mcp-server

MCP server

swatplus-builder exposes a Model Context Protocol (MCP) tool server so an AI agent operates the pipeline through typed tools rather than free-form code. This is the design's intended mode: the agent operates, the package governs.

Start the server

pip install "swatplus-builder[mcp]"
swat mcp                 # stdio transport

Before starting, run a pre-flight check to confirm the server and all tools load correctly in your active Python environment:

swat mcp-check           # exits 0 if ready, 1 if not
swat mcp-check --json    # machine-readable output

This is especially useful when you have multiple Python environments (conda, venv, pyenv) — mcp-check tells you exactly which import is missing and in which environment, rather than failing silently when the server spawns.

Register it with an agent client

Claude Desktop / Claude Code (claude_desktop_config.json) — recommended form (when engine is installed via swat setup engine --path, no env vars needed):

{
  "mcpServers": {
    "swatplus-builder": {
      "command": "swat",
      "args": ["mcp"]
    }
  }
}

If you installed the engine manually via SWATPLUS_EXE instead:

{
  "mcpServers": {
    "swatplus-builder": {
      "command": "swat",
      "args": ["mcp"],
      "env": {
        "SWATPLUS_EXE": "/path/to/swatplus_exe"
      }
    }
  }
}

!!! tip "Mixed conda/venv environments" If you get ModuleNotFoundError: No module named 'mcp' when the server starts, the swat binary is resolving from a different Python than where mcp is installed. Pin the exact Python interpreter instead:

```json
{
  "mcpServers": {
    "swatplus-builder": {
      "command": "/opt/miniconda3/bin/python",
      "args": ["-m", "swatplus_builder.mcp.server"],
      "env": {
        "SWATPLUS_EXE": "/usr/local/bin/swatplus_exe",
        "SWATPLUS_BUILDER_ARTIFACTS": "/data/artifacts"
      }
    }
  }
}
```