MineCode MCP

Version-accurate Minecraft data for AI assistants

MineCode is a local Model Context Protocol (MCP) server that gives AI assistants like GitHub Copilot and Claude real-time access to version-accurate Minecraft data, documentation, vanilla presets, and your Minecraft logs.

Written for a hackathon about MCP sponsored by dust, alpic, and others. Prototyped using the worst ai model ;-; (note I rewrote it now it is actually viable). Please star if you'd like to help out, and open issues for anything broken.

🎯 The problem this solves

AI assistants get Minecraft syntax wrong constantly, and they do it confidently. The reason is simple: Minecraft's datapack format changed substantially and repeatedly, and every model's training data is older than the current game.

• 1.20.5 replaced item NBT with typed components. Old NBT is now a hard parse error.

• 1.21 renamed every datapack folder to singular (advancements/ → advancement/). A pack with the old names loads with no error and no content — it silently does nothing.

• 1.21.2 dropped the generic. prefix from every attribute ID.

• 1.21.4 turned custom_model_data from an integer into an object.

• 1.21.5 made text components strictly typed.

minecraft.wiki documents only the latest version, so consulting it for an older pack actively makes this worse.

MineCode attacks this from four directions:

1. minecraft_start_session — detects the target version from pack.mcmeta before any code is written, so nothing downstream is guessing.

2. get_technical_changes — returns what actually changed between two versions, from misode/technical-changes plus a curated table of the traps agents fall into most.

3. get_command_usage / validate_command — command syntax compiled from the game's own Brigadier grammar, and a parser to check the agent's output against it.

4. Honest tool descriptions — every wiki tool states up front that it covers the latest version only, and names the version-exact alternative.

Related MCP server: Minecraft Dev MCP

🚀 Installation

Requires Python 3.10 or newer. Check with python --version (Windows: py --version).

Windows

py -m pip install --upgrade pip
py -m pip install minecode-mcp

Verify:

py -m minecode.server --help 2>$null; py -c "import minecode; print('ok')"

If py is not recognised: Python isn't installed or wasn't added to PATH. Reinstall from python.org with "Add python.exe to PATH" ticked. Avoid the Microsoft Store build — it sandboxes file access, which breaks reading pack.mcmeta and Minecraft logs from arbitrary paths.

macOS

python3 -m pip install --upgrade pip
python3 -m pip install minecode-mcp

If your Python is Homebrew-managed you'll hit error: externally-managed-environment. Use a venv (see below) or pipx:

brew install pipx && pipx install minecode-mcp

Linux

python3 -m pip install --upgrade pip
python3 -m pip install minecode-mcp

Most modern distributions (Arch, Debian 12+, Ubuntu 23.04+, Fedora) mark the system Python as externally managed and will refuse the command above. That protection is correct — don't override it with --break-system-packages. Use one of:

# Option A: pipx — recommended, isolated but still on PATH
sudo pacman -S python-pipx        # Arch
sudo apt install pipx             # Debian/Ubuntu
sudo dnf install pipx             # Fedora
pipx install minecode-mcp

# Option B: user install
python3 -m pip install --user minecode-mcp

# Option C: a venv you point the client at (see Configuration)

Isolated install (any platform)

Works everywhere and never touches system Python. Note the absolute path it prints — you'll need it for the client config.

# Linux / macOS
python3 -m venv ~/.minecode-venv
~/.minecode-venv/bin/pip install minecode-mcp
echo ~/.minecode-venv/bin/minecode

# Windows
py -m venv $HOME\.minecode-venv
& $HOME\.minecode-venv\Scripts\pip.exe install minecode-mcp
Write-Output "$HOME\.minecode-venv\Scripts\minecode.exe"

Upgrading and uninstalling

pip install --upgrade minecode-mcp   # or: pipx upgrade minecode-mcp
pip uninstall minecode-mcp           # or: pipx uninstall minecode-mcp

Upgrading doesn't clear the response cache. That's intentional — version-pinned data can't go stale. To clear it anyway, call the cache_status tool with clear=true, or delete the directory shown by cache_status.

Which Python am I actually using?

The single most common setup failure is installing into one interpreter and pointing the client at another. When in doubt, get the absolute path and use it verbatim in your client config:

python3 -c "import sys; print(sys.executable)"   # Linux/macOS
py -c "import sys; print(sys.executable)"        # Windows

▶️ Running the server

MineCode is an MCP server, not an app you sit in front of. It speaks JSON-RPC over stdin/stdout and is normally launched by your AI client, not by you. You rarely need to start it manually — but you do need to know how, because that's how you check the install before wiring up a client.

The two ways to launch it

minecode                  # console script, installed by pip
python -m minecode.server # module form — identical, works even if the script isn't on PATH

On Windows use py -m minecode.server.

What "working" looks like

Running it directly looks like a hang. That is correct:

$ minecode
[INFO] Loaded assistant preprompt from .../assistant_preprompt.txt
[INFO] Starting MineCode MCP server
[INFO] MineCode MCP server starting (stdio)
[INFO] Registered 30 tools, 1 prompts, 2 resources

…and then nothing. The server is waiting for JSON-RPC on stdin. This is a healthy server, not a freeze. Press Ctrl+C to stop it.

The line that matters is Registered 30 tools. If you see it, the install is good. Logs go to stderr, so they never corrupt the protocol stream on stdout.

Verifying the install without a client

python -c "
from minecode import tools
print(f'{len(tools.TOOLS)} tools, {len(tools.HANDLERS)} handlers')
assert {t.name for t in tools.TOOLS} == set(tools.HANDLERS)
print('registry consistent')
"

To exercise a tool without any MCP client at all:

python -c "
from minecode import handlers
r = handlers.handle_get_command_usage('1.21.4', 'give')
print(r['usage'])
"

Expected: ['/give <targets> <item>', '/give <targets> <item> <count>']

A full protocol handshake, if you want to be thorough:

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"t","version":"1"}}}' \
  '{"jsonrpc":"2.0","method":"notifications/initialized"}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  | python -m minecode.server 2>/dev/null | tail -1 | head -c 300

Normal usage

Configure your client (next section), then restart it. The client spawns the server itself and keeps it alive for the session. From then on you just talk to your assistant — start with something like "set up my datapack and tell me what version it targets", which triggers minecraft_start_session.

Troubleshooting

⚙️ Configuration

Claude Desktop / Claude Code

{
  "mcpServers": {
    "minecode": {
      "command": "minecode"
    }
  }
}

VS Code (GitHub Copilot)

Add to User Settings (Ctrl+Shift+P → "MCP: Open User Configuration"), or create .vscode/mcp.json in your workspace:

{
  "servers": {
    "minecode": {
      "type": "stdio",
      "command": "minecode",
      "args": []
    }
  },
  "inputs": []
}

The minecode script is installed alongside the package and already points at the right interpreter, so this form works for pipx, venv, and --user installs alike.

The module form is equivalent only if the package is importable from whichever python the client resolves:

{
  "servers": {
    "minecode": {
      "type": "stdio",
      "command": "python",
      "args": ["-m", "minecode.server"]
    }
  },
  "inputs": []
}

⚠️ VS Code spawns the server with a bare python from PATH — not your selected Python interpreter, and not a pipx or venv environment. With a pipx install this fails with No module named minecode, since the package lives in pipx's own venv. Use "command": "minecode" above, or the absolute interpreter path from the next section.

On Windows use "command": "py" for the module form. py is the Windows launcher and does not exist on macOS or Linux.

If minecode isn't on PATH

Common with venv, pipx, and --user installs. Give the absolute path to the interpreter that has the package, and let it run the module:

{
  "mcpServers": {
    "minecode": {
      "command": "/home/you/.minecode-venv/bin/python",
      "args": ["-m", "minecode.server"]
    }
  }
}

Get the exact path with python3 -c "import sys; print(sys.executable)" from the environment where you installed it.

Windows JSON: backslashes must be escaped — C:\\Users\\... — or use forward slashes, which also work.

Restart the client fully after editing the config. Most MCP clients read it only at startup, so a reload isn't enough.

🛠️ Tools

Start here

Version correctness

Commands

Spyglass (authoritative, version-exact)

Misode (real vanilla data)

Minecraft Wiki — ⚠️ latest version only

Datapack Emulator — run the pack without the game

Backed by Datapack Emulator (installed with MineCode from PyPI; its source is a submodule at datapack-emulator/). It loads the pack as a server of the chosen version would, ticks it, and reports what happened. Results carry ok and a one-line summary; ok is false for the silent failures the game only warns about (functions that do not parse, functions/ folders on 1.21+).

Not modelled: physics, mob AI, redstone, gameplay-triggered advancements. See the datapack-emulator skill.

Other

Prompts and resources

🧭 Skills

Six guides written for AI agents live in skills/, one folder each with a SKILL.md:

Claude Code loads them as skills when the repository is installed as a plugin (this also registers the MCP server from .mcp.json, so minecode must be on PATH — see Installation):

/plugin marketplace add AnCarsenat/minecode-mcp
/plugin install minecode@minecode

Every other MCP client gets the same text through the server: the get_minecode_guide tool (agents call it themselves — minecraft_start_session points at it), a guide_<name> prompt, and a minecode://guide/<name> resource. The wheel carries skills/ inside the package, so a plain pip install minecode-mcp has them too.

To add a guide: create skills/<name>/SKILL.md with name and description frontmatter (the description is what an agent chooses by — say when to use it). tests/test_guides.py checks it is found and served.

💡 Example prompts

"Set up my datapack for 1.21.4 and tell me what changed since 1.20.4"

"Why does my datapack do nothing on 1.21?"

"What's the correct /give syntax with enchantments for this pack's version?"

"Convert this 1.20.4 loot table to 1.21.4"

"Check my Minecraft logs for errors"

🧠 How version knowledge works

Two layers, deliberately:

The curated table (minecode/knowledge/migrations.json) holds ~16 breaking changes as concrete before/after code pairs — the ones where a model's training data actively fights the correct answer. It's small, offline, instant, and every entry carries a verify_with field naming the tool that confirms it. It is a fast first-pass signal, never an authority.

The changelog (misode/technical-changes) is exhaustive and community-maintained across every snapshot. get_technical_changes queries it live.

This split is on purpose. A hand-written document covering every version's changes would be stale the day it was written, impossible to keep current against Minecraft's snapshot cadence, and far too large to fit in context. Keeping the curated layer small and querying the maintained source for everything else is what makes it sustainable.

Adding a migration

Add an entry to migrations.json:

{
  "id": "kebab-case-id",
  "title": "Short description",
  "changed_in": "1.21.5",
  "affects": ["give", "item"],
  "severity": "breaking",
  "confidence": "high",
  "before": "the old syntax",
  "after": "the new syntax",
  "explanation": "What changed and what happens if you get it wrong.",
  "detect": [
    {"pattern": "regex", "kind": "command|json|path|any", "message": "What to do instead"}
  ],
  "verify_with": "get_technical_changes(from_version='1.21.4', to_version='1.21.5')"
}

Then add tests to tests/test_knowledge.py — one for detection and one for the false-positive case. A checker that flags correct modern syntax trains the agent to ignore it, which is worse than having no checker at all.

🧑‍💻 Development

Setup

Linux / macOS

git clone --recurse-submodules https://github.com/AnCarsenat/minecode-mcp.git
cd minecode-mcp
python3 -m venv venv
source venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ./datapack-emulator    # optional: the emulator from the submodule, to change both
python -m pip install -e ".[dev]"

Already cloned without --recurse-submodules: git submodule update --init. Skipping the submodule is fine: pip install -e ".[dev]" installs the emulator from PyPI. Install the submodule first only to work on both at once.

Windows (PowerShell)

git clone --recurse-submodules https://github.com/AnCarsenat/minecode-mcp.git
cd minecode-mcp
py -m venv venv
.\venv\Scripts\Activate.ps1
py -m pip install --upgrade pip
py -m pip install -e .\datapack-emulator
py -m pip install -e ".[dev]"

If activation is blocked: Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

The -e (editable) install means source edits take effect immediately — no reinstall between changes. But your MCP client must point at this venv's interpreter, not a system one, or you'll be testing the published package instead of your working copy.

Everyday workflow

pytest -m "not network"    # ~0.4s — run before every commit
python -m minecode.server  # smoke test; "Registered N tools" then a hang is correct

pytest -m network          # live API tests, hits volunteer-run services
pytest tests/test_knowledge.py -v          # one file
pytest -k "migration" -v                   # by name
pytest -m "not network" --lf               # only last-failed

Please don't run the network suite in a loop — Spyglass, misode and minecraft.wiki are volunteer-funded. The offline suite covers the logic; the network suite only checks that upstream response shapes haven't changed.

Testing against a real datapack

The most valuable check, and how the redirect and backtracking bugs were found:

python - <<'EOF'
import pathlib
from minecode import handlers

PACK = pathlib.Path("path/to/your/datapack")
info = handlers.handle_minecraft_start_session(str(PACK))
version = info["target_version"]
print("target:", version, "| multi-version:", info["multi_version"])

print("structure:", handlers.handle_check_pack_structure(str(PACK))["issue_count"], "issues")

bad = 0
for f in PACK.rglob("*.mcfunction"):
    for n, line in enumerate(f.read_text(errors="ignore").splitlines(), 1):
        line = line.strip()
        if not line or line.startswith(("#", "$")):
            continue
        result = handlers.handle_validate_command(line, version)
        if not result.get("valid"):
            bad += 1
            print(f"{f.name}:{n} {line[:80]}\n   -> {result.get('error')}")
print("invalid commands:", bad)
EOF

A false positive here is a bug worth reporting — a validator that cries wolf gets ignored, which is worse than having none.

Environment variables

Project layout

server.py is wiring only (transport, dispatch, prompts, resources). tools.py holds schemas plus the name→handler registry. handlers.py holds behaviour. scrappers/ talks to the outside world. Nothing else should make HTTP calls.

Adding a tool

1. Write handle_<name> in handlers.py. Return a dict with success, never a bare string.

2. Add a Tool to TOOLS in tools.py.

3. Add the entry to HANDLERS in the same file.

4. pytest -m "not network"

Step 3 is not optional and not forgettable — tools.py asserts at import time that TOOLS and HANDLERS match exactly, so a missing entry fails immediately rather than months later. That assertion exists because four working changelog functions sat unreachable in misode.py for exactly that reason.

Description guidance, learned from what actually goes wrong:

• Say when to call it, not just what it does. Agents match situation to description.

• Put limitations first. A caveat at the end isn't read in time to change the decision.

• Name the better tool when one exists — "use X instead for Y" prevents the wrong choice a neutral description invites.

Adding a version migration

See How version knowledge works. Every entry needs two tests: one proving detection fires, one proving it does not fire on correct modern syntax.

Code conventions

• Handlers return dicts; the dispatcher does the JSON encoding

• Every version parameter goes through packmeta.resolve_version first

• Scrapers go through cache.cached_fetch

• Never return [] for a failure — an empty list reads as a real "none found" answer. Raise instead

• Report truncation explicitly. A silently capped list reads as complete

On the mcp dependency: pinned to >=1.25.0,<2. mcp 2.0 removed the low-level decorator API this server is built on; installing 2.x raises AttributeError at import. Migrating to the 2.x MCPServer API is open work — PRs welcome.

Contributing

Branch off main, keep commits scoped to one concern, run pytest -m "not network" before pushing. CI runs the offline suite on Python 3.10/3.11/3.12 for every PR; live tests run nightly.

📦 PyPI publishing

Publishing uses Trusted Publishing (OIDC). There is no API token anywhere — no PYPI_API_TOKEN secret to create, paste, rotate, or leak. GitHub proves its identity to PyPI directly.

One-time setup

1. Create the GitHub environment

Repo → Settings → Environments → New environment → name it exactly pypi.

Optionally add yourself under "Required reviewers". That makes every publish need a manual click — a good safety net, since a tag push would otherwise publish immediately and a version number burned on PyPI can never be reused.

2. Register the publisher on PyPI

Log in at pypi.org.

• If minecode-mcp already exists: go to the project → Manage → Publishing.

• For a brand-new project: Account settings → Publishing → Add a pending publisher.

Fill in exactly these values:

The workflow filename and environment name must match character for character. This is the most common place setup goes wrong, and the resulting error is an opaque 403.

That's it. No token is generated and nothing is pasted into GitHub.

Releasing

⚠️ Step 1 is bumping the version in pyproject.toml. Do not skip it.

A version number on PyPI is permanent. Once 0.2.0 is published, that number can never be reused or overwritten — even if you delete the release. A bad publish can only be followed by a new version, never a replacement. Re-tagging an already-published version fails at the upload step.

1. Bump the version. Edit version in pyproject.toml:

version = "0.2.1"   # was 0.2.0

Which digit to move:

(0.2.0 was a minor bump because it removed two tools and renamed one.)

2. Commit, tag, and push. The tag must be v + the exact version:

git add pyproject.toml
git commit -m "Release 0.2.1"
git tag v0.2.1
git push origin main --tags

3. Approve the deployment. The workflow builds, then pauses. Go to the Actions tab → the running Publish to PyPI run → Review deployments → tick pypi → Approve and deploy. GitHub also emails you an approve link.

Publishing takes about 30 seconds after approval.

What the workflow checks before it will publish

1. The tag matches pyproject.toml — this is the safety net for a forgotten bump. Tagging v0.2.1 while pyproject.toml still says 0.2.0 fails the build with Tag v0.2.1 does not match pyproject.toml version 0.2.0, and nothing is published

2. Offline test suite passes

3. Wheel and sdist build

4. twine check on the metadata

5. The preprompt, config, and migration table are actually inside the wheel

If any of these fail, the approval button never appears — there is nothing to approve.

If you forgot to bump

The build fails at step 1 and nothing ships. Recover by deleting the tag, bumping properly, and re-tagging:

git tag -d v0.2.1
git push origin :refs/tags/v0.2.1
# bump pyproject.toml, commit, then tag again

This is safe precisely because nothing was published.

Optional: TestPyPI first

Register a second pending publisher at test.pypi.org with the same values but environment testpypi, create a matching GitHub environment, then add this job to publish.yml:

  publish-test:
    needs: build
    if: github.ref == 'refs/heads/main'
    runs-on: ubuntu-latest
    environment:
      name: testpypi
    permissions:
      id-token: write
    steps:
      - uses: actions/download-artifact@v4
        with: { name: dist, path: dist/ }
      - uses: pypa/gh-action-pypi-publish@release/v1
        with:
          repository-url: https://test.pypi.org/legacy/
          skip-existing: true

This catches packaging errors before they reach real PyPI, where a version number is burned permanently — you can't re-upload 0.2.0 after a bad publish, only bump to 0.2.1.

Troubleshooting

📁 Project structure

minecode-mcp/
├── minecode/
│   ├── server.py              # Transport, dispatch, prompts, resources
│   ├── guides.py              # Serves skills/ as a tool, prompts, resources
│   ├── tools.py               # Tool schemas + name->handler registry
│   ├── handlers.py            # Tool behaviour
│   ├── brigadier.py           # Command tree rendering and validation
│   ├── packmeta.py            # pack.mcmeta reading, version resolution
│   ├── cache.py               # Disk cache
│   ├── knowledge/
│   │   ├── __init__.py        # Version comparison, syntax checking
│   │   └── migrations.json    # Curated breaking changes
│   ├── preprompts/
│   │   └── assistant_preprompt.txt
│   ├── config/
│   └── scrappers/
│       ├── spyglass.py        # Version-exact registries, commands, mcdoc
│       ├── misode.py          # Vanilla presets + technical changelogs
│       ├── minecraftwiki.py   # Wiki (latest version only)
│       ├── mojira.py          # Bug tracker
│       └── minecraft_logs.py  # Multi-launcher log reader
├── skills/<name>/SKILL.md     # Agent guides (Claude Code skills; also served over MCP)
├── .claude-plugin/            # Claude Code plugin + marketplace manifest
├── tests/
├── example/crystal_dimension/
└── pyproject.toml

🌐 Data sources

Spyglass, misode, and the wiki are volunteer-run. MineCode caches aggressively — version-pinned data permanently, since it cannot change — to keep request volume low. Please don't disable the cache in automated setups.

📄 License

GPL-3.0 — see LICENSE