ArcLibrary

Dify / Coze (visual agent platforms)

Drag-and-drop platforms that turn RAG / Workflow / Skills into a no-code AI app builder.

DifyCozeNoCode
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#

PlatformDeploymentBest forStrength
DifySelf-host / SaaSTeams + data-sensitive scenariosOpen-source, strong workflow, enterprise-ready
Coze (CN / international)SaaSChinese ecosystem / Douyin / Feishu botsModels + channels integrated
n8n + AI nodesSelf-hostAlready use n8n for automationGeneric automation + AI
Flowise / LangFlowSelf-hostWant LangChain visualisedDirectly 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.
Code framework
**Write code.**
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.

Further reading#