Dify
Listed at https://dify.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Sandbox 免费:200 GPT-4 调用 + 5MB vector 存储 + 10 文档 + 5,000 API 调用/月;Professional $59/workplace/月:100MB向量 + 100文档 + 50 工作流 | per million tokens |
| Output | Team $159/workplace/月:500MB向量 + 500文档 + 200 工作流 + 500K 触发事件;Enterprise 合同制(SSO/私有部署/SLA) | per million tokens |
🤖 Supported Models (100)
✨ Pros
- ✓GitHub 150,010 stars — the de-facto open-source LLM app platform (langgenius/dify, Apache-style license)
- ✓Clear 4-tier pricing: Sandbox free → Professional $59 → Team $159 → Enterprise custom
- ✓Visual Workflow Builder + RAG pipeline (all doc formats: PDF/PPTX/Word/code, etc.)
- ✓Agent framework with 50+ built-in tools + Function Calling + ReAct dual engines
- ✓100+ model integrations (OpenAI/Anthropic/Google/all Chinese providers + Ollama custom)
- ✓Full LLMOps: annotations/feedback/latency analytics + plugin marketplace + MCP integration + SSE streaming
⚠️ Cons
- ×Sandbox only 5,000 API calls/month — production needs Professional ($59) or higher
- ×Not an independent LLM provider — users bring their own API keys (except 200 free GPT-4 calls in Sandbox)
- ×Workflow feature limited to Professional and up — Sandbox only gets 2 workflows
- ×Multi-workspace support requires Enterprise; Team plan (single workspace) limits team segregation
- ×Dify Cloud international servers are outside China — domestic direct access may have high latency (self-host recommended)
- ×PartnerStack affiliate transparency is limited — no click-rate or conversion funnel in partner dashboard
🎯 Best For
Teams needing to quickly build AI apps (visual workflow + RAG + Agent all-in-one platform); developers wanting to deploy a complete LLM app stack on China-hosted cloud; ops teams needing unified API key and prompt management across multiple LLM providers
💰 Pricing & Plans
| Plan | Price | What's Included |
|---|---|---|
| Sandbox (Free) | $0 | 200 GPT-4 calls + 5MB vector + 10 docs + 2 workflows + 5,000 API calls/mo (30-day logs) |
| Professional | $59/workplace/mo | 100MB vector + 100 docs + 50 workflows + unlimited team members + SSO |
| Team | $159/workplace/mo | 500MB vector + 500 docs + 200 workflows + 500K trigger events + audit logs |
| Enterprise | Custom contract | SSO/SAML, private deploy, SLA, custom data retention, commercial license, dedicated support |
🔧 API & Developer Experience
- •API Style: REST + Server-Sent Events for streaming chat responses; OpenAI-compatible /v1/chat/completions shim via Dify's API gateway since v0.6+. Each Dify app exposes a unique API key and a stable app_id-based endpoint.
- •Backend-as-a-Service: Dify is not a raw LLM provider — it ships a managed backend for AI apps (workflow runtime, RAG pipeline, agent loop, conversation store) so developers don't build them. The hosted Dify Cloud or self-hosted Docker compose replaces ~3,000 lines of glue code.
- •Visual Builder: Drag-and-drop canvas for Workflow nodes (LLM, knowledge retrieval, code execution, HTTP, conditionals, iterations) and Agent graphs with 50+ built-in tools. Variables, conversation memory, and tool schemas are edited inline; no JSON config files.
- •RAG Pipeline: Built-in ingestion for PDF/PPTX/Word/Markdown/HTML/Excel/code/notion; multiple retrieval modes (vector / full-text / hybrid) with rerank support; per-dataset embedding model selection. No external vector DB required for the Sandbox tier.
- •Model Integrations: 100+ LLMs via unified interface — OpenAI, Anthropic, Google, AWS Bedrock, Azure OpenAI, Cohere, Mistral, Ollama, plus all major Chinese providers (DeepSeek, Qwen, GLM, ERNIE, Hunyuan, Doubao, MiniMax, Moonshot). Switch models without re-architecting.
- •Agent Framework: Single-Agent and Multi-Agent graphs with Function Calling and ReAct dual engines; tool marketplace with custom plugin uploads; MCP server integration since v1.0.0-alpha supports model-context-protocol for external tool discovery.
- •Observability & Ops: Built-in logs dashboard with token/latency/cost per request, annotation + feedback collection, prompt versioning, dataset management for evaluation. Exportable to OpenTelemetry-compatible backends.
🛠️ Visual AI App Backend (BaaS)
Dify's defining product is a Backend-as-a-Service for LLM applications — a managed runtime that combines workflow orchestration, RAG pipeline, agent loop, conversation storage, observability, and a visual builder into a single Docker-deployable product. Where raw LLM providers ship a /chat/completions endpoint and expect the developer to assemble every surrounding component, Dify ships all of them pre-integrated and exposes them through a drag-and-drop canvas backed by YAML-exportable DSL. The trade-off is abstraction: teams needing exotic RAG heuristics or low-level tool orchestration often outgrow Dify's node graph and migrate to custom LangChain / LlamaIndex code. For the 80% case of "production-ready AI app in a week," Dify replaces several weeks of glue code with a single self-hosted container.
🌐 China Access & Latency
Dify has first-class China availability: the parent company LangGenius is China-based, native Chinese UI/UX, full Chinese documentation, and a maintained Alibaba Cloud one-click deploy template. The Dify Cloud international edition (cloud.dify.ai) is hosted overseas — direct access from China averages 300-800ms with occasional timeouts; the China-hosted Cloud (dify.ai / enterprise SaaS) is generally <100ms for domestic users. Self-hosting inside Alibaba Cloud, Tencent Cloud, or Huawei Cloud via the official Terraform / docker-compose templates is the recommended production path for latency-sensitive Chinese deployments; Alibaba Cloud Marketplace also offers a managed Dify image with auto-scaling. China's ICP filing is straightforward because LangGenius is a registered mainland entity, simplifying enterprise procurement.