Dify

Type: low-code · Vendor: LangGenius, Inc. · Language: Python, TypeScript · License: Apache-2.0 (modified — multi-tenant SaaS and LOGO removal forbidden) · Status: active · Status in practice: mature · First released: 2023-04-12

Links: homepage docs repo

Provide an open-source LLM-app development platform on which a builder visually composes AI workflows, RAG pipelines, and tool-using Agents, and ships them as hosted apps, embedded APIs, or MCP servers from a single self-hostable Docker stack.

Description. Dify is an open-source LLM-app development platform built by LangGenius. Per the README: 'Dify is an open-source LLM app development platform. Its intuitive interface combines AI workflow, RAG pipeline, agent capabilities, model management, observability features … and more, letting you quickly go from prototype to production.' Ships in three editions: hosted Dify Cloud at cloud.dify.ai, self-hostable Community (Docker Compose / Helm / K8s), enterprise. The Chinese README is parallel and uses 中文 terminology throughout: '您可以基于 LLM 函数调用或 ReAct 定义 Agent' (You can define Agents based on LLM function calling or ReAct). License is a modified Apache-2.0 that forbids multi-tenant resale and LOGO removal but otherwise tracks Apache-2.0.

Agent loop shape. Apps on a visual workflow canvas of typed nodes (User Input → LLM / Knowledge Retrieval / Agent / Tool / IF-ELSE / Iteration / Template / Output) plus standalone modes (Chatbot, Agent, Chatflow, Workflow). Agent loop lives in two places: the Agent app (chat-style, reasoning + tools) and the Agent node inside a workflow. Both expose two pluggable strategies — Function Calling (native) and ReAct (Thought → Action → Observation) — bounded by Maximum Iterations safety cap. Workflows can chain into other workflows; LLM nodes support retry + fallback model chains; any Dify app can be published as an MCP server.

Primary use cases

no-code/low-code visual LLM-app building
RAG with Knowledge Base + hybrid search + rerank + automatic citations
Agent apps with Function Calling or ReAct strategy
publishing apps as MCP servers consumable by Claude Desktop / Cursor

flowchart TD U[User input / chat / API / MCP client] --> APP{App type} APP -->|Chatbot / Chatflow| LLM[LLM node] APP -->|Agent app| AGENT[Agent loop] APP -->|Workflow| WF[Workflow graph] AGENT -->|strategy| FC[Function Calling] AGENT -->|strategy| RE[ReAct: Thought->Action->Observation] FC --> TL[Tool plugins / MCP tools / built-in] RE --> TL TL --> AGENT WF --> LLM WF --> KR[Knowledge Retrieval node] WF --> AGENT WF --> TL KR --> KB[(Knowledge Base<br/>hybrid + semantic + rerank)] KB --> LLM LLM --> SO[Structured Output<br/>JSON Schema editor] LLM -.retry+backoff.-> LLM LLM -.fallback.-> LLM2[Alternative model / default / error route] AGENT --> PUB[Publish as MCP Server] PUB --> CD[Claude Desktop / Cursor]

Key concepts

App (docs) — Publishable unit — Chatbot / Agent / Chatflow / Workflow modes.
Workflow (docs) — Visual canvas of typed nodes.
LLM node (docs) — Invokes a language model with prompt, context, structured-output schema, retries, fallback.
Agent node / Agent Strategy → react (docs) — Autonomous tool-calling loop driven by Function Calling or ReAct.
Knowledge Base (知识库) → agentic-rag (docs) — Vectorised document corpus queried at runtime with hybrid search + rerank.
Tool / Plugin → tool-use (docs) — External capability — built-in, custom HTTP, plugin-distributed, or MCP-server-backed.
Model provider (docs) — Pluggable LLM/embedding/rerank vendor; provider-agnostic.

Dify

Neighbourhood

Instantiates

Anti-patterns avoided

Alternatives & relatives

Listed as alternative by (9)

References

Provenance