell you what. Up close, monitors and screens tell you exactly how.

Proven with real agents

This is not a mockup. These screenshots come from a real run with Claude agents on a sample repo, driven entirely through the game: a goal typed into the console, questions and permission prompts answered in game, merges reviewed in the diff screen, including a merge conflict sent back to the worker and resolved. The game was restarted mid run and the Foreman was taken offline and brought back. Six features landed in the repo with its tests passing.

Marlow asks which default export format to use.	A worker's monitor streaming its live log.	Back online after a Foreman restart, the team resumes.

Safe on real repos

AgentCraft is built to point at code you care about.

• Worktrees, always. Every task runs in its own git worktree on agentcraft/<agent>/<task>. Your checkout is never touched by an agent.
• You merge, nobody else. A merge happens only when you approve it. It is refused if your checkout has uncommitted changes. Conflicts go back to the worker, who resolves them and asks again.
• Nothing is ever pushed. Agents get no git network access at all. This is enforced inside git itself, not just by a command filter, so even a push hidden in a test script or a hook fails.
• Risky commands ask first. Reads and edits inside the worktree are allowed. Anything else (writing outside it, network access, destructive commands) becomes an in game permission prompt that shows exactly what “Always allow” would cover.
• Clear authorship. Agents commit as AgentCraft <Name>. Only the merge you approve is made as you.

Quick start

You need: Windows 10 or 11, or macOS, Java 25, Node 22+, git, and a copy of Minecraft: Java Edition.

For the real agents you need Claude API access, either of these:

• ANTHROPIC_API_KEY: create a key at console.anthropic.com, then setx ANTHROPIC_API_KEY sk-ant-... and open a new terminal.
• A cloud provider supported by the Agent SDK: Amazon Bedrock (CLAUDE_CODE_USE_BEDROCK=1), Google Vertex AI (CLAUDE_CODE_USE_VERTEX=1) or Microsoft Foundry (CLAUDE_CODE_USE_FOUNDRY=1), with that provider’s usual credentials.

Then:

git clone https://github.com/blendi-remade/agentcraft
cd agentcraft

tools\launch.ps1 -Backend sim                    # try it first: a simulated team, no API usage
tools\launch.ps1 -Repo C:\path\to\your\repo      # real agents on your repo
tools\stop.ps1                                   # stop everything launch.ps1 started

On macOS, install Java 25 with brew install openjdk@25, then run from the checkout (the script selects that JDK without changing your system Java):

node tools/mac.mjs launch --backend sim                 # try the studio without API usage
node tools/mac.mjs stop --profile sim
node tools/mac.mjs launch --repo /path/to/your/repo --use-claude-login
node tools/mac.mjs stop

The macOS launcher installs npm dependencies on first run, downloads Minecraft and Fabric through Gradle, starts the Foreman and game in the background, and waits for the studio world. Use --dev for a muted client that does not take focus; --no-game starts only the Foreman. See tools/README.md for options and logs.

The first launch installs npm dependencies and lets Gradle download Minecraft and Fabric, which takes a few minutes. After that, a launch reaches the studio in under a minute. The HQ builds itself the first time; in an older world, rebuild it with /agentcraft hq.

Personal use with Claude Code. If you already use Claude Code, tools\launch.ps1 -UseClaudeLogin runs the agents on your own claude CLI login instead of an API key. Anthropic does not allow third party tools to offer claude.ai login to their users, so this is off by default and meant for running AgentCraft yourself. To make it permanent for yourself, put {"claude": {"useClaudeLogin": true}} in ~/.agentcraft/config.json.

Your name. The agents call you by your OS user name. Change it with -ForemanArgs '--user-name','Sam', AGENTCRAFT_USER_NAME, or {"userName": "Sam"} in ~/.agentcraft/config.json.

Controls

All keys can be rebound in Options, Controls.

Console commands. Plain text starts a new goal. @name message talks to an agent.

How it works

flowchart LR
    subgraph game ["Minecraft (Fabric mod)"]
        HQ["HQ, agents, monitors,<br/>Task Wall, podium"]
        UI["Console, decisions,<br/>diff review, library"]
    end
    subgraph foreman ["Foreman (Node + TypeScript)"]
        Team["Lead + workers<br/>(Claude Agent SDK)"]
        State["Task graph, messages,<br/>memory, decisions"]
        Git["Worktrees, diffs,<br/>approved merges"]
    end
    Repo[("Your git repo")]
    game <-->|"WebSocket, localhost only"| foreman
    Team --> Git --> Repo

• The Foreman (foreman/) runs the agents and owns all the state: tasks and their dependencies, messages, shared memory, decisions and worktrees. Everything is saved to disk and Claude sessions resume by id, so it survives restarts and crashes.
• The mod (mod/) is the window and the controls. It draws what the Foreman knows and sends back what you decide. If the game closes, no work is lost.
• The sim backend is a scripted team that exercises every feature with real git edits. It powers the demo, the screenshot QA and development, without any API usage.

Costs

The claude backend bills per token to your Anthropic or cloud provider account. A small goal costs a few dollars. For a cheaper team:

tools\launch.ps1 -Repo C:\path\to\repo -ForemanArgs '--model','sonnet','--effort','low','--workers','juniper,kit'

Measured with those settings on the sample repo: three two task goals, including two merge conflicts the workers resolved, took 2 to 10 minutes each and about $6 in total. The sim backend is free.

Project layout

Development

cd foreman; npm test                       # 482 tests
cd mod; .\gradlew.bat build                # the mod
node tools/qa.mjs --home .agentcraft-home  # capture the 10 shot QA gallery

Every visual in this README was captured in game through the DevBridge, a small localhost API in the mod that scripts use to move the camera and take screenshots. Start with mod/DEV.md and mod/FEATURES.md for the mod, foreman/README.md for the orchestrator and docs/QA.md for the screenshot suite.

Status

AgentCraft is young and has been used by one person on one machine. Today it is:

• Windows and macOS development launchers. Both platforms have desktop notifications when the agents need a decision. macOS has been tested on Apple Silicon; Intel Macs are not yet tested.
• Singleplayer, one studio per world, on Minecraft 26.3.
• Run through the development client (gradlew runClient). A regular mod release for normal launchers is planned.

Issues and ideas are welcome.

License

MIT. All art (skins, block textures and UI sprites) is generated by the scripts in assets-src/ and is covered by the same license. Minecraft is not included: Gradle downloads it from Mojang for development, and players need their own copy of Minecraft: Java Edition. AgentCraft is not affiliated with Mojang, Microsoft or Anthropic.