Dify

You build in the Workflow Studio, a canvas where nodes are model calls, tool invocations, conditionals, code steps and retrieval from a knowledge base. A separate Knowledge Pipeline handles getting documents into a searchable state, which is the part of RAG that usually consumes the most engineering time and the least attention. A marketplace supplies models, tools and integrations rather than requiring a connector per vendor.

Once an application works, you publish it as a hosted app, an embeddable widget or an API that your own frontend calls. The same definition runs on Dify Cloud, in a VPC deployment or on self-hosted Docker, and moving between them is a deployment choice rather than a rebuild. Cloud plans are priced per workspace with a monthly allowance of message credits, and the free Sandbox tier exists to try the whole thing without a card.

• •Visual workflow studio - Agentic workflows built as a canvas of model calls, tools, branches and code steps, which makes the control flow of an agent reviewable by people who will not read the orchestration code.
• •Knowledge pipeline for RAG - Document ingestion, chunking and retrieval as a first-class part of the platform, with per-plan limits on documents, storage and request rate published on the pricing page.
• •Model and tool marketplace - Integrations for models, tools and data sources supplied as marketplace entries rather than written per vendor, so swapping a model is configuration.
• •Three deployment targets from one definition - Dify Cloud, VPC and self-hosted Docker run the same application, so a security requirement discovered late is a deployment change rather than a rewrite.
• •Collaborative workspace - Multiple team members on one workspace with shared projects and app limits per plan, aimed at teams rather than individual builders.
• •Published API and embeddable apps - Each application is exposed as an API or an embeddable widget, so the platform can sit behind your own product instead of being the product.

Teams who want an agent platform with a visual builder and are willing to adopt an opinionated stack to get one, especially where non-engineers need to read or modify the workflow. It is a credible alternative to assembling LangGraph or CrewAI yourself when the goal is a shipped internal application rather than a framework you own. The VPC and self-hosted options make it viable in places a pure SaaS platform is not, and the enterprise vertical pages suggest the company genuinely sells into regulated industries rather than only to GitHub.

It is the wrong choice if you want an unencumbered open source licence, because the multi-tenant and branding restrictions are real and should be read before you build a business on it. It is also the wrong choice if you want a library rather than a platform: Dify owns the application shape, and teams who want to express an agent in their own code will find the canvas constraining rather than helpful. Note too that cloud pricing is per workspace rather than per seat, which reads cheap at a glance and stops reading cheap the moment you need a second workspace.