This project implements an MCP server for Apache Airflow, enabling users to interact with their orchestration platform using natural language.
With a few minutes of setup, you should be able to use Claude Desktop or any MCP-enabled LLM to ask questions like:
- "What DAGs do we have in our Airflow cluster?"
- "What is our latest failed DAG?"
- "Show me all versions of the transform_sales_aggregator DAG"
- "Which assets are produced by our load DAGs?"
The Model Context Protocol (MCP) is an open standard creating secure connections between data sources and AI applications. This repository provides a custom MCP server for Apache Airflow that transforms how teams interact with their orchestration platform through natural language.
Supports Airflow 2 and Airflow 3. Set airflow_version=3 in your environment to enable the Airflow 3 API (/api/v2) and JWT authentication automatically.
- Query pipeline statuses through natural language
- Troubleshoot DAG failures efficiently
- Retrieve comprehensive DAG information
- Trigger DAGs based on their status
- Monitor execution results
- Analyze DAG components and configurations
- Browse Airflow variables, pools, and connections
- Inspect task instances for any DAG run
- Airflow 3: DAG version history, Asset (Dataset) catalog
If you already have an Airflow instance and want to use our prebuilt Docker image, you only need:
- Docker
- Access to your Apache Airflow instance (Airflow 2 or Airflow 3)
- LLM access (Claude, ChatGPT, or AWS Bedrock)
This repository also provides a local setup for Apache Airflow, which you can use for demo purposes.
Configure Claude Desktop to connect to your Airflow instance. If you haven't configured Claude Desktop for use with MCP before, follow the Claude Desktop documentation.
Airflow 2:
{
"mcpServers": {
"airflow_mcp": {
"command": "docker",
"args": ["run", "-i", "--rm",
"-e", "airflow_base_url",
"-e", "airflow_username",
"-e", "airflow_password",
"-e", "airflow_version",
"pondered/airflow-mcp:latest"],
"env": {
"airflow_base_url": "http://host.docker.internal:8088",
"airflow_username": "airflow",
"airflow_password": "airflow",
"airflow_version": "2"
}
}
}
}Airflow 3 (uses /api/v2 and JWT authentication automatically):
{
"mcpServers": {
"airflow_mcp": {
"command": "docker",
"args": ["run", "-i", "--rm",
"-e", "airflow_base_url",
"-e", "airflow_username",
"-e", "airflow_password",
"-e", "airflow_version",
"pondered/airflow-mcp:latest"],
"env": {
"airflow_base_url": "http://host.docker.internal:8080",
"airflow_username": "airflow",
"airflow_password": "airflow",
"airflow_version": "3"
}
}
}
}-
Clone this repository:
git clone https://github.com/ponderedw/airflow-mcp -
If you don't have a running Airflow environment, start one with:
just airflowThis will start an Airflow instance on port 8088, with username
airflowand passwordairflow.You can access Airflow at http://localhost:8088/ and see multiple DAGs configured:
These DAGs have complex dependencies, some running on a schedule and some using Airflow's Dataset/Asset functionality.
-
Configure Claude Desktop as shown in the Quick Start section above.
-
Test your setup by asking Claude: "What DAGs do we have in our Airflow cluster?"
| Variable | Required | Description |
|---|---|---|
airflow_base_url |
Yes | Base URL of your Airflow instance (e.g. http://localhost:8080) |
airflow_username |
Yes | Airflow username |
airflow_password |
Yes | Airflow password |
airflow_version |
No | 2 (default) or 3 β selects API version and auth method |
POST_MODE |
No | true to enable DAG trigger and pause/unpause tools (default: false) |
AIRFLOW_INSIGHTS_MODE |
No | true to enable schedule-prediction tools (requires the schedule-insights plugin) |
TOKEN |
No | Bearer token to protect the MCP server's own SSE endpoint |
TRANSPORT_TYPE |
No | sse or stdio (default: stdio) |
AIRFLOW_ASSISTENT_AI_CONN |
No | Alternative connection string: username:password@base_url |
| Tool | Description |
|---|---|
get_dags |
List all DAGs |
get_dag |
Full details for a single DAG |
get_dags_tasks |
All tasks defined in a DAG |
get_dags_task |
Details for a single task |
get_dags_script |
Source code of a DAG file |
get_all_the_runs_for_dag |
25 most recent runs for a DAG |
get_all_the_runs |
Filtered runs across multiple DAGs |
get_task_instances |
Task instances for a specific DAG run |
get_connections |
All Airflow connections |
get_variables |
All Airflow variables |
get_variable |
A single Airflow variable by key |
get_pools |
All resource pools and slot usage |
| Tool | Description |
|---|---|
trigger_dag |
Trigger a new DAG run |
change_dags_pause_status |
Pause or unpause one or all DAGs |
| Tool | Description |
|---|---|
get_dag_versions |
All recorded versions of a DAG |
get_dag_version |
A specific DAG version by number |
get_assets |
All Assets (formerly Datasets) |
| Tool | Description |
|---|---|
insights_get_dags_next_run |
Predicted next run time for a DAG |
Requires the airflow-schedule-insights plugin installed in your Airflow environment.
- "What DAGs do we have in our Airflow cluster?"
- "Identify all DAGs with failed status in their most recent execution and trigger a new run for each one"
- "What operators are used by the transform_forecast_attendance DAG?"
- "Has the transform_forecast_attendance DAG ever completed successfully?"
- "Show me all versions of the load_ticket_sales DAG and what changed"
- "Which assets does the load_box_office_reports DAG produce?"
- "What variables are configured in Airflow?"
You can use from langchain_mcp_adapters.client import MultiServerMCPClient to add Airflow MCP as one of your tools in a LangChain app.
mcp_host = os.environ.get('mcp_host', 'mcp_sse_server:8000')
mcps = {
"AirflowMCP": {
"url": f"http://{mcp_host}/sse",
"transport": "sse",
"headers": {"Authorization": f"Bearer {os.environ.get('MCP_TOKEN')}"}
}
}Install the library first: pip install airflow-mcp-ponder
mcps = {
"AirflowMCP": {
'command': "python",
'args': ["-m", "airflow_mcp_ponder.mcp_airflow"],
"transport": "stdio",
'env': {k: v for k, v in {
'AIRFLOW_ASSISTENT_AI_CONN': os.getenv('AIRFLOW_ASSISTENT_AI_CONN'),
'airflow_base_url': os.getenv('airflow_base_url'),
'airflow_username': os.getenv('airflow_username'),
'airflow_password': os.getenv('airflow_password'),
'airflow_version': os.getenv('airflow_version'),
'AIRFLOW_INSIGHTS_MODE': os.getenv('AIRFLOW_INSIGHTS_MODE'),
'POST_MODE': os.getenv('POST_MODE'),
'TRANSPORT_TYPE': 'stdio',
}.items() if v is not None}
}
}client = MultiServerMCPClient(mcps)
tools = await client.get_tools()We enthusiastically invite the community to contribute to this open-source initiative! Whether you're interested in:
- Adding new features
- Improving documentation
- Enhancing compatibility with different LLM providers
- Reporting bugs
- Suggesting improvements
Please feel free to submit pull requests or open issues on our GitHub repository.
