What are AI workflows, and how do you install one?

An "AI workflow" here doesn't mean one downloadable file. It's a full platform or framework you run yourself, and what "install" means depends entirely on which one you pick. Two are visual, drag-and-drop builders you self-host with Docker. Two are Python frameworks you pip install and write code against. This guide covers both shapes and gets you running one, whichever fits your task.

Pick the shape you actually need

Activepieces is the closest thing here to a Zapier alternative: a visual canvas with 280+ integrations, and every one of those integrations doubles as an MCP tool you can hand to an agent in Claude Desktop, Cursor, or Windsurf. Dify and Langflow are also visual, but built specifically for LLM apps: Dify bundles a full RAG pipeline, model management, and LLMOps in one workspace, while Langflow is the lighter of the two, fully MIT-licensed, with one-click export of a flow as an API or MCP server.

If you'd rather write code than drag nodes, CrewAI and LangGraph are both code-first Python frameworks for multi-agent systems, no visual canvas at all. CrewAI is the more opinionated of the two: you define Crews (role-based agents collaborating) and Flows (event-driven control). LangGraph is lower-level: you build your own agent abstraction out of graph nodes and edges, with durable execution and persistent memory built in.

Before you start

For Activepieces, Dify, or Langflow: Docker with the Compose plugin, and git. For CrewAI or LangGraph: Python 3.10 or newer and pip. You don't need both sets of tooling unless you plan to try more than one.

Run a visual builder

Activepieces is the simplest of the three to get running locally:

git clone https://github.com/activepieces/activepieces
cd activepieces
docker compose up

The first run pulls every image, so it takes a few minutes. Dify and Langflow follow the same clone-and-compose pattern; check their asset pages for the exact commands, since the services each one starts differ.

Checkpoint: once the containers report healthy, open the app in your browser and create the first admin account. If the page won't load, check that Docker is actually running and that nothing else is bound to the same port.

Run a code-first framework

LangGraph installs like any Python package:

pip install -U langgraph

Checkpoint: run python -c "import langgraph". No error means it's ready to import in your own script. CrewAI's asset page has the equivalent install command and a minimal Crew example to run right after.

Use one

On a visual builder, you build a flow on the canvas: trigger, steps, branches, optionally a code step when the built-in pieces aren't enough. In Activepieces specifically, you can also open any piece's MCP settings and paste the generated URL into an agent client instead of building a flow by hand. On a code-first framework, you write Python: define your Crew/Flow in CrewAI, or your graph's nodes and edges in LangGraph, then run the script directly.

Two things worth knowing before you commit to one

Dify's license isn't plain Apache-2.0. It's Apache-2.0 with two added conditions: you can't remove the frontend logo/copyright, and you can't run it as a multi-tenant SaaS for others without a separate commercial license. Normal self-hosted use for your own product is fine. Langflow, CrewAI, and LangGraph are all clean MIT with no such carve-out.

CrewAI and LangGraph look similar from a distance (both code-first, both Python, both for multi-agent systems) but sit at different levels: CrewAI gives you Crew and Flow already defined, LangGraph gives you graph primitives and expects you to build your own abstraction on top. Pick CrewAI to move fast with an opinionated structure, LangGraph when you need to control exactly how state and control flow move between agents.

Browse the full list at AI Workflows for install details specific to each one. These five are what's catalogued today, not a permanent ceiling: this category grows as new self-hostable platforms and frameworks get reviewed and added.