cwprep is a Python-based engine that enables Text-to-PrepFlow generation.

By reverse-engineering the .tfl JSON structure and providing a built-in MCP (Model Context Protocol) server, cwprep acts as a bridge between LLMs (like Claude, Gemini) and Tableau Prep. You can now generate, modify, and build data cleaning flows simply through natural language conversations or Python scripts, without ever opening the GUI!

Author: Cooper Wenhua <imgwho@gmail.com>

pip install cwprep

from cwprep import TFLBuilder, TFLPackager

# Create builder
builder = TFLBuilder(flow_name="My Flow")

# Add database connection
conn_id = builder.add_connection(
    host="localhost",
    username="root",
    dbname="mydb"
)

# Add input tables
orders = builder.add_input_table("orders", "orders", conn_id)
customers = builder.add_input_table("customers", "customers", conn_id)

# Join tables
joined = builder.add_join(
    name="Orders + Customers",
    left_id=orders,
    right_id=customers,
    left_col="customer_id",
    right_col="customer_id",
    join_type="left"
)

# Add output
builder.add_output_server("Output", joined, "My_Datasource")

# Build and save
flow, display, meta = builder.build()
TFLPackager.save_tfl("./my_flow.tfl", flow, display, meta)

By default, both the SDK and MCP output only the final .tfl/.tflx archive. Use save_to_folder() only when you explicitly want the exploded folder for inspection.

See the examples/ directory for complete demos:

• demo_basic.py - Input, Join, Output
• demo_cleaning.py - Filter, Calculate, Rename
• demo_field_operations.py - Quick Calc, Change Type, Duplicate Column
• demo_aggregation.py - Union, Aggregate, Pivot
• demo_comprehensive.py - All features combined
• prompts.md - 8 ready-to-use MCP prompt templates for AI-driven flow generation

cwprep includes a built-in Model Context Protocol server, enabling AI clients (Claude Desktop, Cursor, Gemini CLI, etc.) to generate TFL files directly.

# Local (stdio)
cwprep-mcp

# Remote (Streamable HTTP)
cwprep-mcp --transport streamable-http --port 8000

uv cache clean cwprep

All clients below use the uvx method (recommended). Replace uvx with cwprep-mcp if you prefer a local pip install.

{
  "mcpServers": {
    "cwprep": {
      "command": "uvx",
      "args": ["--from", "cwprep", "cwprep-mcp"]
    }
  }
}

{
  "mcpServers": {
    "cwprep": {
      "command": "uvx",
      "args": ["--from", "cwprep", "cwprep-mcp"]
    }
  }
}

{
  "servers": {
    "cwprep": {
      "command": "uvx",
      "args": ["--from", "cwprep", "cwprep-mcp"]
    }
  }
}

{
  "mcpServers": {
    "cwprep": {
      "command": "uvx",
      "args": ["--from", "cwprep", "cwprep-mcp"]
    }
  }
}

claude mcp add cwprep -- uvx --from "cwprep" cwprep-mcp

{
  "mcpServers": {
    "cwprep": {
      "command": "uvx",
      "args": ["--from", "cwprep", "cwprep-mcp"]
    }
  }
}

mcpServers:
  - name: cwprep
    command: uvx
    args:
      - --from
      - cwprep
      - cwprep-mcp

cwprep-mcp --transport streamable-http --port 8000

This project includes a specialized AI Skill for assistants like Claude or Gemini to help you build flows.

• Location: .agents/skills/tfl-generator/
• Features: MCP server index with fallback SDK usage guide. Detailed API and syntax references are served via MCP Resources from src/cwprep/references/.

cwprep/
├── .agents/skills/      # AI Agent skills (MCP index)
├── src/cwprep/          # SDK source code
│   ├── builder.py       # TFLBuilder class
│   ├── packager.py      # TFLPackager class
│   ├── translator.py    # SQLTranslator class
│   ├── expression_translator.py  # ExpressionTranslator class
│   ├── config.py        # Configuration utilities
│   ├── mcp_server.py    # MCP Server (Tools, Resources, Prompts)
│   └── references/      # MCP Resource documents (.md)
├── examples/            # Demo scripts
├── docs/                # Documentation
└── tests/               # Unit tests

Create config.yaml for default settings:

# MySQL (default)
database:
  host: localhost
  port: 3306
  dbname: mydb
  type: mysql

# SQL Server (Windows Authentication)
# database:
#   host: localhost
#   type: sqlserver
#   authentication: sspi
#   schema: dbo

# PostgreSQL
# database:
#   host: localhost
#   port: 5432
#   dbname: mydb
#   type: postgres

tableau_server:
  url: http://your-server
  default_project: Default

See changelog.md for version history.

AGPL-3.0 License