Dify

Open-source LLM app platform — visually orchestrate agents and workflows

License: Apache-2.0 (with conditions)
Deploy difficulty: Medium
GitHub stars: 141.7k
Last commit: 1d ago

Latest 1.14.1 22.3k forks 778 open issues Synced 2026-05-17

What Dify is#

Dify is an open-source LLM application platform with four app types — chatbot, agent, workflow and text generation. The value is packaging “prompt + knowledge + tools + model” into an app that exposes a stable API, while non-engineers tweak prompts and upload docs through a UI.

Role in a support stack#

Dify usually plays the AI brain:

RAG over a knowledge base
Multi-step business orchestration (lookup order → call refund API → escalate ticket)
Exposes an OpenAI-compatible API for Chatwoot, LibreChat or a custom web widget

Key features#

Document-level RAG with parent-child chunking, hybrid vector + keyword retrieval
Workflow nodes — conditions, loops, HTTP, code
50+ built-in tools plus OpenAPI-defined custom tools
MCP (Model Context Protocol) as both client and server

Operational notes#

Official docker-compose starts on 4 vCPU / 8 GB
For Chinese content use bge-m3 or text-embedding-3-large
Wire into Chatwoot via its inbox-level /v1/chat-messages endpoint

Strengths

Visual builder for agents and workflows — non-engineers can tweak prompts
Four app types out of the box — chatbot, agent, workflow, text generation
200+ models including local Ollama / Xinference
OpenAI-compatible API — Chatwoot, Lark, WeCom can call it directly

Trade-offs

License allows self-hosting but restricts SaaS resale
Self-host has many services (Postgres, Redis, Weaviate, Sandbox, Plugin Daemon)
Complex agent debugging is still better in code frameworks

Best for

Consolidating LLM capabilities behind a single API for support frontends
Teams where ops people maintain prompts and knowledge bases themselves