SuperAPI AI Directory

airflow-mcp-server

Airflow REST API: list DAGs, runs and task instances, tail logs and roll up DAG health. Read-only by default; trigger and clear are write-gated.

Open website

About this tool

Airflow MCP Server is an MCP server for the Airflow 3.x REST API (/api/v2), built on top of @us-all/mcp-toolkit. It works read-only by default, reading DAGs, run records, task instances, and log tails; triggering DAGs (airflow-trigger-dag) and clearing tasks (airflow-clear-task) are write operations and are only enabled when AIRFLOW_ALLOW_WRITE=true is set.

Main capabilities: it provides 7 tools under the airflow and meta categories, plus 2 MCP prompts for DAG triage workflows. These include 1 aggregate tool that can replace the combined operation of "first list run records, then get task instances." There is also a search-tools tool for natural-language tool discovery.

Usage: authentication uses JWT (SimpleAuthManager), requiring AIRFLOW_USERNAME and AIRFLOW_PASSWORD to be provided; the server automatically issues and refreshes tokens; tokens are cached for the session lifetime and refreshed 1 minute before expiration. During configuration, only the host base address needs to be passed in; the server automatically adds /api/v2 internally; if an address ending with /api/v1 or /api/v2 is passed in, it will be stripped. The server uses MCP stdio communication and can be connected to Claude Desktop, Cursor, or any MCP client; setting MCP_TRANSPORT=http enables Streamable HTTP transport (Bearer authentication, providing a /health endpoint).

Use cases: suitable for operations and data engineering scenarios that need to view Airflow DAG health status, run records, task instances, and logs through an MCP client, and to trigger DAGs or clear tasks under controlled conditions. If deeper dbt integration is needed (manifest parsing, run result history, source freshness, column-level test coverage, lineage traversal, custom DQ result tables), the companion @us-all/dbt-mcp can be installed as well.

Limitations: verified against Airflow 3.0–3.2 /api/v2 (with 3.2.1 as the reference). v0.2+ does not support Airflow 2.x because basic auth and /api/v1 have been removed; Airflow 2.x deployments must pin @us-all/airflow-mcp@0.1.0. The project uses the MIT license.