cloudcraft-mcp

Unofficial Model Context Protocol (MCP) server for Cloudcraft — list, read, export, and build cloud architecture blueprints from Claude Desktop and other MCP clients.

Features

Thirteen tools exposed to the MCP host:

Requirements

• uv for the recommended uvx and npm launchers
• Node.js 22+ only when using the npm launcher
• Cloudcraft API key — generate one at https://app.cloudcraft.co/ → User settings → API keys

Install

Use an immutable version in client configuration so upgrades are deliberate.

uvx is delivered by the PyPI package; there is no separate uvx registry. The npm package embeds the byte-identical Python wheel and invokes it through uv, so the npm path requires both Node.js and uv. It does not download code in a postinstall hook.

For development from a checkout:

git clone https://github.com/hypark5540/cloudcraft-mcp.git
cd cloudcraft-mcp
export CLOUDCRAFT_API_KEY='your-key-here'
uv run --frozen cloudcraft-mcp   # Ctrl+C to exit

Claude Desktop integration

Install uv, then add this entry to Claude Desktop's configuration. Prefer Claude's secret storage when available; the literal below is only a portable example.

{
  "mcpServers": {
    "cloudcraft": {
      "command": "uv",
      "args": [
        "tool",
        "run",
        "--isolated",
        "--from",
        "cloudcraft-mcp==0.1.7",
        "cloudcraft-mcp"
      ],
      "env": {
        "CLOUDCRAFT_API_KEY": "your-key-here",
        "CLOUDCRAFT_ENABLE_WRITES": "false",
        "CLOUDCRAFT_ENABLE_DELETES": "false"
      }
    }
  }
}

Restart Claude Desktop. The Developer tab should show cloudcraft as connected.

See client setup for Cursor, Gemini CLI, npm, Docker, and MCPB examples.

Environment variables

Usage examples (in Claude)

Once the server is connected, ask Claude things like:

"List my Cloudcraft blueprints and summarize what each represents."

"Export blueprint f0086b32-... as PNG and save it to my Desktop."

"Take the architecture I just designed and create a new Cloudcraft blueprint called 'Prod 2026'."

"Snapshot ap-northeast-2 of my connected AWS account, excluding security groups."

"Export the monthly budget of blueprint f0086b32-... as CSV in EUR."

Blueprint payload shape

create_blueprint / update_blueprint accept the full Cloudcraft data object. A minimal payload:

{
  "grid": "infinite",
  "projection": "isometric",
  "theme": {"base": "light"},
  "version": 6,
  "nodes": [
    {"id": "...", "type": "ec2", "mapPos": [3, 3], "region": "ap-northeast-2",
     "instanceType": "m7g", "instanceSize": "large", "platform": "linux"},
    {"id": "...", "type": "s3",  "mapPos": [1, 8], "region": "ap-northeast-2",
     "volumeType": "Standard", "dataGb": 100}
  ],
  "edges": [
    {"from": "...ec2-id...", "to": "...s3-id...", "type": "edge",
     "width": 2, "dashed": false, "endCap": "arrow"}
  ],
  "groups": [], "surfaces": [], "text": [], "icons": [],
  "connectors": [], "images": [], "disabledLayers": [],
  "shareDocs": false
}

Refer to Cloudcraft's API docs for the full node-type catalog and service-specific fields.

Development

uv sync --extra dev
uv run pytest            # unit tests (no network)
uv run ruff check .      # lint
uv run mypy src          # type check

Tests mock the HTTP layer with respx so no API key is required.

Coverage

CI uploads coverage.xml from the Python 3.12 matrix cell to Codecov. The project gate is 80% or higher — a PR that drops overall or patch coverage by more than 1 percentage point below that line fails the Codecov check (codecov.yml).

Run the same report locally:

uv run pytest --cov=cloudcraft_mcp --cov-report=term-missing --cov-report=xml

Project layout

cloudcraft-mcp/
├── src/cloudcraft_mcp/
│   ├── __init__.py
│   ├── __main__.py         # python -m cloudcraft_mcp
│   ├── server.py           # MCP tool definitions (FastMCP)
│   ├── client.py           # CloudcraftClient — async httpx wrapper
│   ├── types.py            # TypedDicts for blueprint payloads
│   └── py.typed
├── tests/
│   └── test_client.py
├── bin/cloudcraft-mcp.mjs # npm-to-uv launcher
├── mcpb/manifest.json     # Claude Desktop bundle metadata
├── server.json            # official MCP Registry metadata
├── package.json           # npm distribution
├── server.py               # back-compat shim -> cloudcraft_mcp.server:main
├── pyproject.toml
├── LICENSE
└── README.md

Design notes

• Transport / logic split. client.py is a plain async HTTP client you can import from scripts or CLI tools without pulling the MCP runtime. server.py only owns the MCP tool surface.
• Bearer-token auth. Cloudcraft's API expects Authorization: Bearer <key> (not Apikey). The client sets this automatically.
• No secrets in process args. The API key is read from CLOUDCRAFT_API_KEY; never pass it on the command line.
• One implementation. PyPI, npm, MCPB, and the container all launch the same versioned Python package rather than maintaining language-specific forks.
• Error surface. Non-2xx responses raise CloudcraftError with status and body preserved, re-wrapped as RuntimeError at the MCP boundary so Claude sees a readable message.

Security

• API keys grant full read / write over your Cloudcraft account. Treat them as secrets and rotate regularly.
• MCP tool annotations mark read-only tools and mutating/destructive tools so compatible hosts can present better approval prompts. These annotations are hints, not an enforcement layer.
• Environment gates and exact-ID delete confirmation are defense in depth, not an interactive approval workflow; keep the MCP host's tool approval UX enabled.
• Cloudcraft mutations are disabled by default. Set CLOUDCRAFT_ENABLE_WRITES=true only for clients that need create/update access. Deletes additionally require CLOUDCRAFT_ENABLE_DELETES=true and an exact repeated blueprint ID on every call.
• delete_blueprint is irreversible — when asking Claude to delete, be explicit about the target id.
• export_blueprint_image writes are sandboxed under CLOUDCRAFT_EXPORT_DIR (a process-private temporary subdirectory by default) and refuse to overwrite existing files unless overwrite=True.
• For read-heavy setups, create a dedicated Cloudcraft user with read-only scope (if/when Cloudcraft adds scoped keys) and use that key for MCP.
• Supported versions and private reporting instructions are in SECURITY.md; data flow is documented in PRIVACY.md.

Contributing

Issues and PRs welcome at https://github.com/hypark5540/cloudcraft-mcp. Please run ruff, mypy, and pytest before submitting.

License

MIT — see LICENSE.

Related

• Cloudcraft API documentation
• Model Context Protocol specification
• awslabs/mcp — AWS's collection of MCP servers
• modelcontextprotocol/servers — reference MCP server implementations