Dify

Dify is a platform for building LLM applications, such as chatbots, agents, document Q&A and multi-step workflows, without writing the orchestration code yourself. Instead of wiring model calls, retrieval and tool use together in a framework, you assemble them in a web console and publish the result.

How it works: you build an app on a visual canvas from nodes such as LLM, Knowledge Retrieval, Agent, If-Else, Iteration, Loop, HTTP Request, Human Input and a Code node that runs Python or JavaScript in a sandbox. Knowledge bases handle RAG: you upload files or sync from Notion or websites, choose chunking and indexing, or connect an external retrieval API instead. Model providers are installed as plugins and configured with your own API keys at workspace level, and tools can come from the plugin marketplace or from MCP servers.

What you get: every published app is a shareable web app, an embeddable chat widget, a REST API called with a per-app key, and optionally an MCP server for clients like Claude Desktop or Cursor. Knowledge bases have their own API too. You can use the hosted Dify Cloud, or run the Community Edition yourself with Docker Compose, which starts the API, workers, web frontend, plugin daemon, PostgreSQL, Redis, Weaviate and a code sandbox.

The limits: your agent logic lives in Dify's app definitions rather than in your own codebase, and publishing replaces the live version; separate staging and production environments are an Enterprise feature. Self-hosting needs at least 2 CPU cores and 4 GiB of RAM for a stack of more than a dozen containers. The license is not plain Apache 2.0: running Dify as a multi-tenant service, or removing the logo and copyright from its frontend, requires a commercial license.