[{"data":1,"prerenderedAt":1692},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"compare-autogen-vs-dify":8,"compare-a-autogen":9,"compare-b-dify":578},{"tools":4,"reviews":5},98,27,{"tools":4,"reviews":5,"playbooks":7,"news":5},22,null,{"id":10,"title":11,"alternatives":12,"api_compatible":8,"body":16,"category":541,"chinese_friendly":530,"cover":542,"description":543,"domestic":544,"extension":545,"faq":8,"free":544,"github":522,"languages":546,"lastVerified":548,"meta":549,"models":8,"navigation":550,"notSuitable":8,"opensource":550,"path":551,"pillar":552,"platforms":553,"priceTable":8,"pricing":556,"published":557,"relatedPlaybooks":8,"relatedReviews":8,"score":558,"self_host":544,"seo":561,"seoTitle":562,"slug":563,"sources":564,"stem":567,"suitable":8,"tagline":568,"tags":569,"updated":548,"verdict":576,"website":514,"__hash__":577},"tools\u002Ftools\u002Fagent\u002Fplatform\u002Fautogen.md","AutoGen",[13,14,15],"agent\u002Fplatform\u002Fcrewai","agent\u002Fplatform\u002Flangflow","agent\u002Fplatform\u002Fdify",{"type":17,"value":18,"toc":525},"minimark",[19,24,33,36,39,98,101,104,108,117,141,146,169,172,209,212,352,355,403,407,439,443,449,467,473,482,485,500,503,508],[20,21,23],"h2",{"id":22},"tldr","TL;DR",[25,26,27,28,32],"p",{},"AutoGen 是微软研究院开源的多 Agent 对话框架（MIT 协议），用 Python 代码定义 Agent 角色、对话流程和工具调用。核心是 ",[29,30,31],"code",{},"ConversableAgent"," + Group Chat 模式——多个 Agent 自动对话协作完成任务，支持代码执行、工具调用、人在回路（human-in-the-loop）。AutoGen 0.4+ 重构为事件驱动架构，性能和扩展性大幅提升。",[25,34,35],{},"适合：AI 研究者、需要精细控制 Agent 协作逻辑的高级开发者、多 Agent 实验项目。不适合：快速原型验证（用 CrewAI）、非技术用户（用 Dify \u002F Flowise）、需要 GUI 的场景、追求 API 稳定性的生产项目。",[20,37,38],{"id":38},"核心能力",[40,41,42,50,56,62,68,74,80,86,92],"ul",{},[43,44,45,49],"li",{},[46,47,48],"strong",{},"多 Agent 对话","：Group Chat 模式，多个 Agent 自动对话协作完成任务",[43,51,52,55],{},[46,53,54],{},"代码执行","：内置 Docker 代码执行器，Agent 可写代码 + 运行 + 调试",[43,57,58,61],{},[46,59,60],{},"工具调用","：自定义函数工具，Agent 自动选择和调用",[43,63,64,67],{},[46,65,66],{},"人在回路","：Human-in-the-loop 模式，关键决策需人工确认",[43,69,70,73],{},[46,71,72],{},"事件驱动架构","：0.4+ 重构为 async 事件驱动，支持分布式 Agent",[43,75,76,79],{},[46,77,78],{},"可定制 Agent","：system message \u002F 工具集 \u002F 终止条件全可自定义",[43,81,82,85],{},[46,83,84],{},"多模型支持","：OpenAI \u002F Claude \u002F Azure \u002F Ollama \u002F Gemini 等",[43,87,88,91],{},[46,89,90],{},"Agent 可组合","：嵌套 Agent、层级 Agent、条件路由",[43,93,94,97],{},[46,95,96],{},"可观测性","：集成 OpenTelemetry \u002F LangSmith 追踪 Agent 行为",[20,99,100],{"id":100},"价格",[25,102,103],{},"完全免费、MIT 开源、商用免费。运行成本仅来自所接入的 LLM API 调用费用。",[20,105,107],{"id":106},"体验与评测资料整理","体验与评测（资料整理）",[109,110,111],"blockquote",{},[25,112,113,114],{},"说明：本节基于官方文档与公开评测整理，非本站独立实测环境，具体数据请以官方为准。\n",[46,115,116],{},"亮点：",[40,118,119,122,125,128,135,138],{},[43,120,121],{},"Group Chat 模式让多 Agent 协作真正\"自动化\"，代码 reviewer + coder + tester 角色分工清晰",[43,123,124],{},"Docker 代码执行器安全隔离，Agent 写的代码在沙箱中运行",[43,126,127],{},"0.4+ 的事件驱动架构性能提升明显，异步并发能力强",[43,129,130,131,134],{},"自定义工具集成灵活，Python 函数加 ",[29,132,133],{},"@user_function"," 装饰器即可",[43,136,137],{},"微软背书，学术认可度高，论文引用多",[43,139,140],{},"人在回路模式适合需要人工把关的高风险场景",[25,142,143],{},[46,144,145],{},"踩坑：",[40,147,148,151,154,157,160,163,166],{},[43,149,150],{},"学习曲线非常陡峭，文档虽全但概念密集，新手容易劝退",[43,152,153],{},"0.2 → 0.4 API 大改，迁移成本高，网上旧教程大量失效",[43,155,156],{},"Agent 对话容易\"跑飞\"——无限循环 \u002F 偏离主题，需仔细设计终止条件",[43,158,159],{},"无 GUI，调试全靠日志和 print，排查多 Agent 对话链路费时",[43,161,162],{},"Token 消耗大——多 Agent 对话轮次多，API 费用叠加明显",[43,164,165],{},"错误处理不够健壮，LLM 返回格式异常时容易崩溃",[43,167,168],{},"社区活跃度不如 LangChain \u002F CrewAI，遇到问题搜索不到答案",[20,170,171],{"id":171},"上手",[173,174,175,181,188,194,200,206],"ol",{},[43,176,177,180],{},[29,178,179],{},"pip install autogen-agentchat autogen-ext","（0.4+ 新包名）",[43,182,183,184,187],{},"配置 LLM：设置 ",[29,185,186],{},"OPENAI_API_KEY"," 环境变量或代码内传入",[43,189,190,191],{},"定义 Agent：",[29,192,193],{},"AssistantAgent(name=\"coder\", system_message=\"...\", model_client=...)",[43,195,196,197],{},"创建 Group Chat：",[29,198,199],{},"RoundRobinGroupChat(agents=[agent1, agent2])",[43,201,202,203],{},"发起任务：",[29,204,205],{},"result = await team.run(task=\"写一个贪吃蛇游戏\")",[43,207,208],{},"进阶：加 Docker 代码执行器 + 自定义工具 + 人在回路",[20,210,211],{"id":211},"对比",[213,214,215,236],"table",{},[216,217,218],"thead",{},[219,220,221,225,227,230,233],"tr",{},[222,223,224],"th",{},"维度",[222,226,11],{},[222,228,229],{},"CrewAI",[222,231,232],{},"LangGraph",[222,234,235],{},"Dify",[237,238,239,255,271,288,303,319,335],"tbody",{},[219,240,241,245,248,250,252],{},[242,243,244],"td",{},"形态",[242,246,247],{},"Python 框架",[242,249,247],{},[242,251,247],{},[242,253,254],{},"可视化平台",[219,256,257,260,263,266,268],{},[242,258,259],{},"上手难度",[242,261,262],{},"高",[242,264,265],{},"中",[242,267,262],{},[242,269,270],{},"低",[219,272,273,276,279,282,285],{},[242,274,275],{},"多 Agent",[242,277,278],{},"✅ Group Chat",[242,280,281],{},"✅ Crew 角色",[242,283,284],{},"✅ 图编排",[242,286,287],{},"✅ 工作流",[219,289,290,292,295,298,300],{},[242,291,54],{},[242,293,294],{},"✅ Docker 沙箱",[242,296,297],{},"需自定义",[242,299,297],{},[242,301,302],{},"沙箱",[219,304,305,308,311,314,316],{},[242,306,307],{},"GUI",[242,309,310],{},"❌",[242,312,313],{},"❌（有 CrewAI Studio）",[242,315,310],{},[242,317,318],{},"✅",[219,320,321,324,327,330,333],{},[242,322,323],{},"API 稳定性",[242,325,326],{},"一般（大改过）",[242,328,329],{},"较好",[242,331,332],{},"好",[242,334,332],{},[219,336,337,340,343,346,349],{},[242,338,339],{},"适合场景",[242,341,342],{},"研究 \u002F 复杂协作",[242,344,345],{},"业务自动化",[242,347,348],{},"精确流程控制",[242,350,351],{},"应用构建",[20,353,354],{"id":354},"避坑",[40,356,357,367,373,379,385,391,397],{},[43,358,359,362,363,366],{},[46,360,361],{},"锁定版本","：0.2 和 0.4 API 不兼容，",[29,364,365],{},"pip install"," 时务必指定版本",[43,368,369,372],{},[46,370,371],{},"设计终止条件","：Group Chat 不设终止条件会无限对话，设 max_turns + 终止关键词",[43,374,375,378],{},[46,376,377],{},"Token 成本控制","：多 Agent 对话 token 消耗是单 Agent 的 3-5 倍，用 GPT-4 级模型注意费用",[43,380,381,384],{},[46,382,383],{},"代码执行器一定要用 Docker","：直接本地执行 Agent 生成的代码有安全风险",[43,386,387,390],{},[46,388,389],{},"别指望第一次跑通","：system message 调试 + 工具定义 + 终止条件需要反复迭代",[43,392,393,396],{},[46,394,395],{},"错误处理要完善","：LLM 返回异常格式时手动 catch + 重试",[43,398,399,402],{},[46,400,401],{},"不要用旧教程","：0.4+ 完全重构，网上大部分 AutoGen 教程是 0.2 版本的",[20,404,406],{"id":405},"适合-不适合","适合 \u002F 不适合",[40,408,409,412,415,418,421,424,427,430,433,436],{},[43,410,411],{},"✅ AI 研究者实验多 Agent 协作模式",[43,413,414],{},"✅ 需要代码级精细控制 Agent 行为",[43,416,417],{},"✅ 需要 Agent 代码执行 + 自动调试",[43,419,420],{},"✅ 人在回路的高风险决策场景",[43,422,423],{},"✅ 学术项目 \u002F 论文复现",[43,425,426],{},"❌ 快速原型验证（用 CrewAI，API 更简洁）",[43,428,429],{},"❌ 非技术用户（用 Dify \u002F Flowise）",[43,431,432],{},"❌ 追求 API 稳定性的生产项目（版本变动大）",[43,434,435],{},"❌ 需要可视化调试（无 GUI，全靠日志）",[43,437,438],{},"❌ 预算敏感场景（多 Agent 对话 token 消耗大）",[20,440,442],{"id":441},"faq","FAQ",[25,444,445,448],{},[46,446,447],{},"Q: AutoGen 和 CrewAI 怎么选？","\nA: AutoGen 更底层、更灵活，适合研究和复杂多 Agent 协作实验，但学习成本高。CrewAI API 更简洁直观，角色 + 任务 + 流程的概念更易理解，适合业务自动化场景。研究选 AutoGen，做产品选 CrewAI。",[25,450,451,454,455,458,459,462,463,466],{},[46,452,453],{},"Q: AutoGen 0.2 和 0.4 有什么区别？","\nA: 0.4 是完全重构版本——从同步改为异步事件驱动架构，包名从 ",[29,456,457],{},"pyautogen"," 改为 ",[29,460,461],{},"autogen-agentchat"," + ",[29,464,465],{},"autogen-ext","，API 全面更新。性能和扩展性大幅提升但旧代码无法直接迁移。新项目直接用 0.4+。",[25,468,469,472],{},[46,470,471],{},"Q: 多 Agent 对话成本高吗？","\nA: 高。多 Agent 每轮对话都消耗 token，一个任务 5-10 轮对话是常态，使用 GPT-4 级模型单个任务可能花费 $0.5-2。建议开发调试用便宜模型（GPT-4o-mini），生产再切高级模型。",[25,474,475,478,479,481],{},[46,476,477],{},"Q: 可以接入本地模型吗？","\nA: 可以。通过 ",[29,480,465],{}," 的 OpenAI 兼容客户端接入 Ollama \u002F vLLM \u002F LM Studio 的本地模型端点。但本地模型能力有限，复杂多 Agent 协作效果可能不如 GPT-4 \u002F Claude。",[20,483,484],{"id":484},"相关阅读",[25,486,487,491,492,491,496],{},[488,489,229],"a",{"href":490},"\u002Fagent\u002Fplatform\u002Fcrewai.html"," · ",[488,493,495],{"href":494},"\u002Fagent\u002Fplatform\u002Fflowise.html","Flowise",[488,497,499],{"href":498},"\u002Fcoding\u002Fapi\u002Flangfuse.html","Langfuse",[20,501,502],{"id":502},"来源",[109,504,505],{},[25,506,507],{},"本文的价格、版本号与性能数据均参考以下官方渠道整理，可能随时间变动，请以官方实时信息为准。",[40,509,510,518],{},[43,511,512],{},[488,513,517],{"href":514,"rel":515},"https:\u002F\u002Fmicrosoft.github.io\u002Fautogen",[516],"nofollow","官网",[43,519,520],{},[488,521,524],{"href":522,"rel":523},"https:\u002F\u002Fgithub.com\u002Fmicrosoft\u002Fautogen",[516],"GitHub",{"title":526,"searchDepth":527,"depth":527,"links":528},"",3,[529,531,532,533,534,535,536,537,538,539,540],{"id":22,"depth":530,"text":23},2,{"id":38,"depth":530,"text":38},{"id":100,"depth":530,"text":100},{"id":106,"depth":530,"text":107},{"id":171,"depth":530,"text":171},{"id":211,"depth":530,"text":211},{"id":354,"depth":530,"text":354},{"id":405,"depth":530,"text":406},{"id":441,"depth":530,"text":442},{"id":484,"depth":530,"text":484},{"id":502,"depth":530,"text":502},"platform","\u002Fimg\u002Ftools\u002Fautogen.webp","AutoGen 真实评测：微软开源的多 Agent 对话框架（MIT 协议），通过代码定义 Agent 角色和协作流程，支持多 Agent 对话、工具调用、代码执行。适合需要精细控制多 Agent 协作逻辑的开发者和研究团队。",false,"md",[547],"en","2026-07-30",{},true,"\u002Ftools\u002Fagent\u002Fplatform\u002Fautogen","agent",[554,555],"linux","docker","Free \u002F 开源（MIT）","2026-07-05",{"power":559,"ux":530,"price":560,"cn_support":530,"stability":527},4,5,{"title":11,"description":543},"AutoGen - 微软多 Agent 框架评测与使用 | AIHO","agent\u002Fplatform\u002Fautogen",[565,566],{"title":517,"url":514},{"title":524,"url":522},"tools\u002Fagent\u002Fplatform\u002Fautogen","微软开源多 Agent 对话框架，代码驱动 Agent 协作",[570,571,572,573,574,575],"agent-platform","multi-agent","framework","microsoft","python","opensource","需要代码级精细控制多 Agent 协作逻辑的研究者和高级开发者首选，微软背书 + 代码执行 + Group Chat 模式强大，但学习曲线陡峭、API 稳定性一般、无 GUI，不适合快速原型或非技术用户。","Asmg5F2Nn2KaD4XJWzqRwFA4s1iK_qUmi34M41yhcNE",{"id":579,"title":235,"alternatives":580,"api_compatible":8,"body":584,"category":541,"chinese_friendly":559,"cover":1632,"description":1633,"domestic":544,"extension":545,"faq":8,"free":544,"github":878,"languages":1634,"lastVerified":8,"meta":1637,"models":8,"navigation":550,"notSuitable":8,"opensource":550,"path":1638,"pillar":552,"platforms":1639,"priceTable":1642,"pricing":1659,"published":1660,"relatedPlaybooks":8,"relatedReviews":1661,"score":1666,"self_host":550,"seo":1667,"seoTitle":1668,"slug":15,"sources":1669,"stem":1681,"suitable":8,"tagline":1682,"tags":1683,"updated":1689,"verdict":1690,"website":1574,"__hash__":1691},"tools\u002Ftools\u002Fagent\u002Fplatform\u002Fdify.md",[581,582,583,14],"agent\u002Fplatform\u002Fcoze","agent\u002Fplatform\u002Ffastgpt","agent\u002Fplatform\u002Fn8n",{"type":17,"value":585,"toc":1610},[586,588,613,618,621,626,629,697,700,704,707,721,739,743,746,757,761,769,780,784,787,790,794,803,863,870,874,882,945,948,968,972,975,1035,1041,1045,1139,1142,1171,1174,1311,1329,1368,1371,1444,1446,1449,1469,1472,1501,1503,1565,1567,1598,1606],[20,587,23],{"id":22},[589,590,595,601],"div",{"className":591},[592,593,594],"card","p-5","my-4",[25,596,597,600],{},[46,598,599],{},"一句话："," Dify 是开源 LLMOps 平台的事实标准。GitHub 13 万 star、累计 100 万+ 生产 app（据 chatforest.com 2026 评测引用 Dify 官方数据），把\"可视化工作流编排 + RAG 知识库 + Agent + MCP 协议\"打包成一个 Docker Compose 能跑起来的东西。",[25,602,603,604,607,608,612],{},"最大价值是 ",[46,605,606],{},"完全开源 + 模型不挑食","——同一个工作流里同时调 OpenAI、Anthropic、Ollama 本地、DeepSeek、Qwen 都行。代价是部署比 ",[488,609,611],{"href":610},"\u002Fagent\u002Fplatform\u002Fcoze.html","Coze"," 折腾，新手得读 1-2 小时文档。",[109,614,615],{},[25,616,617],{},"来源说明：本文基于 docs.dify.ai 官方文档、langgenius\u002Fdify GitHub 仓库、第三方评测（besthub.dev \u002F chatforest.com \u002F joshuaopolko.com \u002F zhihu 知名专栏）综合归纳。版本号会变，部署要求请以官方最新文档为准。",[20,619,620],{"id":620},"核心特性",[622,623,625],"h3",{"id":624},"可视化工作流chatflow-workflow","可视化工作流（Chatflow + Workflow）",[25,627,628],{},"Dify 把 LLM 应用拆成两种\"应用类型\"：",[213,630,631,643],{},[216,632,633],{},[219,634,635,638,640],{},[222,636,637],{},"类型",[222,639,339],{},[222,641,642],{},"编排范式",[237,644,645,658,671,684],{},[219,646,647,652,655],{},[242,648,649],{},[46,650,651],{},"Chatbot",[242,653,654],{},"简单对话机器人",[242,656,657],{},"prompt + tools",[219,659,660,665,668],{},[242,661,662],{},[46,663,664],{},"Agent",[242,666,667],{},"自主多步任务",[242,669,670],{},"ReAct \u002F Function Calling",[219,672,673,678,681],{},[242,674,675],{},[46,676,677],{},"Chatflow",[242,679,680],{},"对话型工作流（多轮 + 分支）",[242,682,683],{},"节点 DAG，带聊天上下文",[219,685,686,691,694],{},[242,687,688],{},[46,689,690],{},"Workflow",[242,692,693],{},"单次输入→输出（API 模式）",[242,695,696],{},"节点 DAG，无对话状态",[25,698,699],{},"节点类型覆盖：LLM、知识检索、HTTP 请求、代码执行（Python \u002F JS）、条件分支、迭代、变量聚合、参数提取、问题分类——满足\"用拖拽实现可观测的 LLM pipeline\"。",[622,701,703],{"id":702},"rag-知识库","RAG 知识库",[25,705,706],{},"内置完整 RAG 链路：",[173,708,709,712,715,718],{},[43,710,711],{},"上传文档（PDF \u002F Word \u002F Markdown \u002F 网页）",[43,713,714],{},"自动分块 + embedding（可配置分段策略和 embedding 模型）",[43,716,717],{},"混合检索（向量 + 全文 + 重排）",[43,719,720],{},"引用溯源（回答末尾自动附原文片段）",[25,722,723,724,729,730,733,734,738],{},"注意：根据 ",[488,725,728],{"href":726,"rel":727},"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F1887141987838309480",[516],"知乎 LLM 实战笔记 2025-03 对比"," 的实测，Dify ",[46,731,732],{},"社区版默认是基础语义检索","，企业版才解锁多路召回 + 重排。RAG 极致精度场景仍推荐 ",[488,735,737],{"href":736},"\u002Fagent\u002Fplatform\u002Ffastgpt.html","FastGPT","（实测准确率高 10+ 个百分点），Dify 胜在工作流而非纯 RAG。",[622,740,742],{"id":741},"模型生态40-提供商","模型生态：40+ 提供商",[25,744,745],{},"Dify 通过插件市场接入主流模型——OpenAI、Anthropic、Google Gemini、Azure、AWS Bedrock、Cohere、xAI、DeepSeek、Qwen、智谱、文心、豆包、月之暗面、Ollama、LM Studio、Replicate、Together AI、OpenRouter……几乎你能数出来的 LLM 提供商都在。",[25,747,748,749,751,752,756],{},"国产模型原生支持（不像 ",[488,750,737],{"href":736}," 需要 ",[488,753,755],{"href":754},"\u002Fcoding\u002Fapi\u002Fone-api.html","OneAPI"," 中转），是 Dify 在国内 toB 场景流行的关键。",[622,758,760],{"id":759},"mcp-协议支持","MCP 协议支持",[25,762,763,764,768],{},"Dify 较早接入了 ",[488,765,767],{"href":766},"\u002Fwiki\u002Fmcp.html","MCP（Model Context Protocol）","，工作流可以直接调 MCP Server 暴露的 tools。意味着你可以让 Dify 工作流：",[40,770,771,774,777],{},[43,772,773],{},"通过 MCP 调本地 PostgreSQL \u002F SQLite",[43,775,776],{},"通过 MCP 调 GitHub \u002F Slack \u002F Linear",[43,778,779],{},"通过 MCP 调自家内部系统（写一个 MCP Server 即可）",[622,781,783],{"id":782},"api-first","API-first",[25,785,786],{},"每个 app 自动暴露 REST API，参数和返回结构自动生成 OpenAPI Schema。集成到自家产品里不需要写包装代码，给前端 \u002F 微信小程序 \u002F 飞书机器人调用都方便。",[20,788,789],{"id":789},"价格与运行成本",[622,791,793],{"id":792},"云版difyai","云版（dify.ai）",[25,795,796,797,802],{},"根据 ",[488,798,801],{"href":799,"rel":800},"https:\u002F\u002Fwww.tooljunction.io\u002Fai-tools\u002Fdify-ai",[516],"tooljunction.io 2026 评测"," 引用的官方定价：",[213,804,805,817],{},[216,806,807],{},[219,808,809,812,814],{},[222,810,811],{},"套餐",[222,813,100],{},[222,815,816],{},"主要限制",[237,818,819,830,841,852],{},[219,820,821,824,827],{},[242,822,823],{},"Sandbox",[242,825,826],{},"免费",[242,828,829],{},"200 次模型调用，1 app，5MB 知识库",[219,831,832,835,838],{},[242,833,834],{},"Professional",[242,836,837],{},"$59\u002F月起",[242,839,840],{},"5000 调用\u002F月，多 app，50MB 知识库",[219,842,843,846,849],{},[242,844,845],{},"Team",[242,847,848],{},"$159\u002F月起",[242,850,851],{},"团队协作、SSO",[219,853,854,857,860],{},[242,855,856],{},"Enterprise",[242,858,859],{},"联系销售",[242,861,862],{},"定制 SLA、私有云",[25,864,865,866,869],{},"注意：云版价格只是 Dify 平台费，",[46,867,868],{},"模型 API 费用另算","（自带 OpenAI \u002F Anthropic key）。",[622,871,873],{"id":872},"自托管推荐","自托管（推荐）",[25,875,876,881],{},[488,877,880],{"href":878,"rel":879},"https:\u002F\u002Fgithub.com\u002Flanggenius\u002Fdify",[516],"官方 GitHub 仓库"," 提供 Docker Compose 部署，社区版完全免费可商用：",[883,884,888],"pre",{"className":885,"code":886,"language":887,"meta":526,"style":526},"language-bash shiki shiki-themes github-light github-dark","git clone https:\u002F\u002Fgithub.com\u002Flanggenius\u002Fdify.git\ncd dify\u002Fdocker\ncp .env.example .env\ndocker compose up -d\n# 默认 http:\u002F\u002Flocalhost \u002F 端口可在 .env 调整\n","bash",[29,889,890,906,915,926,939],{"__ignoreMap":526},[891,892,895,899,903],"span",{"class":893,"line":894},"line",1,[891,896,898],{"class":897},"sScJk","git",[891,900,902],{"class":901},"sZZnC"," clone",[891,904,905],{"class":901}," https:\u002F\u002Fgithub.com\u002Flanggenius\u002Fdify.git\n",[891,907,908,912],{"class":893,"line":530},[891,909,911],{"class":910},"sj4cs","cd",[891,913,914],{"class":901}," dify\u002Fdocker\n",[891,916,917,920,923],{"class":893,"line":527},[891,918,919],{"class":897},"cp",[891,921,922],{"class":901}," .env.example",[891,924,925],{"class":901}," .env\n",[891,927,928,930,933,936],{"class":893,"line":559},[891,929,555],{"class":897},[891,931,932],{"class":901}," compose",[891,934,935],{"class":901}," up",[891,937,938],{"class":910}," -d\n",[891,940,941],{"class":893,"line":560},[891,942,944],{"class":943},"sJ8bj","# 默认 http:\u002F\u002Flocalhost \u002F 端口可在 .env 调整\n",[25,946,947],{},"硬件门槛（社区共识，非官方硬性要求）：",[40,949,950,956,962],{},[43,951,952,955],{},[46,953,954],{},"最低","：2 核 4G，纯外接 API 模式",[43,957,958,961],{},[46,959,960],{},"推荐","：4 核 8G + 至少 30GB 磁盘（向量数据 + 文件存储）",[43,963,964,967],{},[46,965,966],{},"企业","：8 核 16G+，单机日活上千",[622,969,971],{"id":970},"真实-tco","真实 TCO",[25,973,974],{},"按一家中小团队 3 年场景估算（基于上面引用的多份评测交叉对比）：",[213,976,977,990],{},[216,978,979],{},[219,980,981,984,987],{},[222,982,983],{},"成本项",[222,985,986],{},"云版 Professional",[222,988,989],{},"自托管",[237,991,992,1003,1013,1024],{},[219,993,994,997,1000],{},[242,995,996],{},"平台费",[242,998,999],{},"~$2,100（3 年）",[242,1001,1002],{},"$0",[219,1004,1005,1008,1010],{},[242,1006,1007],{},"服务器",[242,1009,1002],{},[242,1011,1012],{},"~$50\u002F月 × 36 = $1,800",[219,1014,1015,1018,1021],{},[242,1016,1017],{},"模型 API",[242,1019,1020],{},"与下同",[242,1022,1023],{},"与上同",[219,1025,1026,1029,1032],{},[242,1027,1028],{},"运维人力",[242,1030,1031],{},"0",[242,1033,1034],{},"约 0.2 人月",[25,1036,1037,1040],{},[46,1038,1039],{},"结论","：日活 \u003C 100 用云版省心；> 500 或数据敏感场景自托管 ROI 更好。",[20,1042,1044],{"id":1043},"上手-10-分钟","上手 10 分钟",[883,1046,1048],{"className":885,"code":1047,"language":887,"meta":526,"style":526},"# 1. 自托管（社区版）\ngit clone https:\u002F\u002Fgithub.com\u002Flanggenius\u002Fdify.git\ncd dify\u002Fdocker\ncp .env.example .env\ndocker compose up -d\n\n# 2. 浏览器打开 http:\u002F\u002Flocalhost\n#    首次会让你创建 admin 账号\n\n# 3. 进入\"设置 → 模型供应商\"，配置 OpenAI \u002F 国产模型 API key\n\n# 4. 在主界面\"创建空白应用\"，选 Chatflow 或 Workflow\n# 5. 拖入\"开始 → LLM → 结束\"节点试一下基础 prompt\n# 6. 满意了点右上\"发布\"，自动生成 API endpoint\n",[29,1049,1050,1055,1063,1069,1077,1087,1093,1099,1105,1110,1116,1121,1127,1133],{"__ignoreMap":526},[891,1051,1052],{"class":893,"line":894},[891,1053,1054],{"class":943},"# 1. 自托管（社区版）\n",[891,1056,1057,1059,1061],{"class":893,"line":530},[891,1058,898],{"class":897},[891,1060,902],{"class":901},[891,1062,905],{"class":901},[891,1064,1065,1067],{"class":893,"line":527},[891,1066,911],{"class":910},[891,1068,914],{"class":901},[891,1070,1071,1073,1075],{"class":893,"line":559},[891,1072,919],{"class":897},[891,1074,922],{"class":901},[891,1076,925],{"class":901},[891,1078,1079,1081,1083,1085],{"class":893,"line":560},[891,1080,555],{"class":897},[891,1082,932],{"class":901},[891,1084,935],{"class":901},[891,1086,938],{"class":910},[891,1088,1090],{"class":893,"line":1089},6,[891,1091,1092],{"emptyLinePlaceholder":550},"\n",[891,1094,1096],{"class":893,"line":1095},7,[891,1097,1098],{"class":943},"# 2. 浏览器打开 http:\u002F\u002Flocalhost\n",[891,1100,1102],{"class":893,"line":1101},8,[891,1103,1104],{"class":943},"#    首次会让你创建 admin 账号\n",[891,1106,1108],{"class":893,"line":1107},9,[891,1109,1092],{"emptyLinePlaceholder":550},[891,1111,1113],{"class":893,"line":1112},10,[891,1114,1115],{"class":943},"# 3. 进入\"设置 → 模型供应商\"，配置 OpenAI \u002F 国产模型 API key\n",[891,1117,1119],{"class":893,"line":1118},11,[891,1120,1092],{"emptyLinePlaceholder":550},[891,1122,1124],{"class":893,"line":1123},12,[891,1125,1126],{"class":943},"# 4. 在主界面\"创建空白应用\"，选 Chatflow 或 Workflow\n",[891,1128,1130],{"class":893,"line":1129},13,[891,1131,1132],{"class":943},"# 5. 拖入\"开始 → LLM → 结束\"节点试一下基础 prompt\n",[891,1134,1136],{"class":893,"line":1135},14,[891,1137,1138],{"class":943},"# 6. 满意了点右上\"发布\"，自动生成 API endpoint\n",[20,1140,1141],{"id":1141},"国内使用注意事项",[173,1143,1144,1150,1156,1162],{},[43,1145,1146,1149],{},[46,1147,1148],{},"云版 dify.ai 直连国内访问稳定但需要付款","——支持国际信用卡 \u002F Stripe",[43,1151,1152,1155],{},[46,1153,1154],{},"自托管 + 国产模型"," = 完全国内闭环，是 Dify 在国内最大优势",[43,1157,1158,1161],{},[46,1159,1160],{},"Docker 镜像拉取","：国内可能慢，建议配 Docker registry 镜像（阿里云 \u002F 网易）",[43,1163,1164,1167,1168,1170],{},[46,1165,1166],{},"数据合规","：完全自托管时，数据零外泄；某些金融 \u002F 政府客户因此从 ",[488,1169,611],{"href":610}," 迁到 Dify",[20,1172,1173],{"id":1173},"与同类怎么选",[213,1175,1176,1198],{},[216,1177,1178],{},[219,1179,1180,1182,1184,1188,1192],{},[222,1181,224],{},[222,1183,235],{},[222,1185,1186],{},[488,1187,611],{"href":610},[222,1189,1190],{},[488,1191,737],{"href":736},[222,1193,1194],{},[488,1195,1197],{"href":1196},"\u002Fagent\u002Fplatform\u002Fn8n.html","n8n",[237,1199,1200,1214,1227,1242,1256,1270,1284,1297],{},[219,1201,1202,1205,1207,1209,1211],{},[242,1203,1204],{},"开源",[242,1206,318],{},[242,1208,310],{},[242,1210,318],{},[242,1212,1213],{},"✅（fair-code）",[219,1215,1216,1219,1221,1223,1225],{},[242,1217,1218],{},"私有部署",[242,1220,318],{},[242,1222,310],{},[242,1224,318],{},[242,1226,318],{},[219,1228,1229,1231,1234,1237,1239],{},[242,1230,259],{},[242,1232,1233],{},"★★★☆☆",[242,1235,1236],{},"★★☆☆☆ 最简单",[242,1238,1233],{},[242,1240,1241],{},"★★★★☆",[219,1243,1244,1247,1250,1252,1254],{},[242,1245,1246],{},"工作流编排",[242,1248,1249],{},"★★★★★",[242,1251,1241],{},[242,1253,1233],{},[242,1255,1249],{},[219,1257,1258,1261,1263,1265,1267],{},[242,1259,1260],{},"RAG 精度",[242,1262,1241],{},[242,1264,1233],{},[242,1266,1249],{},[242,1268,1269],{},"★★☆☆☆",[219,1271,1272,1275,1277,1279,1282],{},[242,1273,1274],{},"模型生态",[242,1276,1249],{},[242,1278,1241],{},[242,1280,1281],{},"★★★☆☆（OneAPI 中转）",[242,1283,1241],{},[219,1285,1286,1289,1291,1293,1295],{},[242,1287,1288],{},"中文场景",[242,1290,1241],{},[242,1292,1249],{},[242,1294,1241],{},[242,1296,1233],{},[219,1298,1299,1302,1304,1307,1309],{},[242,1300,1301],{},"字节生态绑定",[242,1303,310],{},[242,1305,1306],{},"✅（飞书\u002F抖音深度集成）",[242,1308,310],{},[242,1310,310],{},[25,1312,1313,1316,1317,1322,1323,1328],{},[46,1314,1315],{},"怎么选","（基于 ",[488,1318,1321],{"href":1319,"rel":1320},"https:\u002F\u002Fwww.besthub.dev\u002Farticles\u002Fcoze-vs-dify-vs-fastgpt-which-ai-agent-platform-fits-your-needs-fa59cf97b798",[516],"BestHub 2025-07"," 和 ",[488,1324,1327],{"href":1325,"rel":1326},"https:\u002F\u002Fwww.cnblogs.com\u002Fuulucias\u002Fp\u002F19449008",[516],"博客园 2026-01"," 两份选型指南综合）：",[40,1330,1331,1337,1345,1352,1359],{},[43,1332,1333,1336],{},[46,1334,1335],{},"数据必须不出内网 + 工作流复杂"," → Dify",[43,1338,1339,1342,1343],{},[46,1340,1341],{},"个人 \u002F 小团队 \u002F 快速原型 + 字节生态"," → ",[488,1344,611],{"href":610},[43,1346,1347,1342,1350],{},[46,1348,1349],{},"核心场景就是企业知识库 QA",[488,1351,737],{"href":736},[43,1353,1354,1342,1357],{},[46,1355,1356],{},"重点是连接外部 SaaS（Slack \u002F Notion \u002F 数据库）",[488,1358,1197],{"href":1196},[43,1360,1361,1342,1364],{},[46,1362,1363],{},"要画图式表达 LangChain pipeline",[488,1365,1367],{"href":1366},"\u002Fagent\u002Fplatform\u002Flangflow.html","Langflow",[20,1369,1370],{"id":1370},"避坑清单",[40,1372,1373,1379,1396,1407,1420,1426,1432,1438],{},[43,1374,1375,1378],{},[46,1376,1377],{},"社区版与企业版差距比想象大","：多路召回 \u002F 重排序 \u002F 单点登录 \u002F 审计日志都在企业版。社区版做生产前心里要有数。",[43,1380,1381,1387,1388,1391,1392,1395],{},[46,1382,1383,1386],{},[29,1384,1385],{},".env"," 文件改完忘 restart","：",[29,1389,1390],{},"docker compose down && up -d","，不是 ",[29,1393,1394],{},"restart","——后者不重新加载 env。",[43,1397,1398,1387,1401,1406],{},[46,1399,1400],{},"大版本升级会破坏数据库 schema",[488,1402,1405],{"href":1403,"rel":1404},"https:\u002F\u002Fdocs.dify.ai\u002Fzh-hans",[516],"官方升级文档"," 有详细 migration 步骤，跨大版本（如 0.x → 1.x）务必先备份 PostgreSQL 卷。生产环境强烈建议跑 staging 完整验证后再升。",[43,1408,1409,1412,1413,1415,1416,1419],{},[46,1410,1411],{},"RAG 文件大小社区版默认 15MB","：根据上述知乎实测，超过会失败。改 ",[29,1414,1385],{}," 的 ",[29,1417,1418],{},"UPLOAD_FILE_SIZE_LIMIT"," 并重启容器。",[43,1421,1422,1425],{},[46,1423,1424],{},"代码节点的 Sandbox 性能差","：内置代码执行节点跑在隔离容器里启动慢、内存小。生产高频用建议改成 HTTP 节点调外部服务。",[43,1427,1428,1431],{},[46,1429,1430],{},"工作流\"迭代节点\"循环上限","：默认 10 次，复杂 ReAct agent 容易撞天花板，需要在节点设置里调高。",[43,1433,1434,1437],{},[46,1435,1436],{},"Dify Plugin 系统是新东西","：1.0 后引入的 Plugin 体系替代了原来的 Tools\u002FModels 配置方式，老教程可能已过时——以最新官方文档为准。",[43,1439,1440,1443],{},[46,1441,1442],{},"国内 Docker 拉取镜像慢","：先配国内 registry，否则首次 pull 可能要 30+ 分钟。",[20,1445,406],{"id":405},[25,1447,1448],{},"✅ 适合：",[40,1450,1451,1454,1457,1460,1463,1466],{},[43,1452,1453],{},"中大型企业 LLM 中台建设",[43,1455,1456],{},"需要私有化部署（金融 \u002F 医疗 \u002F 政府）",[43,1458,1459],{},"想做\"AI 工作流即产品\"的开发团队",[43,1461,1462],{},"同时需要 RAG + Agent + Workflow 三件套",[43,1464,1465],{},"想用国产模型 + 国际模型混合编排",[43,1467,1468],{},"已经接受 Docker + 一定运维投入",[25,1470,1471],{},"❌ 不适合：",[40,1473,1474,1480,1486,1489,1495],{},[43,1475,1476,1477,1479],{},"纯个人玩家做对话机器人（",[488,1478,611],{"href":610}," 更快）",[43,1481,1482,1483,1485],{},"只想做企业知识库 QA（",[488,1484,737],{"href":736}," RAG 更专）",[43,1487,1488],{},"团队完全没运维能力（云版还行，自托管会踩坑）",[43,1490,1491,1492,1494],{},"需要深度对接字节飞书 \u002F 抖音（",[488,1493,611],{"href":610}," 原生）",[43,1496,1497,1498,1500],{},"工作流核心是连接 100+ SaaS（",[488,1499,1197],{"href":1196}," 节点更全）",[20,1502,484],{"id":484},[40,1504,1505,1517,1535,1554],{},[43,1506,1507,1508,1510,1511,1510,1513,1510,1515],{},"同类对比：",[488,1509,611],{"href":610}," \u002F ",[488,1512,737],{"href":736},[488,1514,1197],{"href":1196},[488,1516,1367],{"href":1366},[43,1518,1519,1520,1510,1524,1510,1528,1510,1531],{},"概念基础：",[488,1521,1523],{"href":1522},"\u002Fwiki\u002Fai-agent.html","AI Agent",[488,1525,1527],{"href":1526},"\u002Fwiki\u002Frag.html","RAG",[488,1529,1530],{"href":766},"MCP",[488,1532,1534],{"href":1533},"\u002Fwiki\u002Ffunction-calling.html","Function Calling",[43,1536,1537,1538,1510,1542,1510,1546,1510,1550],{},"模型选型：",[488,1539,1541],{"href":1540},"\u002Fmodels\u002Fgpt-5.html","GPT-5",[488,1543,1545],{"href":1544},"\u002Fmodels\u002Fclaude-sonnet-4.html","Claude Sonnet 4",[488,1547,1549],{"href":1548},"\u002Fmodels\u002Fdeepseek-v3.html","DeepSeek-V3",[488,1551,1553],{"href":1552},"\u002Fmodels\u002Fglm-5.2.html","GLM-5.2",[43,1555,1556,1557,1510,1561],{},"进阶：",[488,1558,1560],{"href":1559},"\u002Fwiki\u002Ffine-tuning-vs-rag.html","Fine-tuning vs RAG",[488,1562,1564],{"href":1563},"\u002Fwiki\u002Fcontext-engineering.html","Context Engineering",[20,1566,502],{"id":502},[40,1568,1569,1576,1582,1588,1595],{},[43,1570,1571,1572],{},"官网：",[488,1573,1574],{"href":1574,"rel":1575},"https:\u002F\u002Fdify.ai",[516],[43,1577,1578,1579],{},"中文文档：",[488,1580,1403],{"href":1403,"rel":1581},[516],[43,1583,1584,1585],{},"GitHub：",[488,1586,878],{"href":878,"rel":1587},[516],[43,1589,1590,1591],{},"官方定价：",[488,1592,1593],{"href":1593,"rel":1594},"https:\u002F\u002Fdify.ai\u002Fpricing",[516],[43,1596,1597],{},"第三方评测：tooljunction.io \u002F chatforest.com \u002F besthub.dev \u002F joshuaopolko.com \u002F 知乎 LLM 实战笔记",[25,1599,1600,1601,1605],{},"本卡片由 AIHO 编辑部根据官方公开资料与第三方评测整理。所有事实点均标注来源；如发现版本号 \u002F 价格 \u002F 功能与最新官方信息不一致，请通过 ",[488,1602,1604],{"href":1603},"mailto:hello@aiho.net","反馈邮箱"," 反馈。",[1607,1608,1609],"style",{},"html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":526,"searchDepth":527,"depth":527,"links":1611},[1612,1613,1620,1625,1626,1627,1628,1629,1630,1631],{"id":22,"depth":530,"text":23},{"id":620,"depth":530,"text":620,"children":1614},[1615,1616,1617,1618,1619],{"id":624,"depth":527,"text":625},{"id":702,"depth":527,"text":703},{"id":741,"depth":527,"text":742},{"id":759,"depth":527,"text":760},{"id":782,"depth":527,"text":783},{"id":789,"depth":530,"text":789,"children":1621},[1622,1623,1624],{"id":792,"depth":527,"text":793},{"id":872,"depth":527,"text":873},{"id":970,"depth":527,"text":971},{"id":1043,"depth":530,"text":1044},{"id":1141,"depth":530,"text":1141},{"id":1173,"depth":530,"text":1173},{"id":1370,"depth":530,"text":1370},{"id":405,"depth":530,"text":406},{"id":484,"depth":530,"text":484},{"id":502,"depth":530,"text":502},"\u002Fimg\u002Ftools\u002Fdify.webp","Dify 2026 真实评测：开源 LLMOps 与 AI Agent 平台，集工作流编排、RAG 知识库、Agent、MCP 和多模型接入于一体。本文对比 Coze、FastGPT、n8n，整理自托管部署、云版价格、适合团队和避坑建议。",[1635,547,1636],"zh","ja",{},"\u002Ftools\u002Fagent\u002Fplatform\u002Fdify",[1640,1641,554,555],"windows","macos",[1643,1647,1651,1655],{"plan":1644,"price":1002,"features":1645,"notes":1646},"Self-hosted（开源版）","Docker 一键部署 + 全部核心功能（工作流 \u002F RAG \u002F Agent \u002F MCP）+ 接任意模型 API","私有部署 \u002F 完全免费 \u002F Apache 2.0",{"plan":1648,"price":1002,"features":1649,"notes":1650},"Cloud Sandbox（免费云）","官方托管试水档，含基础调用配额","免运维 \u002F 试水 POC",{"plan":1652,"price":837,"features":1653,"notes":1654},"Cloud Professional","更高调用额度 + 团队协作 + 商用支持","商用云首选",{"plan":1656,"price":1657,"features":1658,"notes":859},"Cloud Team \u002F Enterprise","Custom","更大配额 + SLA + 私有部署支持 + 合规","云版 SaaS（免费档 \u002F Professional $59\u002F月起） + 开源自托管完全免费","2026-06-18",[1662,1663,1664,1665],"coze-deep-review","coze-vs-dify","dify-deep-review","fastgpt-deep-review",{"power":560,"ux":559,"price":560,"cn_support":559,"stability":559},{"title":235,"description":1633},"Dify 评测 2026：开源 LLMOps 与 AI Agent 平台，自托管指南",[1670,1672,1674,1676,1678],{"title":1671,"url":1403},"Dify 官方文档（中文）",{"title":1673,"url":878},"Dify GitHub",{"title":1675,"url":1593},"Dify 官方定价",{"title":1677,"url":1319},"Coze vs Dify vs FastGPT 选型",{"title":1679,"url":1680},"Dify Self-Hosted Guide 2026","https:\u002F\u002Fjoshuaopolko.com\u002Fdify-self-hosted-guide","tools\u002Fagent\u002Fplatform\u002Fdify","开源 LLMOps 平台，私有部署 Agent 首选",[570,575,1684,1685,1686,1687,1688],"self-host","rag","workflow","llmops","mcp","2026-06-24","想私有部署、想接全球任意模型，Dify 是答案。比 Coze 工程化、上手陡一点；比 FastGPT 工作流强、RAG 略弱。","p5aiXfjt5rD0m3qxj903DVwZMIONVWKdagLa7niYhcE",1785428441007]