核心 · Key Idea
In one line: Dify (open-source) and Coze (ByteDance, mostly cloud) put RAG, Workflow, Tools, Agent behind a graphical canvas. Business / product / ops people can drag-and-drop a working agent — not every line has to come from an engineer.
What it is#
Inside the web UI you see:
- Left: node library (LLM node, knowledge-base node, HTTP node, conditional node, variable node)
- Centre: a canvas where you connect nodes
- Right: a debug panel — feed inputs, watch each node's output
- Top: publish → become an API / embeddable widget / WeChat bot
[Input] → [Intent classification] → [KB retrieval] → [LLM generation] → [Output]
↘ [HTTP call to Order API] ↗
The whole pipeline is codeless.
Analogy#
打个比方 · Analogy
- LangGraph = assemble Lego with code — flexible, requires coding.
- Dify / Coze = a packaged Lego Friends kit — instructions on the box, all parts compatible. Not everyone wants to operate the factory.
Key concepts#
App / Agent / WorkflowThree app types
App = single-prompt chat; Agent = tool loop; Workflow = freeform graph.
KnowledgeKnowledge base
Upload docs → auto-chunk + embed + vector store. One-stop RAG.
Tools / PluginsTool marketplace
Hundreds of built-ins: search, image gen, code, HTTP, API calls…
Variables / BranchesVariables / branches
Node outputs are referenced; canvas supports conditional branches and loops.
Self-host vs CloudSelf-host vs cloud
Dify is fully open-source and self-hostable; Coze is primarily SaaS.
Main options#
| Platform | Deployment | Best for | Strength |
|---|---|---|---|
| Dify | Self-host / SaaS | Teams + data-sensitive scenarios | Open-source, strong workflow, enterprise-ready |
| Coze (CN / international) | SaaS | Chinese ecosystem / Douyin / Feishu bots | Models + channels integrated |
| n8n + AI nodes | Self-host | Already use n8n for automation | Generic automation + AI |
| Flowise / LangFlow | Self-host | Want LangChain visualised | Directly draws LangChain graphs |
How it works#
Visual platforms are Workflow + Agent + RAG underneath — the same primitives as LangGraph, just packaged as a canvas.
Practical notes#
- Visual is best for business / ops. PMs can edit prompts and add nodes themselves — 10× faster than filing a ticket and waiting for a release.
- Complex backends still hybrid. Run the core in LangGraph; expose Dify as the orchestration layer to business — best of both.
- Dify self-host suits B2B. Customer data stays on their infra; the OSS edition has enough features.
- Coze plays a "whole-stack" game. Models, plugins, distribution channels (Douyin / Feishu / Telegram) all integrated — fastest path for "I want a bot, not technology."
- Export workflow JSON when possible. Version control, cross-env migration, disaster backup.
Easy confusions#
Visual platform
**Drag + configure.**
Business / product can edit.
Business / product can edit.
Code framework
**Write code.**
Engineer-only, but maximally flexible.
Engineer-only, but maximally flexible.
Dify / Coze
Generic platform — multi-model + multi-channel + self-host.
ChatGPT GPTs
Embedded in OpenAI — only inside ChatGPT.
Wide reach, limited extensibility.
Wide reach, limited extensibility.