airflow-unfactor

An MCP server that converts Apache Airflow DAGs into Prefect flows. Point it at a DAG, and the LLM generates idiomatic Prefect code. Not a template with TODOs — working code. Built with FastMCP.

Install

Claude Code — one line:

claude mcp add airflow-unfactor -- uvx airflow-unfactor

Claude Desktop and other clients — see manual config below.

Then ask your LLM: "Convert the DAG in dags/my_etl.py to a Prefect flow."

How It Works

The server exposes seven tools over MCP. The LLM reads raw DAG source code, looks up translation knowledge, and generates the Prefect flow.

No AST parsing. No template engine. The LLM reads the code directly, just like a developer would.

Manual config

The buttons above and the claude mcp add command both register the server with uvx, which downloads it on first run — no separate pip install needed. To install the package directly anyway: pip install airflow-unfactor or uv pip install airflow-unfactor.

{
  "mcpServers": {
    "airflow-unfactor": {
      "command": "uvx",
      "args": ["airflow-unfactor"]
    }
  }
}

{
  "mcpServers": {
    "airflow-unfactor": {
      "command": "uvx",
      "args": ["airflow-unfactor"]
    }
  }
}

{
  "mcpServers": {
    "airflow-unfactor": {
      "command": "uvx",
      "args": ["airflow-unfactor"]
    }
  }
}

Example

Airflow DAG:

from airflow import DAG
from airflow.operators.python import PythonOperator

def extract():
    return {"users": [1, 2, 3]}

def transform(ti):
    data = ti.xcom_pull(task_ids="extract")
    return [u * 2 for u in data["users"]]

with DAG("my_etl", ...) as dag:
    t1 = PythonOperator(task_id="extract", python_callable=extract)
    t2 = PythonOperator(task_id="transform", python_callable=transform)
    t1 >> t2

Generated Prefect flow:

from prefect import flow, task

@task
def extract():
    return {"users": [1, 2, 3]}

@task
def transform(data):
    return [u * 2 for u in data["users"]]

@flow(name="my_etl")
def my_etl():
    data = extract()
    result = transform(data)
    return result

The >> dependency chain becomes explicit data passing through return values. XCom is gone. It's just Python.

Translation Knowledge

The server ships with 78 pre-compiled Airflow→Prefect translation entries covering operators, patterns, connections, and core concepts. These are compiled by Colin from live Airflow source and Prefect documentation.

When the pre-compiled knowledge doesn't cover something, search_prefect_docs queries the Prefect documentation MCP server at docs.prefect.io in real time.

Documentation

Full docs: gabcoyne.github.io/airflow-unfactor

Development

git clone https://github.com/gabcoyne/airflow-unfactor.git
cd airflow-unfactor
uv sync

# Run tests
uv run pytest

# Lint
uv run ruff check --fix

# Compile translation knowledge
cd colin && colin run

License

MIT — see LICENSE.