MCP server that gives LLMs efficient access to the Pywr-DRB water resource model codebase. Uses static analysis (Python ast) to extract structure and data from source files without importing pywrdrb or its heavy dependencies.

• uv installed
• Pywr-DRB source code cloned locally (the src/pywrdrb/ directory is read at startup)

uv run python -m pywrdrb_mcp.server

uv run pytest tests/ -v

The server reads the Pywr-DRB source tree via static analysis (Python ast) at startup. It never imports or executes pywrdrb code, so pywrdrb's heavy dependencies (numpy, pandas, pywr, etc.) are not required.

Set PYWRDRB_ROOT if your Pywr-DRB checkout is not in the default sibling location:

# Linux/macOS
export PYWRDRB_ROOT=/path/to/Pywr-DRB/src/pywrdrb

# Windows (PowerShell)
$env:PYWRDRB_ROOT = "C:\path\to\Pywr-DRB\src\pywrdrb"

claude mcp add --transport stdio pywrdrb-mcp -- \
  uv run --directory /path/to/pywrdrb-mcp python -m pywrdrb_mcp.server

Add to .claude/settings.json (project or user level):

{
  "mcpServers": {
    "pywrdrb-mcp": {
      "type": "stdio",
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/pywrdrb-mcp",
        "python", "-m", "pywrdrb_mcp.server"
      ]
    }
  }
}

Edit the Claude Desktop config file:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

(Open via Claude Desktop: Settings > Developer > Edit Config)

{
  "mcpServers": {
    "pywrdrb-mcp": {
      "command": "uv",
      "args": [
        "run",
        "--directory", "/path/to/pywrdrb-mcp",
        "python", "-m", "pywrdrb_mcp.server"
      ],
      "env": {
        "PYWRDRB_ROOT": "/path/to/Pywr-DRB/src/pywrdrb"
      }
    }
  }
}

Windows paths: Use forward slashes (C:/Users/...) or escaped backslashes (C:\\Users\\...) in the JSON.

After saving, fully quit and restart Claude Desktop (system tray > Exit, not just close the window).

In practice, this means the tools are the primary interface on Claude Desktop. The domain knowledge resources (FFMP rules, STARFIT rules, flood operations) and the how_to_* prompts work but require manual user action to invoke. In Claude Code, resources and prompts are more seamlessly integrated.

Topology & Data

API References

Domain Knowledge Guides

Project

Parameterized Templates

All prompts use the how_to_* naming convention to indicate they are instructional guides.

pywrdrb-mcp/
├── src/pywrdrb_mcp/
│   ├── server.py          # FastMCP entry point
│   ├── config.py          # PYWRDRB_ROOT path
│   ├── index/             # AST-based static analysis engine
│   │   ├── builder.py     # PywrDRBIndex — topology, lists, capacity, rating curves
│   │   ├── ast_utils.py   # Safe AST extraction utilities
│   │   └── file_utils.py  # File system utilities
│   ├── tools/             # 14 MCP tools (8 modules)
│   ├── resources/         # 22 MCP resources
│   ├── prompts/           # 6 MCP prompt templates
│   └── content/           # Hand-written domain knowledge (7 markdown files)
└── tests/

The server reads from the Pywr-DRB source tree at startup via ast.parse() and ast.literal_eval() — it never imports or executes pywrdrb code.