Langflow

You start either by installing the Python package or by downloading Langflow Desktop, the free native build the homepage's primary call to action points at. Both give you the same editor. You build a flow by dragging components onto the canvas and connecting them: an input, a model, a prompt, a retriever, a vector store, a tool. Each component exposes its real parameters, so a model node shows the model name, the API key field and the temperature rather than hiding them behind a wizard.

When the visual layer stops being enough you write Python. Components are Python underneath and can be customised or replaced outright, which is the difference between a low-code tool you grow out of and one you keep. Finished flows run as agents, singly or as a fleet with every component available to them as a tool, and any flow can be exposed as an API endpoint so the rest of your stack calls it over HTTP. Langflow also builds MCP servers, so a flow becomes a tool an external agent can call. Deployment is either your own infrastructure or a hosted cloud account, and the project states the same Langflow runs in both.

• •Visual Flow Canvas - Components are dragged, connected and configured on a canvas, with hundreds of pre-built flows and components to start from rather than a blank page.
• •Python All The Way Down - Every component is Python and can be customised or swapped, so the canvas is a convenience rather than a ceiling.
• •Agents and Fleets - Run a single agent or a fleet of them, with the components in your workspace exposed to them as callable tools.
• •MCP Server Generation - Flows can be published as MCP servers, which puts your pipeline inside any MCP-speaking agent or editor.
• •Flow as an API - Any flow can be exposed as an HTTP endpoint for the rest of an application to call, which is what turns a prototype into something a product can depend on.
• •Broad Connector Library - Major model providers, vector stores and data sources are supported out of the box, so swapping a model or a store is a node change rather than a rewrite.
• •MIT Licensed - Self-hosting is unrestricted and the source can be forked, which is the clearest practical difference from the alternatives in this category.

Developers and teams prototyping RAG applications and agents who want to see the graph rather than hold it in their heads, and who want to hand a working flow to a colleague who does not write Python. The named testimonials on the site come from engineering and product roles at BetterUp, WinWeb and Athena Intelligence, which is the kind of adoption this is aimed at. It also suits anyone who needs to self-host without a licence conversation: MIT means you can run it, change it and ship it.

It is a poor fit in a few cases. If you want a visual builder with a published price and a clear commercial support path, Langflow's own site gives you neither, and the IBM relationship is visible in the cookie banner but not explained anywhere that sets out what the commercial offering is. If you already run Dify and are happy with it, the capability overlap is large enough that switching for its own sake is hard to justify. And if your pipeline is simple and stable, a canvas adds a layer between you and the code that you may not want.