Dagster+ MCP Server

The Dagster+ MCP Server connects your AI assistant to Dagster+, the managed cloud edition of Dagster. Dagster is an open-source Python tool for orchestrating data pipelines, developed on GitHub under the Apache-2.0 license since 2018. Orchestration means Dagster decides when each computation runs, in which order and what happens when something fails. MCP, the Model Context Protocol, is an open standard that lets assistants such as Claude Code or Codex work with external services. The Dagster team runs the server itself and documents it in the official docs.

Who is behind Dagster?

Dagster was built by Dagster Labs, founded by Nick Schrock, who previously helped create GraphQL at Facebook. In July 2026, Prefect Technologies acquired Dagster Labs. According to the acquisition announcement and the customer FAQ, Dagster and Dagster+ remain separate products, nobody has to move to Prefect, and pricing and contracts stay the same. That is why Skill Road lists the server under the provider Prefect, next to the equally official Prefect MCP Server.

What sets Dagster apart from other orchestrators?

Dagster does not think in individual tasks but in assets. An asset is a piece of data that a pipeline produces, such as a table in your data warehouse, a file or a trained model. Dagster knows which assets depend on each other and when each was last updated. A run is a single execution that recomputes one or more assets. Once you have that model in mind, the MCP server is easy to understand, because it works with exactly these building blocks.

What it can do, according to the docs

The documentation includes a table showing what the server may do with each object. You can view, launch and terminate runs, but not edit them afterwards. Run logs are readable. Assets can be viewed and materialized, meaning you trigger a recomputation; they cannot be edited or deleted. Deployments, the environments of your project, are read-only. Alert policies, the rules for when Dagster+ notifies you by email, Slack or Teams, can be fully created, updated and deleted, and the same goes for Dagster+ Issues. For runs, assets and deployments, the server also returns metrics from Dagster+ Insights, for example on run times and failure rates.

How does it help day to day?

The typical case is troubleshooting. Instead of clicking through the UI and the logs, you ask: Why did last night's run for the revenue table fail? The assistant reads the logs, spots, say, a renamed column in the source system and, because it can also see your code in the editor, fixes the cause right there. It then triggers a recomputation of the affected asset. Setting up alerts in plain language or getting a quick overview of which assets have been running unusually long lately is just as handy.

Setup and sign-in

The server is hosted only, so there is nothing to install. For organizations in the US region, the address is https://mcp.agent.dagster.cloud/mcp, and for the EU region it is https://mcp.agent.eu.dagster.cloud/mcp. Sign-in uses OAuth, meaning you approve access in the browser without handing over a key. The server then gets exactly your permissions in Dagster+. For clients without OAuth, there is a second route using a user token and your organization name as headers. If you want the assistant to have different permissions from your own, the documentation recommends creating a dedicated service user. Claude Code and Codex are described explicitly; Cursor and other clients use the general route via the address.

Cost and status

The server only works with Dagster+, not with self-hosted open-source Dagster. According to the pricing page, Dagster+ has no permanently free tier but paid plans with a 30-day trial, plus a Pro tier on request. The MCP server is included in every plan. The documentation calls it beta and the pricing page calls it a preview, so features may still change. According to the pricing page, Dagster+ Issues and further AI features are only part of the Pro tier, as is the EU region.

Who is the Dagster+ MCP Server for?

It suits data engineers and analytics teams who already use Dagster+ and want to monitor and fix their pipelines from the editor. Because the server can launch and terminate runs and delete alerts, you should deliberately give it only the permissions it needs. If you run Dagster yourself, you cannot use the server and are better served by the Dagster command line. For teams that use Prefect alongside Dagster, the Prefect MCP Server completes the picture.

A Dagster+ organization (US or EU region) and a client that supports remote MCP servers over HTTP, ideally with OAuth. Without OAuth, a Dagster+ user token and the organization name.

With the Dagster+ MCP Server, your AI assistant reads runs, logs and assets straight from Dagster+ and can, if you ask, launch runs or recompute assets. The Dagster team hosts the server, so there is nothing to install. This guide shows the setup in Claude Code and Codex and gives tips for working safely day to day.

Requirements

You need an organization in Dagster+, the managed cloud edition of Dagster. The server does not work with self-hosted open-source Dagster. Also check which region your organization is in: US organizations use https://mcp.agent.dagster.cloud/mcp, EU organizations use https://mcp.agent.eu.dagster.cloud/mcp.

Step 1: Connect Claude Code

Run claude mcp add --transport http dagster-plus https://mcp.agent.dagster.cloud/mcp in your terminal. Then start a Claude Code session, type /mcp, pick dagster-plus and choose Authenticate. Sign in to Dagster+ in the browser and approve access. From then on, the server works with your permissions.

Step 2: Connect Codex

In Codex, the command is codex mcp add dagster-plus --url https://mcp.agent.dagster.cloud/mcp. Sign-in opens in the browser right away. When you then type /mcp in a Codex session, dagster-plus should show as connected.

Step 3: Other clients and tokens

If your client supports OAuth, just enter the address and follow the sign-in. Otherwise, pass two headers: Authorization with the value Bearer followed by your user token, and Dagster-Cloud-Organization with your organization name. You create the token in Dagster+ under user tokens. In Claude Code, append --header twice to the command.

First questions

Start with questions such as: Which runs failed in the last 24 hours? Read the logs of the last failed run and name the cause. Which assets are running unusually long according to Insights? Create an alert that notifies me on Slack when the revenue table asset fails.

Work safely

The server can launch and terminate runs, trigger asset recomputations and delete alert policies. Accidentally recomputing large assets can noticeably drive up data warehouse costs. Dagster therefore recommends creating a dedicated service user with suitable permissions for the assistant instead of using your own account. Always have your client ask for confirmation before write actions.

claude mcp add --transport http dagster-plus https://mcp.agent.dagster.cloud/mcp

OAuth in the browser with your Dagster+ account; the server inherits the signed-in user's permissions. Alternatively, a user token as a bearer header plus the Dagster-Cloud-Organization header. For different permissions, Dagster recommends a service user.

View, launch and terminate runs; read run logs; view and materialize assets; read deployments only; create, update and delete alert policies and Dagster+ issues; read Insights metrics. Always limited to the permissions of the user in use.

The server runs hosted by Dagster in your organization's region (US or EU). Logs, asset names and metrics that end up in the chat also reach the client's AI provider.

Medium. The server can launch and terminate runs, trigger recomputations and delete alert policies; a mistake can drive up data warehouse costs or silence notifications. Logs may contain sensitive data. Use a service user with narrow permissions and require confirmation for write actions.