#Introduction to Dify

#What is Dify?

Dify is an open-source platform for building LLM applications, maintained by LangGenius. In Studio you drag-and-drop agents, agentic workflows, and chatbots grounded on your own data, then publish them as a web app or a REST API (and, if you want, as an MCP server).

Official one-liner: build AI apps on your data, then put them in users' hands. The name Dify comes from Do It For You.

┌─────────────────────────────────────────────────────────────┐
│                      Dify platform                           │
└─────────────────────────────────────────────────────────────┘
   Model providers   Knowledge (RAG)     Studio
   OpenAI /          chunk · embed        Chatbot / Agent
   DeepSeek /        hybrid retrieve      Workflow / Chatflow
   Ollama …          citations
         │                 │                    │
         └─────────────────┴─────────┬──────────┘
                                     ▼
                      Web App  ·  REST API  ·  MCP

#App types (official picture)

When you create an app, current docs recommend Workflow or Chatflow (visual graph, reusable nodes). Three simpler “basic” types still exist:

This tutorial goes Chatbot → knowledge → workflow → agent, the visual twin of the code path in the LangChain tutorial.

#Dify vs LangChain vs Coze

All three combine models, tools, and knowledge. The split is who writes the orchestration, where data lives, and whether you can self-host.

When to pick which:

• Ship a knowledge-grounded assistant in days, and let non-engineers edit prompts → Dify
• Fine-grained state, tests, deep Python integration → LangChain (same RAG / agent loop)
• Stay inside a China-cloud bot ecosystem and channels → evaluate Coze; keep data on-prem → self-host Dify

They compose: Dify orchestrates and publishes; an HTTP or code node can call your LangChain service.

#Good fit / caveats

Good fit: internal FAQ and policy Q&A (knowledge + citations); fixed classify/generate pipelines (Workflow / Chatflow); tool-using assistants (Agent); one app for both browser users and backends (Web App + API).

Watch out: very complex typed state machines belong in LangChain / LangGraph; ultra-low latency pays a studio/sandbox tax; self-host needs CPU/RAM/disk and real volumes backups (see Installation).

#How you run it

Repo: github.com/langgenius/dify. Pair local models with the Ollama tutorial. To expose a Dify app to an IDE, see the MCP tutorial.

A workspace is the unit of sharing: models, knowledge, and apps live together. Only Owner / Admin can add provider keys; everyone else consumes whatever is already configured. Export DSL (YAML) when you copy an app between instances, and keep secrets in environment variables so the file stays safe to share.

#Next steps

• Installation
• Quick Start
• Side-by-side: LangChain introduction