Labsco
MCP SERVER

Airflow MCP Server

by abhishekbhakat ↗

Control Airflow 3 from an assistant through its own API — read-only by default with `--safe`, and tools grouped by category so a large API does not arrive as one flat list.

Data Platform: Pipelines, Warehousing, BI & Governance
Summary
`--safe` is the flag to reach for first, because unsafe is the default.

Started plain, the server will happily trigger, pause and modify things in your Airflow instance; `--safe` restricts it to GET requests, which is the right posture for anything pointed at production while you are still learning what the model does with it. Hierarchical discovery is the other default worth keeping — Airflow's API is large enough that a flat tool list crowds out the conversation.

What it is

An MCP server for Apache Airflow 3. It fetches the OpenAPI spec from your Airflow instance at startup and turns it into tools, so the surface tracks the API your deployment actually exposes. Two switches shape it: an operation mode that can restrict everything to read-only, and a discovery mode that groups tools by category — DAGs, tasks, connections and so on — rather than exposing all of them at once.

What you get
  • Airflow's own API surface as tools — DAGs, tasks, connections and the rest of the categories your instance exposes — generated from the spec at `${base_url}/openapi.json`
  • Safe mode with `--safe`, which allows only read-only GET operations and makes modification impossible; `--unsafe` is the default and allows everything
  • Hierarchical discovery by default: browse categories first and then pick a tool, which keeps a large API manageable; `--static-tools` exposes everything immediately for programmatic use
  • A folder of Markdown guides mounted as read-only MCP resources — each top-level `.md` or `.markdown` file becomes a `file:///<slug>` resource, titled from its first heading
  • Both transports: stdio for direct process communication, and Streamable HTTP with `--http` for web deployments and multiple clients
Requirements

An Airflow 3 instance and a JWT token — Airflow 3.0 dropped cookie and basic auth, so a valid token is the only way in. Run it with `uvx airflow-mcp-server`, passing `--base-url` and `--auth-token`, or set `AIRFLOW_BASE_URL` and `AUTH_TOKEN`. `--base-url` must be the root Airflow URL with no `/api/v2` suffix, because the server appends the OpenAPI path itself. `--http` switches to Streamable HTTP with `--port` (default 3000) and `--host` (default localhost); `--sse` still exists but is deprecated in favour of HTTP. `--resources-dir`, or `AIRFLOW_MCP_RESOURCES_DIR`, points at the Markdown guides, and the list refreshes on restart. Pagination defaults to 100 items, changed via `maximum_page_limit` in the `[api]` section of `airflow.cfg`. Published to PyPI as `airflow-mcp-server` and registered as `io.github.abhishekbhakat/airflow-mcp-server`.

Setup effort

One command plus a key — airflow-mcp-server --safe --base-url http://localhost:8080 --auth-token <jwt>, then supply credentials