MCP Tools

Five tools for LLM-assisted Airflow→Prefect conversion. The LLM reads raw code, looks up translation knowledge, and generates Prefect flows.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      MCP Client (LLM)                           │
├─────────────────────────────────────────────────────────────────┤
│  1. read_dag(path)              → Raw DAG source code           │
│  2. lookup_concept(name)        → Airflow→Prefect mappings      │
│  3. search_prefect_docs(query)  → Live Prefect documentation    │
│  4. LLM generates complete Prefect flow                         │
│  5. validate(original, flow)    → Syntax check + both sources   │
│  6. scaffold(directory)         → Project structure              │
└─────────────────────────────────────────────────────────────────┘

Tools

read_dag

Read an Airflow DAG file and return the raw source code with metadata.

Input:

{
  "path": "dags/etl_pipeline.py",
  "content": "# optional: inline DAG code"
}

Output:

• source: The raw DAG source code
• file_path: Resolved file path
• file_size_bytes: File size
• line_count: Number of lines

The LLM reads the source directly — no AST intermediary.

lookup_concept

Query pre-compiled Airflow→Prefect translation knowledge. Covers operators, patterns, connections, and core concepts. Backed by 78 Colin-compiled entries with built-in fallback mappings.

Input:

{
  "concept": "PythonOperator"
}

Output includes:

• concept_type: operator, pattern, connection, or concept
• Airflow details and Prefect equivalent
• translation_rules: How to convert
• source: "colin" (pre-compiled) or "fallback" (built-in)

search_prefect_docs

Search current Prefect documentation via the Prefect MCP server. For real-time queries beyond what Colin pre-compiled.

Input:

{
  "query": "task retries and caching"
}

Output: Search results from docs.prefect.io, or an error with fallback guidance.

validate

Syntax-check generated Prefect code and return both sources for structural comparison.

Input:

{
  "original_dag": "dags/etl_pipeline.py",
  "converted_flow": "# generated Prefect code..."
}

Output:

• original_source: The original DAG code
• converted_source: The generated flow code
• syntax_valid: Whether the generated code parses
• syntax_errors: Line/column/message if invalid
• comparison_guidance: Checklist for the LLM to verify fidelity

scaffold

Generate a Prefect project directory structure following prefecthq/flows conventions. Does not generate flow code.

Input:

{
  "output_directory": "/path/to/project",
  "project_name": "my-project",
  "workspace": "default",
  "flow_names": ["my_flow"],
  "include_docker": true,
  "include_github_actions": true,
  "schedule_interval": "@daily"
}

Output: Created directories, files, prefect.yaml template, and next steps.

generate_deployment

Write a prefect.yaml deployment configuration from DAG metadata. Call after generating flow.py. Produces a complete prefect.yaml with YAML anchors, schedule config, parameter defaults, and TODO stubs for work pool and pull step configuration.

Input:

{
  "output_directory": "/path/to/project",
  "flows": [
    {
      "flow_name": "my_flow",
      "entrypoint": "deployments/default/my_flow/flow.py:my_flow",
      "schedule": "0 6 * * *",
      "parameters": {"date": null},
      "tags": ["etl"]
    }
  ],
  "workspace": "default"
}

Output: JSON with created_file, deployment_names, and next_steps.

generate_migration_report

Write MIGRATION.md — a human-readable record of a DAG conversion. Call as the final step after generate_deployment. Documents every conversion decision, produces a before-production checklist with Prefect doc links, and suggests adding the Prefect MCP server.

Input:

{
  "output_directory": "/path/to/project",
  "dag_path": "dags/etl_pipeline.py",
  "flow_path": "deployments/default/my_flow/flow.py",
  "decisions": [
    {
      "component": "PythonOperator",
      "outcome": "converted",
      "rationale": "Mapped to @task decorator"
    }
  ],
  "manual_actions": ["setup_work_pool", "migrate_connections"]
}

Output: JSON with created_file and checklist_items_count.

External MCP Integration

The search_prefect_docs tool connects to the Prefect MCP server via FastMCP Client:

Configure via environment:

export MCP_PREFECT_ENABLED=true
export MCP_PREFECT_URL="https://docs.prefect.io/mcp"

Next Steps

• Tool Design — Design principles
• Examples — Usage examples
• Schemas — Input/output schemas