mcp-dagster: A Dagster MCP Server

What is mcp-dagster?

mcp-dagster is a Model Context Protocol server that enables AI agents to interact with Dagster, the data orchestration platform. It bridges LLMs and data engineering workflows by allowing agents to explore pipelines, monitor runs, and manage assets.

How to use mcp-dagster?

Install the server from PyPI (mcp-server-dagster) and run it with uv. Start a Dagster instance with your pipeline, then launch the MCP server using SSE transport (uv run examples/open-ai-agent/run_sse_mcp.py). Finally, start the agent loop with uv run ./examples/open-ai-agent/agent.py. The server connects to http://localhost:3000/graphql by default.

Key features of mcp-dagster

• list_repositories – Lists all available Dagster repositories
• list_jobs – Lists all jobs in a specific repository
• list_assets – Lists all assets in a specific repository
• recent_runs – Gets recent Dagster runs (default limit: 10)
• get_run_info – Gets detailed information about a specific run
• launch_run – Launches a Dagster job run
• materialize_asset – Materializes a specific Dagster asset
• terminate_run – Terminates an in-progress Dagster run
• get_asset_info – Gets detailed information about a specific asset

Use cases of mcp-dagster

• Explore available assets and their descriptions in a Dagster instance
• Check the status of recent runs and summarize failures
• Materialize a specific asset on demand via natural language
• Launch or terminate Dagster job runs programmatically
• Create new assets that depend on existing ones

FAQ from mcp-dagster

What is the default GraphQL endpoint?

The default endpoint is http://localhost:3000/graphql.

What transport does the server use?

It uses SSE (Server-Sent Events) transport.

How can I install mcp-dagster?

It is available on PyPI as mcp-server-dagster and can be run with uv.

What is the default limit for recent runs?

The default limit for recent_runs is 10 runs.

How do I start interacting with Dagster through mcp-dagster?

Start a Dagster instance, run the MCP server with SSE transport, then start the agent loop as shown in the quickstart.