MCP: Prefect
by Prefect
Prefect's official read-only MCP server for inspecting flow runs, deployments and automations on Prefect Cloud or self-hosted, with a docs proxy for current CLI and SDK guidance.
Astro Airflow MCP Server
by Astronomer
Released
Apache-2.0
Summary
Astronomer's MCP server for Apache Airflow: 30+ tools over the Airflow REST API, with consolidated helpers that explore a DAG, diagnose a failed run or summarise system health in a single call. Works with Airflow 2.x and 3.x.
Features
- Consolidated explore_dag, diagnose_dag_run and get_system_health tools answer in one call what would otherwise take five
- 30+ fine-grained tools for DAGs, tasks, pools, Variables, Connections, assets, plugins and providers
- MCP resources for version, providers, plugins and config, plus troubleshooting and onboarding prompts
- Automatic Airflow version detection; supports both Airflow 2.x and 3.x
- Runs standalone via uvx, or as an Airflow 3.x plugin for a single shared deployment
- Basic-auth or bearer-token authentication; stdio or HTTP transport
Installation
Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.
Description
Debugging a failed Airflow run is a scavenger hunt through the UI — find the DAG, find the run, find the failed task, open the log, check the connection, check the pool. The Astro Airflow MCP Server is Astronomer's Model Context Protocol server for Apache Airflow, and its most useful design decision is that it does not stop at wrapping one endpoint per tool.
Consolidated tools, because agents pay per round trip
Alongside 30+ fine-grained tools it registers three agent-optimised ones:
- `explore_dag` — everything about a DAG in one response instead of five calls.
- `diagnose_dag_run` — the failed-run investigation, assembled: the run, its tasks, and the failure context.
- `get_system_health` — a system overview with health metrics.
That shape matters because each extra tool call is latency, tokens and another chance for the agent to lose the thread. A single diagnose_dag_run gets an assistant to a hypothesis in one hop.
Everything else it covers
The fine-grained tools span DAG operations (list, details, source code, statistics, trigger, pause and unpause), task management, pools, Variables and Connections, assets and datasets, plugin and provider inventories, and configuration. It also publishes MCP resources — airflow://version, airflow://providers, airflow://plugins, airflow://config — and MCP prompts for troubleshooting workflows, health checks and onboarding.
Deployment
It is a Python package (3.10+) on PyPI and runs with no install step:
{
"mcpServers": {
"airflow": { "command": "uvx", "args": ["astro-airflow-mcp", "--transport", "stdio"] }
}
}Point it at a webserver with AIRFLOW_API_URL plus either AIRFLOW_USERNAME/AIRFLOW_PASSWORD or AIRFLOW_AUTH_TOKEN. It detects the Airflow version automatically and supports both 2.x and 3.x, and on Airflow 3.x it can run as an Airflow plugin instead of a standalone process — one deployment for the whole team rather than credentials on every laptop. MCP_TRANSPORT, MCP_HOST and MCP_PORT switch it to HTTP.
Who it is for
Data engineers on call for a scheduler. The read paths (explore, diagnose, health) are where it earns its keep; triggering and pausing DAGs is available, so give the API user the grants you would give a junior on-call engineer and no more.
Apache 2.0, maintained by Astronomer.
Related MCP Servers
See all Astro Airflow MCP Server alternatives
The official W&B MCP server: query experiment runs, Weave LLM traces, artifacts and registries in natural language, and write findings back as a W&B report.
Nutanix's official open-source MCP server: 1,000+ Prism Central V4 API operations across 19 namespaces, read-only by default and audit-logged.
MCP: Fivetran MCP Server
by Fivetran
Fivetran's own MCP server exposes the Fivetran REST API to agents as discoverable tools, so an assistant can check sync status, inspect connections and destinations, and trigger syncs from chat.