01Context
Dagster: Organize the context.
Ingest, transform, catalog and organize your context into structured knowledge, so agents read information that is current and approved.
Orchestrate every step, from the data your agents read to the work they do.
Trusted in Production by:
01Context
Ingest, transform, catalog and organize your context into structured knowledge, so agents read information that is current and approved.
02Connectivity
Build and deploy MCP servers in minutes. Deploy them with a FastMCP Gateway for observability and governance.
03Outcome
Workflows that run each step, pause for a person and verify the result, with durable execution that resumes instead of repeating.
Context from Dagster, tools through the FastMCP Gateway, and a Prefect run that pauses for a person and verifies the result. Drag the steps and trace an edge.
From startups to enterprises, teams trust Prefect with the workflows their businesses depend on.
Prefect, Dagster and FastMCP are open source Python frameworks for workflow orchestration, data orchestration and MCP servers, and every managed service runs the same engine rather than a fork.
Prefect is the name of the company and of its workflow orchestration product. The company makes three open source products: Prefect, for orchestrating workflows and AI agents; Dagster, for data orchestration; and FastMCP, for connecting agents to tools through the Model Context Protocol (MCP). Each has a managed service built on the same engine as its open source core.
Workflow orchestration coordinates the steps of a process: the order they run in, what happens when one fails, and who has to approve what before the work moves on. Prefect does this for Python data pipelines and AI agent workflows, with scheduling, retries, durable execution and human review built into the engine instead of into your code.
Data orchestration builds and maintains the data a business runs on: pulling it in from source systems, transforming it, testing it and keeping it fresh, with a record of where every table came from. In Dagster, each table, file or ML model is an asset defined in Python, so you can see what depends on what, which version is current and whether its quality checks passed. People and agents downstream all work from that checked version.
An MCP gateway is one governed endpoint between AI clients and your MCP servers. It authenticates every client, authorizes each tool call per agent and role, and logs the result, so agents reach the tools they are allowed to use and nothing else. FastMCP Gateway is the managed MCP gateway and MCP hosting service from the team behind FastMCP.
They answer three different questions about an agent. Dagster handles what it knows: your data and business knowledge, curated into context. FastMCP handles what it can do: which tools and systems each agent may reach, and the record of every call. Prefect handles what its goal is: the workflow that runs the work, checks every result and brings in a person when it should.
No. Each product works on its own and most teams start with one. Data teams usually start with Dagster, teams building MCP servers start with FastMCP, and teams automating workflows or agents start with Prefect. Using them together adds decision lineage: every outcome traces back to the context it used, the tools it called and who approved it.
Prefect workflows are plain Python functions that can branch, loop and create tasks at runtime, where Airflow asks you to declare a DAG up front. Durable execution is built into Prefect's engine, so retries and crash recovery are not code you write. And on Prefect's managed service there is no scheduler for your team to host or get paged about.
Durable execution means a workflow keeps its progress when something fails. When a step raises an error or a run is retried, work that already finished is reused instead of repeated, so the run picks up at the step that failed. In Prefect you get this from decorator arguments and settings rather than custom code: retries on flows and tasks, persisted task results that let a retried run skip completed steps, and pauses that wait for a person before continuing.
No. Prefect works on ordinary Python. Add the flow and task decorators to functions you already have, then set retries, timeouts or result persistence as arguments. Prefect does not replay your code, so there are no determinism rules: your code can call APIs and LLMs directly, branch, loop and create tasks at runtime. Existing libraries and agent frameworks keep working, and a coding agent can make the change for you.
No. Prefect can run your flows on Prefect-managed infrastructure or push them to serverless platforms such as Google Cloud Run, AWS ECS and Modal, with no worker to operate. If your code needs to run next to your data, a worker in your own environment runs it while Prefect handles scheduling and visibility. Managed and push execution aren't part of the open-source Prefect server.
Yes. A flow can pause until a person responds in the Prefect UI, with typed input such as an approval or a form. For waits of hours or days, suspend the run instead: it stops running until it is resumed, by a person or by an automation when an event arrives. Every task run and state change is recorded, so you can see where a run is and why.
Yes. Agent runs are long, call unreliable tools and often need approvals. Run each tool call as a task and Prefect retries the calls that fail, reuses calls that already succeeded once results are persisted, pauses when a person must approve, and records every step the agent took. Pydantic AI agents can be wrapped with PrefectAgent for durable execution.
Yes. A Prefect workflow is ordinary Python with two decorators, so any coding agent that writes Python can produce a working flow, add retries, and deploy it without learning a DSL. Your engineers review and own the result, and the same is true for humans: if you can write a Python function, you can ship a Prefect flow.
Yes. All three are open source under the Apache 2.0 license, and each managed service runs the same engine rather than a fork. FastMCP 1.0 was folded into the official MCP Python SDK.
Only if you want to. Every product has a managed service. Prefect can run your flows on serverless compute, so you can start with no infrastructure, or use the hybrid model: workers inside your environment run the code next to your data while the managed control plane handles scheduling and visibility. Dagster offers the same choice of serverless or hybrid deployment. FastMCP Gateway can host your MCP servers or govern servers that stay in your own cloud. All three open source projects can also be self-hosted end to end.
Prefect acquired Dagster Labs in July 2026. Nothing changes for users of either product: Dagster keeps its name and open source license, and no customer needs to take any action. The full announcement is at prefect.io/prefect-acquires-dagster.