Dagster+ MCP Server

The Dagster+ MCP server connects an AI agent to a Dagster+ deployment over a hosted HTTP endpoint, so questions about a data platform can be answered against the live orchestrator instead of a stale runbook.

It is read-and-write. On the reading side an agent can pull runs and run logs, asset details, deployment state, alert policies, Insights metrics and Dagster+ Issues — which covers most of what someone actually does when a pipeline breaks at 3am: find the failed run, read its logs, see which assets went stale downstream. On the writing side it can launch runs and create, update or delete assets, deployments, alert policies and Issues.

Because it is a remote server at https://mcp.agent.dagster.cloud/mcp, there is nothing to install or keep updated. Authentication is OAuth by default, which means the agent inherits exactly the permissions of the user who authorised it — a meaningful safety property when the same tool can also delete a deployment. Service-user tokens are available for automation that should run under a narrower custom role, and personal user tokens for quick local setups.

Two constraints are worth knowing before adopting it. Dagster Labs ships this as a preview feature under active development, explicitly not recommended for production, with feature gaps and APIs that may change. It is also unavailable to organisations in the EU region. A Dagster+ account is required; the open-source Dagster distribution is not covered.