[{"data":1,"prerenderedAt":2808},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"review-autoglm-deep-review":9,"review-related-autoglm-deep-review":557},{"tools":4,"reviews":5},77,25,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,13,{"id":10,"title":11,"body":12,"cover":537,"description":538,"extension":539,"meta":540,"navigation":541,"path":542,"published":543,"relatedTools":544,"seo":548,"stem":549,"tags":550,"updated":543,"verdict":555,"__hash__":556},"review\u002Freview\u002Fautoglm-deep-review.md","AutoGLM 深度评测：智谱通用 Agent 能打几分",{"type":13,"value":14,"toc":519},"minimark",[15,19,40,47,62,66,73,83,96,108,112,122,127,133,155,167,173,177,186,212,215,218,228,248,251,257,261,264,270,276,282,288,292,295,350,356,362,372,375,408,411,422,428,434,440,454,458,464,474,480,486,489],[16,17,18],"h2",{"id":18},"一句话结论",[20,21,22,23,27,28,31,32,39],"p",{},"如果你要做",[24,25,26],"strong",{},"中文 App 自治","——让 AI 操作微信、淘宝、美团、小红书完成真实任务——",[24,29,30],{},"AutoGLM 在 2026 年是国内开源 GUI Agent 里覆盖度最高、最值得上手的一个","。智谱清言 + 清华大学合作、32 个月研发、2025-12 开源 Open-AutoGLM（",[33,34,38],"a",{"href":35,"rel":36},"https:\u002F\u002Fxiao9905.github.io\u002FAutoGLM",[37],"nofollow","MIT 模型 + Apache 2.0 代码","），AutoGLM-Phone-9B 单卡 4090 就能跑，覆盖 50+ 中文 App 示例。",[20,41,42,43,46],{},"但它不是万能的：",[24,44,45],{},"9B 模型在复杂多步任务上推理不及 GPT-5 \u002F Claude、iOS 支持空缺、生产级稳定性仍在演进","。它的甜点区是\"中文 App 自治 + 私有化部署 + 学术研究\"，不是\"跨平台桌面自动化\"——后者去 Anthropic Computer Use \u002F Claude Desktop 更合适。",[48,49,54],"div",{"className":50},[51,52,53],"card","p-5","my-4",[20,55,56,57,61],{},"选型建议：学术研究 \u002F 国内 App 自治 \u002F 私有化部署 → AutoGLM 开源版直接上手（成本 = 0 + 一张 4090）。跨平台桌面自动化 \u002F 英文场景 \u002F 生产级稳定性 → 优先 Anthropic Computer Use 或 Claude Desktop。要云端通用 Agent（深度研究、长报告）→ 看 ",[33,58,60],{"href":59},"\u002Fagent\u002Fgeneral\u002Fmanus.html","Manus","。",[16,63,65],{"id":64},"autoglm-真正在解决的问题","AutoGLM 真正在解决的问题",[20,67,68,69,72],{},"社区讨论\"为什么 AutoGLM 重要\"经常停在\"智谱开源、清华背书\"。但深一层看，AutoGLM 是在解决",[24,70,71],{},"中文 GUI Agent 长期以来的三个空白","：",[20,74,75,78,79,82],{},[24,76,77],{},"第一个空白：中文 App 的 GUI Agent。"," Anthropic Computer Use、Claude Desktop 这类产品强在英文桌面场景，对中文 App（微信、美团、淘宝、小红书）的 UI 元素识别和操作链路支持很弱。AutoGLM 把 ",[24,80,81],{},"50+ 中文 App 的任务示例","直接开源出来——微信发消息、淘宝下单、美团点外卖、小红书搜攻略——这是商业产品给不了的中文场景覆盖度。",[20,84,85,88,89,92,93,95],{},[24,86,87],{},"第二个空白：可私有化的手机 Agent。"," 手机里装着通讯录、聊天记录、支付信息，把这些数据交给云端 Agent 跑，合规和隐私都是硬问题。AutoGLM 的 ",[24,90,91],{},"9B 模型 + 框架可以完全本地部署","——模型、日志、权限全在你自己的机器上，数据零外泄。这对金融、政府、医疗等合规敏感场景是刚需，也是 ",[33,94,60],{"href":59}," 这类云端闭源 Agent 给不了的。",[20,97,98,101,102,107],{},[24,99,100],{},"第三个空白：学术研究的开放基准。"," GUI Agent 是前沿研究方向，但商业产品（Manus、Claude Desktop）都是黑盒，研究者拿不到模型权重、看不到推理链路、复现不了 benchmark。AutoGLM 把模型权重、框架代码、Android 适配层、50+ App 任务示例、VAB-WebArena-Lite benchmark 全部释放出来——这是国内学术圈难得的开放姿态，也是它能进 arXiv 论文（",[33,103,106],{"href":104,"rel":105},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.00820",[37],"2411.00820","）的核心价值。",[16,109,111],{"id":110},"agent-能力手机-浏览器双线","Agent 能力：手机 + 浏览器双线",[20,113,114,115,118,119,61],{},"AutoGLM 的 Agent 能力分两条线：",[24,116,117],{},"手机 GUI Agent（Open-AutoGLM 开源版）"," 和 ",[24,120,121],{},"浏览器 Agent（智谱清言 Chrome 扩展）",[123,124,126],"h3",{"id":125},"手机-gui-agent","手机 GUI Agent",[20,128,129,130,72],{},"这是 AutoGLM 最核心、最区别于其他通用 Agent 的能力。工作机制是",[24,131,132],{},"截屏 + UI 树双输入，规划与 grounding 分离",[134,135,136,143,149],"ul",{},[137,138,139,142],"li",{},[24,140,141],{},"规划阶段","：模型看屏幕截图和 UI 树，决定下一步该点哪个元素、输什么文字",[137,144,145,148],{},[24,146,147],{},"Grounding 阶段","：把\"点登录按钮\"这种高层指令映射到屏幕上具体坐标",[137,150,151,154],{},[24,152,153],{},"Safe Operations","：敏感操作（登录、验证码、支付）触发人审接管，避免误操作",[20,156,157,158,162,163,166],{},"按 ",[33,159,161],{"href":35,"rel":160},[37],"AutoGLM 项目主页"," 公布的数据，VAB-WebArena-Lite benchmark ",[24,164,165],{},"首次成功率 55.2%，二次尝试 59.1%","。注：这是学术 benchmark，需第三方验证，但即便打八折也仍是开源 GUI Agent 里的头部。",[20,168,169,172],{},[24,170,171],{},"Remote ADB 是被低估的工程细节","：初次 USB 配置后，通过 WiFi 控制设备，长跑实验摆脱数据线束缚。这对要做批量任务回归的工程团队是实质提升。",[123,174,176],{"id":175},"浏览器-agentchrome-扩展","浏览器 Agent（Chrome 扩展）",[20,178,179,180,185],{},"Web 端通过智谱清言 Chrome 扩展（",[33,181,184],{"href":182,"rel":183},"https:\u002F\u002Fchromewebstore.google.com\u002Fdetail\u002Fmnpdbmgpebfihcndnpgdaihnkmloclkd",[37],"Chrome Web Store，10 万+ 用户","）提供，能力偏\"页面增强\"而非\"全程自治\"：",[134,187,188,194,200,206],{},[137,189,190,193],{},[24,191,192],{},"页面总结","：打开长网页一键总结",[137,195,196,199],{},[24,197,198],{},"划线助手","：选中文字解释 \u002F 翻译 \u002F 续写",[137,201,202,205],{},[24,203,204],{},"写作助手","：在任意输入框唤起 AI 写作",[137,207,208,211],{},[24,209,210],{},"高级检索","：知网 \u002F 知乎 \u002F 小红书跨源检索合成",[20,213,214],{},"注意：Chrome 扩展的定位是\"浏览器内的 AI 助手\"，不是\"全程替你操作浏览器\"的自治 Agent。要全程自治的浏览器 Agent，开源版的 WebArena benchmark 路线更接近，但稳定性仍弱于商业产品。Chrome 扩展评分中等（3.4\u002F5），功能尚可但稳定性一般。",[16,216,217],{"id":217},"多模态屏幕理解",[20,219,220,221,224,225,72],{},"GUI Agent 的核心技术难点是",[24,222,223],{},"屏幕理解","——模型要\"看懂\"屏幕上有什么、能操作什么。AutoGLM 的方案是",[24,226,227],{},"截屏 + UI 树双输入",[134,229,230,236,242],{},[137,231,232,235],{},[24,233,234],{},"截屏（视觉）","：捕捉屏幕像素，理解布局、图标、图片按钮等无文字标签的元素",[137,237,238,241],{},[24,239,240],{},"UI 树（结构）","：通过 Android Accessibility Service 获取 UI 元素的层级、类型、文字，提供精确的可操作坐标",[137,243,244,247],{},[24,245,246],{},"双输入融合","：视觉负责\"这是什么\"，结构负责\"在哪能点\"",[20,249,250],{},"这种双输入方案比纯视觉方案（只看截图）更准——UI 树提供了精确的可点击坐标，避免了纯视觉方案的\"点歪了\"问题。也比纯结构方案（只读 UI 树）更鲁棒——遇到 UI 树缺失或乱序的元素（图片按钮、Canvas 绘制），视觉能补上。",[20,252,253,256],{},[24,254,255],{},"9B 模型的多模态能力上限","：这是 AutoGLM 的核心权衡。9B 参数能在单卡 4090（24GB 显存）跑，门槛友好，但复杂多步任务的推理能力不及 GPT-5 \u002F Claude Opus。社区实测：简单任务（打开 App、搜索、点按钮）成功率高，复杂多步任务（多 App 联动、长链路表单填写）成功率下降明显。要提升上限，可以叠加外部规划模型（用大模型做规划、9B 做 grounding）。",[16,258,260],{"id":259},"中文体验autoglm-的护城河","中文体验：AutoGLM 的护城河",[20,262,263],{},"中文体验是 AutoGLM 相对海外 GUI Agent 最深的护城河，体现在三层：",[20,265,266,269],{},[24,267,268],{},"第一层：中文 App 覆盖度。"," 内置 50+ 中文 App 示例：微信、淘宝、美团、京东、支付宝、抖音、小红书、网易云、大众点评。这些 App 的 UI 元素、操作链路、常见任务模式都被预先适配过。海外产品面对这些 App 基本是从零开始，AutoGLM 是\"开箱即跑\"。",[20,271,272,275],{},[24,273,274],{},"第二层：中文语境理解。"," 模型由智谱清言训练，中文语义理解原生——\"在小红书搜罗马旅游攻略并总结景点\"这种中文指令的意图解析、结果组织，比海外模型更贴中文用户习惯。",[20,277,278,281],{},[24,279,280],{},"第三层：国内合规与数据自主。"," 完全本地部署时数据零外泄，符合国内数据合规要求。读取 IM、通讯录这类敏感操作仍需用户授权 + 符合当地法律，但至少数据不离开你的机器——这是云端 Agent 给不了的。",[20,283,284,287],{},[24,285,286],{},"但中文体验也有边界","：国行 App 的反爬监测和账号风控会偶尔拦截自动化操作，验证码、滑块验证会触发，必要时需要 human-in-the-loop 接管。这不是 AutoGLM 的锅，是国行 App 的反自动化机制——任何 GUI Agent 都会遇到。",[16,289,291],{"id":290},"价格开源的经济学","价格：开源的经济学",[20,293,294],{},"AutoGLM 的价格结构对学术和工程团队极度友好：",[296,297,298,314],"table",{},[299,300,301],"thead",{},[302,303,304,308,311],"tr",{},[305,306,307],"th",{},"形态",[305,309,310],{},"价格",[305,312,313],{},"关键点",[315,316,317,329,339],"tbody",{},[302,318,319,323,326],{},[320,321,322],"td",{},"Open-AutoGLM（开源版）",[320,324,325],{},"$0",[320,327,328],{},"MIT 模型 + Apache 2.0 代码，私有化部署",[302,330,331,334,336],{},[320,332,333],{},"智谱清言 Chrome 扩展",[320,335,325],{},[320,337,338],{},"Chrome Web Store 装即用",[302,340,341,344,347],{},[320,342,343],{},"智谱清言 App 内置版",[320,345,346],{},"商业服务",[320,348,349],{},"完整 GUI agent 能力，价格随订阅",[20,351,352,355],{},[24,353,354],{},"真实成本 = 0（开源）+ GPU 推理","。9B 模型一张 RTX 4090（24GB 显存）就能跑，这是开源 GUI Agent 里门槛最低的之一。",[20,357,358,359,361],{},"对比 ",[33,360,60],{"href":59},"（$20-$200\u002F月、credit 烧得快、大陆屏蔽）和 Anthropic Computer Use（按 token 计费、需海外账号），AutoGLM 的开源模式对学术研究、PoC 验证、私有化部署的成本优势是数量级的。",[20,363,364,367,368,371],{},[24,365,366],{},"但要算清\"隐性成本\"","：开源版需要你会 ADB + Python 部署、要处理 App 版本更新后的 UI 适配、要自己搭推理服务。这些工程成本对纯用户是门槛，对工程团队则可控。",[24,369,370],{},"Open-AutoGLM 的能力 ≠ 智谱清言 App 内置版的能力","——别拿开源版的体验去推断商业整合版，反之亦然。",[16,373,374],{"id":374},"适用场景",[134,376,377,384,390,396,402],{},[137,378,379,380,383],{},"✅ ",[24,381,382],{},"GUI Agent 学术研究 \u002F 工程团队","——开放模型权重 + benchmark + 任务示例，可复现可改进",[137,385,379,386,389],{},[24,387,388],{},"国内 App 自治","——微信 \u002F 美团 \u002F 淘宝 \u002F 小红书等 50+ 中文 App 任务",[137,391,379,392,395],{},[24,393,394],{},"想本地跑 phone agent + 数据自主","——9B 单卡可跑，数据零外泄",[137,397,379,398,401],{},[24,399,400],{},"私有化部署 + 合规需求","——金融 \u002F 医疗 \u002F 政府场景，代码可审计",[137,403,379,404,407],{},[24,405,406],{},"教学 \u002F Demo","——门槛低、中文友好、开源可改",[16,409,410],{"id":410},"不推荐场景",[20,412,413,416,417,421],{},[24,414,415],{},"跨平台桌面自动化","：AutoGLM 主要面向 Android + Web，桌面 agent 能力空缺。要操作 macOS \u002F Windows 桌面应用，去 ",[33,418,420],{"href":419},"\u002Fagent\u002Fdesktop\u002Fclaude-desktop.html","Claude Desktop","（MCP 生态）或 Anthropic Computer Use（跨平台桌面原生）。",[20,423,424,427],{},[24,425,426],{},"iOS App 自治","：iOS 支持完全空缺。iOS 的沙箱和 Accessibility 限制比 Android 严，需要走别的方案，AutoGLM 目前帮不上。",[20,429,430,433],{},[24,431,432],{},"生产级稳定性需求","：学术 \u002F 开源版本仍在演进，App 版本更新后 UI 元素变化要重新适配，复杂多步任务成功率下降明显。要 production-grade 稳定性，目前所有开源 GUI Agent 都还差一截，得等生态成熟或上商业产品。",[20,435,436,439],{},[24,437,438],{},"不懂 ADB \u002F Python 部署的纯用户","：Open-AutoGLM 的部署链路是 git clone → pip install → 下模型 → 配 ADB → 跑示例，对没有 Python \u002F 命令行经验的人门槛偏高。纯用户建议直接用智谱清言 App 内置版，别碰开源版。",[20,441,442,445,446,448,449,453],{},[24,443,444],{},"要云端通用 Agent（深度研究、长报告）","：AutoGLM 是 GUI Agent（操作界面），不是通用 Research Agent（深度调研、写报告）。要\"开着任务下班、明早看带引用的研究报告\"这种场景，去 ",[33,447,60],{"href":59}," 或 ",[33,450,452],{"href":451},"\u002Fagent\u002Fgeneral\u002Fgenspark.html","Genspark","。两者定位不同，别混用。",[16,455,457],{"id":456},"faq","FAQ",[20,459,460,463],{},[24,461,462],{},"Q：Open-AutoGLM 和智谱清言 App 里的 AutoGLM 什么关系？","\nA：原 AutoGLM 是智谱 2024-10 发布的闭源整合产品（手机 + Web GUI agent），首次在真实手机环境跑通完整自治链路。Open-AutoGLM 是 2025-12 开源版本，把核心模型（AutoGLM-Phone-9B）+ 框架代码 + Android 适配层 + 50+ 中文 App 任务示例释放出来。整合版闭源、商业；开源版用于学术 \u002F 私有化。两者能力不等价，别混用。",[20,465,466,469,470,61],{},[24,467,468],{},"Q：AutoGLM 和 Manus 有什么区别？","\nA：定位完全不同。AutoGLM 是 GUI Agent——操作手机 \u002F 浏览器界面完成真实任务（点按钮、填表单、搜索）。Manus 是通用 Research Agent——云端异步完成深度调研、数据分析、写报告。一个操作界面、一个产出报告，详见 ",[33,471,473],{"href":472},"\u002Fcompare\u002Fautoglm-vs-manus.html","AutoGLM vs Manus 对比",[20,475,476,479],{},[24,477,478],{},"Q：9B 模型够用吗？","\nA：简单任务够用，复杂多步任务上限明显。单卡 4090 能跑是门槛优势，但 9B 在推理深度上不及 GPT-5 \u002F Claude。要提升上限可以叠加外部规划模型——大模型做规划、9B 做 grounding 的分工模式是社区常见做法。",[20,481,482,485],{},[24,483,484],{},"Q：国内能直接用吗？","\nA：能。开源版完全本地部署，不依赖外网。智谱清言 Chrome 扩展和 App 都是国内服务，直连无障碍。这是 AutoGLM 相对 Manus（大陆屏蔽）、Anthropic Computer Use（需海外账号）的天然优势。",[16,487,488],{"id":488},"相关阅读",[134,490,491,497,502,508,513],{},[137,492,493],{},[33,494,496],{"href":495},"\u002Fagent\u002Fdesktop\u002Fautoglm.html","AutoGLM 工具卡：智谱 GUI 自治智能体",[137,498,499],{},[33,500,501],{"href":472},"AutoGLM vs Manus：国产通用 Agent 怎么选",[137,503,504],{},[33,505,507],{"href":506},"\u002Freview\u002Fmanus-deep-review.html","Manus 深度评测：通用 Agent 天花板值不值 $20\u002F月",[137,509,510],{},[33,511,512],{"href":419},"Claude Desktop 工具卡：MCP 桌面 Agent",[137,514,515],{},[33,516,518],{"href":517},"\u002Fwiki\u002Fai-agent.html","什么是 AI Agent",{"title":520,"searchDepth":521,"depth":521,"links":522},"",3,[523,525,526,530,531,532,533,534,535,536],{"id":18,"depth":524,"text":18},2,{"id":64,"depth":524,"text":65},{"id":110,"depth":524,"text":111,"children":527},[528,529],{"id":125,"depth":521,"text":126},{"id":175,"depth":521,"text":176},{"id":217,"depth":524,"text":217},{"id":259,"depth":524,"text":260},{"id":290,"depth":524,"text":291},{"id":374,"depth":524,"text":374},{"id":410,"depth":524,"text":410},{"id":456,"depth":524,"text":457},{"id":488,"depth":524,"text":488},"\u002Fog\u002Freview\u002Fautoglm.png","AutoGLM 真实评测：智谱清言 + 清华大学合作的 GUI 自治智能体，2025-12 开源 Open-AutoGLM（MIT 模型 + Apache 2.0 代码），AutoGLM-Phone-9B 覆盖微信 \u002F 淘宝 \u002F 美团等 50+ 中文 App。本文写它真正解决的问题、手机与浏览器操控能力、多模态屏幕理解、中文体验优势、私有部署经济学，以及 5 类不推荐场景。AIHO 编辑部基于官方论文与公开评测整理。","md",{},true,"\u002Freview\u002Fautoglm-deep-review","2026-07-04",[545,546,547],"agent\u002Fdesktop\u002Fautoglm","agent\u002Fgeneral\u002Fmanus","agent\u002Fgeneral\u002Fgenspark",{"title":11,"description":538},"review\u002Fautoglm-deep-review",[551,552,553,554],"AutoGLM","智谱AI","AI Agent","深度评测","国内『AI 操作手机 \u002F 浏览器』开源标杆——智谱 32 个月研发、清华学术背景、MIT 开源模型权重、50+ 中文 App 覆盖领先。中文 App 自治 + 私有化部署 + 学术研究首选。短板是 9B 模型推理上限不及商业大模型、iOS 空缺、生产稳定性仍在演进。","TGPnQJ1Yt5gNeWoqd4d4IimMryaOoVd_aGgNNzhABic",[558,1265,1840],{"id":559,"title":551,"alternatives":560,"api_compatible":564,"body":565,"category":1198,"chinese_friendly":799,"cover":1199,"description":1200,"domestic":1201,"extension":539,"faq":1202,"free":1201,"github":564,"languages":1215,"meta":1218,"models":564,"navigation":541,"notSuitable":564,"opensource":541,"path":1219,"pillar":1220,"platforms":1221,"priceTable":1225,"pricing":1237,"published":1238,"relatedPlaybooks":1239,"relatedReviews":564,"score":1241,"self_host":541,"seo":1242,"seoTitle":564,"slug":545,"sources":1243,"stem":1253,"suitable":564,"tagline":1254,"tags":1255,"updated":1246,"verdict":1263,"website":35,"__hash__":1264},"tools\u002Ftools\u002Fagent\u002Fdesktop\u002Fautoglm.md",[561,562,563],"agent\u002Fdesktop\u002Fclaude-desktop","agent\u002Fdesktop\u002Fopenclaw","coding\u002Fagent\u002Fmanus",null,{"type":13,"value":566,"toc":1186},[567,571,574,577,580,646,648,668,674,678,683,703,708,731,735,896,899,1046,1049,1099,1103,1129,1131,1151,1154,1182],[16,568,570],{"id":569},"tldr","TL;DR",[20,572,573],{},"AutoGLM 是智谱清言 + 清华大学合作的 GUI 自治智能体研究项目（arXiv 2411.00820）。2025-12-08 开源 Open-AutoGLM：AutoGLM-Phone-9B 模型（MIT 协议，HuggingFace \u002F ModelScope 可下载）+ Apache 2.0 框架代码 + Android 适配 + Remote ADB + 50+ 中文 App 示例。VAB-WebArena-Lite 55.2% \u002F 59.1% 成功率，覆盖微信 \u002F 淘宝 \u002F 美团 \u002F Gmail \u002F Google Maps 等。Web 端通过智谱清言 Chrome 扩展（10 万+ 用户）提供页面总结 \u002F 高级检索能力。",[20,575,576],{},"适合：研究 GUI agent 的学术 \u002F 工程团队；国内 App 自治场景；要中文模型 + 私有化部署；想本地跑 phone agent 不依赖外网。不适合：跨平台桌面自动化（用 Anthropic Computer Use \u002F Claude Desktop）；纯英文 web 任务（成功率不如商业产品）；要 production-grade 稳定性（学术 \u002F 开源版本仍在演进）。",[16,578,579],{"id":579},"核心能力",[134,581,582,588,594,600,605,611,617,623,629,634,640],{},[137,583,584,587],{},[24,585,586],{},"AutoGLM-Phone-9B","：MIT 协议开源模型，HuggingFace \u002F ModelScope 可下载",[137,589,590,593],{},[24,591,592],{},"Android 适配层","：原生 Android API + Accessibility Service",[137,595,596,599],{},[24,597,598],{},"Remote ADB","：初次配置后通过 WiFi 控制设备，无需 USB",[137,601,602,604],{},[24,603,217],{},"：截屏 + UI 树双输入，规划 + grounding 分离",[137,606,607,610],{},[24,608,609],{},"50+ 中文 App 任务示例","：微信 \u002F 淘宝 \u002F 美团 \u002F 京东 \u002F 支付宝 \u002F 抖音 \u002F 小红书 \u002F 大众点评等",[137,612,613,616],{},[24,614,615],{},"英文 App 覆盖","：Gmail \u002F Google Maps \u002F X \u002F Reddit \u002F OneStopShop",[137,618,619,622],{},[24,620,621],{},"Web 浏览器智能体","：Chrome 扩展 + VAB-WebArena-Lite 55.2% \u002F 59.1%",[137,624,625,628],{},[24,626,627],{},"页面总结 + 划线助手 + 写作助手 + 高级检索","（知网 \u002F 知乎 \u002F 小红书）",[137,630,631,633],{},[24,632,153],{},"：敏感操作（登录 \u002F 验证码）确认后由人接管",[137,635,636,639],{},[24,637,638],{},"Python API","：几行代码就能跑通自动化任务",[137,641,642,645],{},[24,643,644],{},"私有化部署","：模型 + 框架本地跑，数据 \u002F 日志 \u002F 权限完全自控",[16,647,310],{"id":310},[134,649,650,656,662],{},[137,651,652,655],{},[24,653,654],{},"Open-AutoGLM","：$0；模型 MIT + 代码 Apache 2.0",[137,657,658,661],{},[24,659,660],{},"Chrome 扩展","：$0；Chrome Web Store 装即用",[137,663,664,667],{},[24,665,666],{},"整合产品（智谱清言 App）","：商业服务，价格随订阅",[669,670,671],"blockquote",{},[20,672,673],{},"真实成本 = 0（开源）+ GPU 推理（9B 模型 RTX 4090 一张即可跑）",[16,675,677],{"id":676},"实测开源版本-pixel-6a-ubuntu-2204","实测（开源版本 \u002F Pixel 6a + Ubuntu 22.04）",[20,679,680],{},[24,681,682],{},"亮点：",[134,684,685,688,691,694,697,700],{},[137,686,687],{},"32 个月研发的『手机 GUI agent』终于开源，是国内学术圈难得的开放姿态",[137,689,690],{},"中文 App 覆盖度领先：微信 \u002F 美团 \u002F 淘宝跑通率高",[137,692,693],{},"模型 9B 可在单卡 4090 + 24GB 显存跑，门槛友好",[137,695,696],{},"Remote ADB 让长跑实验摆脱 USB 线",[137,698,699],{},"学术 benchmark 数据透明（55.2% \u002F 59.1%）",[137,701,702],{},"safe-operation 设计对登录 \u002F 支付环节做人审接管，避免误操作",[20,704,705],{},[24,706,707],{},"踩坑：",[134,709,710,713,716,719,722,725,728],{},[137,711,712],{},"App 版本更新后 UI 元素变化要重新适配",[137,714,715],{},"9B 模型在复杂多步骤任务上推理不如商业大模型（GPT-5.4 \u002F Claude Opus 4.6）",[137,717,718],{},"主要面向 Android，iOS 支持空缺",[137,720,721],{},"文档以 README + 论文为主，工程化最佳实践薄",[137,723,724],{},"国行 App 隐私 \u002F 反爬监测会偶尔拦截",[137,726,727],{},"Chrome 扩展 3.4 \u002F 5 评分中等，功能尚可但稳定性一般",[137,729,730],{},"整合版（智谱清言）和开源版差异要注意，别混用",[16,732,734],{"id":733},"上手open-autoglm","上手（Open-AutoGLM）",[736,737,741],"pre",{"className":738,"code":739,"language":740,"meta":520,"style":520},"language-bash shiki shiki-themes github-light github-dark","git clone https:\u002F\u002Fgithub.com\u002Fzai-org\u002FOpen-AutoGLM\ncd Open-AutoGLM && pip install -r requirements.txt\n\n# 下载 9B 模型\nhuggingface-cli download THUDM\u002FAutoGLM-Phone-9B\n\n# 配置 ADB（USB 一次）\nadb tcpip 5555 && adb connect \u003Cdevice-ip>:5555\n\n# 跑示例任务\npython examples\u002Frun_phone_agent.py \\\n  --model THUDM\u002FAutoGLM-Phone-9B \\\n  --task \"在小红书搜索罗马旅游攻略并总结景点\"\n","bash",[742,743,744,760,785,790,797,809,814,820,854,859,865,877,888],"code",{"__ignoreMap":520},[745,746,749,753,757],"span",{"class":747,"line":748},"line",1,[745,750,752],{"class":751},"sScJk","git",[745,754,756],{"class":755},"sZZnC"," clone",[745,758,759],{"class":755}," https:\u002F\u002Fgithub.com\u002Fzai-org\u002FOpen-AutoGLM\n",[745,761,762,766,769,773,776,779,782],{"class":747,"line":524},[745,763,765],{"class":764},"sj4cs","cd",[745,767,768],{"class":755}," Open-AutoGLM",[745,770,772],{"class":771},"sVt8B"," && ",[745,774,775],{"class":751},"pip",[745,777,778],{"class":755}," install",[745,780,781],{"class":764}," -r",[745,783,784],{"class":755}," requirements.txt\n",[745,786,787],{"class":747,"line":521},[745,788,789],{"emptyLinePlaceholder":541},"\n",[745,791,793],{"class":747,"line":792},4,[745,794,796],{"class":795},"sJ8bj","# 下载 9B 模型\n",[745,798,800,803,806],{"class":747,"line":799},5,[745,801,802],{"class":751},"huggingface-cli",[745,804,805],{"class":755}," download",[745,807,808],{"class":755}," THUDM\u002FAutoGLM-Phone-9B\n",[745,810,812],{"class":747,"line":811},6,[745,813,789],{"emptyLinePlaceholder":541},[745,815,817],{"class":747,"line":816},7,[745,818,819],{"class":795},"# 配置 ADB（USB 一次）\n",[745,821,823,826,829,832,834,836,839,843,846,848,851],{"class":747,"line":822},8,[745,824,825],{"class":751},"adb",[745,827,828],{"class":755}," tcpip",[745,830,831],{"class":764}," 5555",[745,833,772],{"class":771},[745,835,825],{"class":751},[745,837,838],{"class":755}," connect",[745,840,842],{"class":841},"szBVR"," \u003C",[745,844,845],{"class":755},"device-i",[745,847,20],{"class":771},[745,849,850],{"class":841},">",[745,852,853],{"class":755},":5555\n",[745,855,857],{"class":747,"line":856},9,[745,858,789],{"emptyLinePlaceholder":541},[745,860,862],{"class":747,"line":861},10,[745,863,864],{"class":795},"# 跑示例任务\n",[745,866,868,871,874],{"class":747,"line":867},11,[745,869,870],{"class":751},"python",[745,872,873],{"class":755}," examples\u002Frun_phone_agent.py",[745,875,876],{"class":764}," \\\n",[745,878,880,883,886],{"class":747,"line":879},12,[745,881,882],{"class":764},"  --model",[745,884,885],{"class":755}," THUDM\u002FAutoGLM-Phone-9B",[745,887,876],{"class":764},[745,889,890,893],{"class":747,"line":8},[745,891,892],{"class":764},"  --task",[745,894,895],{"class":755}," \"在小红书搜索罗马旅游攻略并总结景点\"\n",[16,897,898],{"id":898},"对比",[296,900,901,918],{},[299,902,903],{},[302,904,905,908,910,912,915],{},[305,906,907],{},"维度",[305,909,551],{},[305,911,420],{},[305,913,914],{},"OpenClaw",[305,916,917],{},"Anthropic Computer Use",[315,919,920,937,954,971,985,1001,1015,1029],{},[302,921,922,925,928,931,934],{},[320,923,924],{},"平台",[320,926,927],{},"Android + Web",[320,929,930],{},"macOS \u002F Windows + MCP",[320,932,933],{},"macOS\u002FLinux\u002FWin",[320,935,936],{},"跨平台",[302,938,939,942,945,948,951],{},[320,940,941],{},"开源",[320,943,944],{},"✅ MIT 模型 + Apache 代码",[320,946,947],{},"闭源",[320,949,950],{},"✅ MIT",[320,952,953],{},"❌",[302,955,956,959,962,965,968],{},[320,957,958],{},"中文 App",[320,960,961],{},"✅ 50+ 示例",[320,963,964],{},"–",[320,966,967],{},"部分",[320,969,970],{},"弱",[302,972,973,976,979,981,983],{},[320,974,975],{},"Phone agent",[320,977,978],{},"✅ 9B 模型",[320,980,953],{},[320,982,953],{},[320,984,953],{},[302,986,987,990,992,995,998],{},[320,988,989],{},"桌面 agent",[320,991,953],{},[320,993,994],{},"✅ MCP",[320,996,997],{},"✅ Gateway",[320,999,1000],{},"✅ 原生",[302,1002,1003,1006,1009,1011,1013],{},[320,1004,1005],{},"学术 benchmark",[320,1007,1008],{},"✅ 公开",[320,1010,964],{},[320,1012,964],{},[320,1014,967],{},[302,1016,1017,1020,1022,1024,1027],{},[320,1018,1019],{},"整合 IM",[320,1021,964],{},[320,1023,964],{},[320,1025,1026],{},"✅ 20+ 平台",[320,1028,964],{},[302,1030,1031,1034,1037,1040,1043],{},[320,1032,1033],{},"适合",[320,1035,1036],{},"国内 App \u002F 研究 \u002F 私有化",[320,1038,1039],{},"MCP 生态桌面",[320,1041,1042],{},"IM 触发个人 agent",[320,1044,1045],{},"跨平台桌面",[16,1047,1048],{"id":1048},"避坑",[134,1050,1051,1057,1063,1069,1075,1081,1087,1093],{},[137,1052,1053,1056],{},[24,1054,1055],{},"App 版本敏感","：UI 改动后效果下降，要 pin App 版本或重训样本",[137,1058,1059,1062],{},[24,1060,1061],{},"9B 模型推理上限","：复杂多步任务上不及 GPT-5.4 \u002F Claude，可叠加外部规划模型",[137,1064,1065,1068],{},[24,1066,1067],{},"iOS 暂缺","：需要 iOS agent 走别的方案",[137,1070,1071,1074],{},[24,1072,1073],{},"Remote ADB 安全","：开放端口必须限内网 + 防火墙",[137,1076,1077,1080],{},[24,1078,1079],{},"国行 App 反爬","：账号风控 \u002F 验证码会触发，必要时 human-in-the-loop",[137,1082,1083,1086],{},[24,1084,1085],{},"学术 vs 整合版","：开源版能力 ≠ 智谱清言 App 内置版能力，PoC 别混淆",[137,1088,1089,1092],{},[24,1090,1091],{},"隐私 \u002F 合规","：本地部署是优势，但读取 IM \u002F 通讯录仍要符合用户授权 + 当地法律",[137,1094,1095,1098],{},[24,1096,1097],{},"MIT 模型 + Apache 代码","：商用务必看 LICENSE，企业合规审计要走流程",[16,1100,1102],{"id":1101},"适合-不适合","适合 \u002F 不适合",[134,1104,1105,1108,1111,1114,1117,1120,1123,1126],{},[137,1106,1107],{},"✅ GUI agent 学术研究 \u002F 工程团队",[137,1109,1110],{},"✅ 国内 App 自治（微信 \u002F 美团 \u002F 淘宝 \u002F 小红书）",[137,1112,1113],{},"✅ 想本地跑 phone agent + 数据自主",[137,1115,1116],{},"✅ 私有化部署 + 合规需求",[137,1118,1119],{},"❌ 跨平台桌面自动化（Claude Desktop \u002F Anthropic Computer Use）",[137,1121,1122],{},"❌ iOS App 自治",[137,1124,1125],{},"❌ 要 production-grade 稳定性（仍在演进）",[137,1127,1128],{},"❌ 不懂 ADB \u002F Python 部署的纯用户",[16,1130,488],{"id":488},[134,1132,1133,1139,1145],{},[137,1134,1135],{},[33,1136,1138],{"href":1137},"\u002Ftools\u002Fagent\u002Fdesktop\u002Fclaude-desktop","Claude Desktop 评测",[137,1140,1141],{},[33,1142,1144],{"href":1143},"\u002Ftools\u002Fagent\u002Fdesktop\u002Fopenclaw","OpenClaw 评测",[137,1146,1147],{},[33,1148,1150],{"href":1149},"\u002Ftools\u002Fcoding\u002Fagent\u002Fmanus","Manus 评测",[16,1152,1153],{"id":1153},"来源",[1155,1156,1157,1163,1170,1176],"ol",{},[137,1158,1159,1160],{},"AutoGLM 项目主页（含 Open-AutoGLM 2025-12-08 发布说明 \u002F benchmark 数据）",[33,1161,35],{"href":35,"rel":1162},[37],[137,1164,1165,1166],{},"腾讯新闻 — 智谱开源 AutoGLM（32 个月研发 \u002F MIT + Apache 2.0）",[33,1167,1168],{"href":1168,"rel":1169},"https:\u002F\u002Fnews.qq.com\u002Frain\u002Fa\u002F20251209A05V1G00",[37],[137,1171,1172,1173],{},"AutoGLM arXiv 论文 ",[33,1174,104],{"href":104,"rel":1175},[37],[137,1177,1178,1179],{},"Chrome Web Store — 智谱清言 ChatGLM & AutoGLM 工作学习 AI 助手 ",[33,1180,182],{"href":182,"rel":1181},[37],[1183,1184,1185],"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 .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}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":520,"searchDepth":521,"depth":521,"links":1187},[1188,1189,1190,1191,1192,1193,1194,1195,1196,1197],{"id":569,"depth":524,"text":570},{"id":579,"depth":524,"text":579},{"id":310,"depth":524,"text":310},{"id":676,"depth":524,"text":677},{"id":733,"depth":524,"text":734},{"id":898,"depth":524,"text":898},{"id":1048,"depth":524,"text":1048},{"id":1101,"depth":524,"text":1102},{"id":488,"depth":524,"text":488},{"id":1153,"depth":524,"text":1153},"desktop","\u002Fimg\u002Ftools\u002Fautoglm.webp","AutoGLM 真实评测：智谱清言 + 清华大学合作的 GUI 自治智能体研究产品（arXiv 2411.00820）。2025-12-08 Open-AutoGLM 开源版本（MIT\u002FApache 2.0）发布，含 AutoGLM-Phone-9B 模型（HuggingFace \u002F ModelScope 可下载）+ Android 适配层 + 50+ 中文 App 任务示例。VAB-WebArena-Lite 55.2% 成功率（二次尝试 59.1%），中文 App 支持 Gmail\u002F微信\u002F淘宝\u002F美团等。",false,[1203,1206,1209,1212],{"q":1204,"a":1205},"Open-AutoGLM 和原 AutoGLM 什么关系？","原 AutoGLM 是智谱 2024-10 发布的闭源整合产品（手机 + Web GUI agent），首次在真实手机环境跑通完整自治链路。Open-AutoGLM 是 2025-12 开源版本，把核心模型（AutoGLM-Phone-9B）+ 框架代码 + Android 适配层 + 50+ 中文 App 任务示例释放出来，方便开发者本地部署 + 二次开发。整合版闭源、商业；开源版用于学术 \u002F 私有化。",{"q":1207,"a":1208},"支持哪些 App？","Open-AutoGLM 内置 50+ 中文 App 示例：微信 \u002F 淘宝 \u002F 美团 \u002F 京东 \u002F 支付宝 \u002F 抖音 \u002F 小红书 \u002F 网易云 \u002F 大众点评 + 英文 Gmail \u002F Google Maps \u002F X \u002F Reddit 等。任何 Android App 理论可用，但效果取决于 UI 元素的可识别度，新版本 App 需要重新适配。",{"q":1210,"a":1211},"Web 端能力？","Web 端通过 Chrome 扩展（智谱清言：ChatGLM & AutoGLM 工作学习 AI 助手）提供——10 万+ 用户安装，含页面总结 \u002F 划线助手 \u002F 写作助手 \u002F 高级检索（知网 \u002F 知乎 \u002F 小红书）。VAB-WebArena-Lite benchmark 55.2% 首次成功率，二次尝试 59.1%。",{"q":1213,"a":1214},"和 Anthropic Computer Use \u002F Claude Desktop 怎么选？","AutoGLM 强在中文 App + 学术研究 + 开源权重 + 国内合规；Anthropic Computer Use 强在跨平台桌面 + 英文场景 + 模型推理；Claude Desktop 强在 MCP 生态 + 一键 .mcpb 安装。国内 App 自治 + 私有化 → AutoGLM；英文桌面 + 通用桌面操作 → Anthropic \u002F Claude Desktop。",[1216,1217],"zh","en",{},"\u002Ftools\u002Fagent\u002Fdesktop\u002Fautoglm","agent",[1222,1223,1224],"android","chrome-extension","web",[1226,1229,1232],{"plan":654,"price":325,"features":1227,"notes":1228},"AutoGLM-Phone-9B 模型（MIT）+ 框架代码（Apache 2.0）+ Android 适配 + Remote ADB + 50+ 中文 App 示例","开源 \u002F 私有化部署",{"plan":333,"price":325,"features":1230,"notes":1231},"Web 划线 \u002F 总结 \u002F 高级检索 \u002F 写作助手 \u002F 页面对话","Chrome Web Store",{"plan":1233,"price":1234,"features":1235,"notes":1236},"智谱清言 App 内置","Custom","完整 GUI agent 能力（Phone \u002F Web）+ 整合产品体验","国行 App \u002F 商业服务","Open-AutoGLM 开源免费（MIT 模型 + Apache 2.0 代码）\u002F 智谱清言 App 内置版商业","2026-06-19",[1240],"onboarding\u002Fgui-agent-engineering",{"power":792,"ux":792,"price":799,"cn_support":799,"stability":521},{"title":551,"description":1200},[1244,1247,1249,1251],{"name":1245,"url":35,"accessed":1246},"AutoGLM 项目主页（清华 \u002F Z.AI）","2026-06-24",{"name":1248,"url":1168,"accessed":1246},"腾讯新闻 — 智谱开源 AutoGLM",{"name":1250,"url":104,"accessed":1246},"AutoGLM arXiv 论文 2411.00820",{"name":1252,"url":182,"accessed":1246},"Chrome Web Store — 智谱清言扩展（10万+用户）","tools\u002Fagent\u002Fdesktop\u002Fautoglm","智谱清言出品 + 清华团队 GUI 自治智能体——Open-AutoGLM 开源 9B 手机模型 + Chrome 浏览器扩展",[1256,1257,1258,1259,1260,1261,1262],"gui-agent","mobile-agent","browser-agent","opensource","chinese","zhipu","chatglm","国内『AI 操作手机 \u002F 浏览器』开源标杆——智谱 32 个月研发、清华学术背景、MIT 开源模型权重。学术研究 \u002F 国内 App 自治 \u002F 私有化部署首选。生产稳定性 + 跨平台桌面仍弱于 Anthropic Computer Use \u002F Claude Desktop。","UpD6U6glOmQSbK7O4bjqwqdn7ZYE7JmmNZwkfE87YLc",{"id":1266,"title":452,"alternatives":1267,"api_compatible":564,"body":1270,"category":1782,"chinese_friendly":792,"cover":1783,"description":1784,"domestic":1201,"extension":539,"faq":1785,"free":1201,"github":564,"languages":1798,"meta":1800,"models":564,"navigation":541,"notSuitable":564,"opensource":1201,"path":1801,"pillar":1220,"platforms":1802,"priceTable":1804,"pricing":1815,"published":1238,"relatedPlaybooks":1816,"relatedReviews":564,"score":1818,"self_host":1201,"seo":1819,"seoTitle":564,"slug":547,"sources":1820,"stem":1829,"suitable":564,"tagline":1830,"tags":1831,"updated":1246,"verdict":1837,"website":1838,"__hash__":1839},"tools\u002Ftools\u002Fagent\u002Fgeneral\u002Fgenspark.md",[1268,561,1269],"agent\u002Fgeneral\u002Fflowith","agent\u002Fgeneral\u002Falice",{"type":13,"value":1271,"toc":1770},[1272,1274,1277,1280,1282,1356,1358,1384,1389,1393,1397,1420,1424,1450,1453,1473,1475,1632,1634,1690,1692,1718,1720,1738,1740],[16,1273,570],{"id":569},[20,1275,1276],{},"Genspark 是 2024 年创立于 Palo Alto 的 AI Super Agent 公司，CEO Eric Jing（前 Microsoft Bing \u002F 小冰），团队来自 Microsoft + Google。2026-04 ARR 突破 $250M（12 个月内）。差异点：『多 Agent 架构』路由不同 agent + 模型组合，而非单 LLM + Super Agent 自治执行覆盖研究 \u002F 内容 \u002F 调用 \u002F 建站 \u002F 视频 \u002F 数据分析 \u002F 真实电话 + Sparkpage 可视化输出。Free 100-200 daily credits \u002F Plus $25 月度 12k credits \u002F Pro $249 月度 125k credits。",[20,1278,1279],{},"适合：要一站式 Super Agent 完成完整工作流（研究 → 文档 → 演示 → 网站）；重度内容创作者 + 营销 \u002F 销售（Call For Me 跟进）；中级专业用户（Plus $25 \u002F 月划算）。不适合：纯研究问答（Perplexity \u002F ChatGPT 更便宜）；预算极紧（credit 消耗快）；需要透明引用 \u002F 学术严谨（Sparkpage 引用不如 Perplexity 清晰）；要画布多线程（用 Flowith）。",[16,1281,579],{"id":579},[134,1283,1284,1290,1296,1302,1308,1314,1320,1326,1332,1338,1344,1350],{},[137,1285,1286,1289],{},[24,1287,1288],{},"Super Agent","：自主规划 + 多步执行 + 工具调用",[137,1291,1292,1295],{},[24,1293,1294],{},"多 Agent 架构","：搜索 \u002F 研究 \u002F 内容 \u002F 调用 \u002F 数据分析 agent 协同",[137,1297,1298,1301],{},[24,1299,1300],{},"Sparkpage","：可视化研究输出（含图表 + 引用 + 可分享链接）",[137,1303,1304,1307],{},[24,1305,1306],{},"AI Slides","：10-15 张幻灯片含图表",[137,1309,1310,1313],{},[24,1311,1312],{},"AI Sites","：完整 landing page 一键生成",[137,1315,1316,1319],{},[24,1317,1318],{},"AI Video","：30-60s 短视频生成",[137,1321,1322,1325],{},[24,1323,1324],{},"AI Pods","：10 分钟 AI podcast 生成",[137,1327,1328,1331],{},[24,1329,1330],{},"Call For Me","：调用真实电话完成预订 \u002F 跟进",[137,1333,1334,1337],{},[24,1335,1336],{},"AI Sheets","：CSV 数据分析 + 自动图表",[137,1339,1340,1343],{},[24,1341,1342],{},"Plus 不限聊天","：o3-Pro \u002F Claude \u002F Gemini 顶级模型 chat 无限",[137,1345,1346,1349],{},[24,1347,1348],{},"iOS \u002F Android App","：移动端可用",[137,1351,1352,1355],{},[24,1353,1354],{},"企业 SSO \u002F API","（Pro+）：集成内部系统",[16,1357,310],{"id":310},[134,1359,1360,1366,1372,1378],{},[137,1361,1362,1365],{},[24,1363,1364],{},"Free","：100-200 daily credits（~3-8 简单任务）",[137,1367,1368,1371],{},[24,1369,1370],{},"Plus","：$25\u002F月（年付 $240，~$20\u002F月）；12,000 月度 credits + 顶级模型不限 chat",[137,1373,1374,1377],{},[24,1375,1376],{},"Pro","：$249\u002F月（年付 $2388，~$199\u002F月）；125,000 credits + 优先 + 早期 + 高端 agent",[137,1379,1380,1383],{},[24,1381,1382],{},"Extra credits","：仅 Pro 可买，$10 \u002F 5,000 credits",[669,1385,1386],{},[20,1387,1388],{},"Plus 一月做 10-15 个严肃任务（slides \u002F 研究 \u002F sites）；重度做视频 \u002F 通话 \u002F 多 sites 建议 Pro。",[16,1390,1392],{"id":1391},"实测营销-销售-内容生产","实测（营销 \u002F 销售 \u002F 内容生产）",[20,1394,1395],{},[24,1396,682],{},[134,1398,1399,1402,1405,1408,1411,1414,1417],{},[137,1400,1401],{},"一句话『做一份 B2B SaaS 竞品分析 + 5 张幻灯片 + landing page』全自动跑通",[137,1403,1404],{},"AI Sheets 处理 CSV 数据 + 自动出图 + 写入 Sparkpage 一步到位",[137,1406,1407],{},"Call For Me 跟进客户 \u002F 餐厅预订实际能用，时效感最强",[137,1409,1410],{},"Plus 顶级模型不限 chat 性价比超高，比 ChatGPT Plus $20 + Claude Pro $20 + Perplexity Pro $20 全订便宜",[137,1412,1413],{},"多 Agent 在复杂任务上确实优于单 LLM 的 ChatGPT",[137,1415,1416],{},"Sparkpage 输出可直接分享 + 嵌入",[137,1418,1419],{},"ARR $250M 增长说明产品市场契合度（PMF）非常好",[20,1421,1422],{},[24,1423,707],{},[134,1425,1426,1429,1432,1435,1438,1441,1444,1447],{},[137,1427,1428],{},"credit 消耗不透明，复杂任务前难估算成本",[137,1430,1431],{},"用户反馈 customer support + refund 有问题",[137,1433,1434],{},"生成内容偶尔『AI 味』重，要人工编辑润色",[137,1436,1437],{},"引用 \u002F 来源不如 Perplexity 严谨，学术场景慎用",[137,1439,1440],{},"AI Sites \u002F Video 编辑灵活度低，定制要走外部工具",[137,1442,1443],{},"地域限制：印度 \u002F 巴西 \u002F 巴基斯坦 \u002F 尼日利亚 \u002F 印尼等部分国家 Plus \u002F Pro 不可用",[137,1445,1446],{},"Call For Me 隐私 \u002F 合规要小心（GDPR \u002F 加州 CCPA \u002F 录音同意）",[137,1448,1449],{},"Pro $249 月价格陡，中端断档",[16,1451,1452],{"id":1452},"上手",[1155,1454,1455,1458,1461,1464,1467,1470],{},[137,1456,1457],{},"genspark.ai 注册 → Free 拿 100-200 daily credits",[137,1459,1460],{},"Super Agent 试『一句话目标』：『做一份 AI Agent 2026 竞品分析 Sparkpage + 5 张 Slides』",[137,1462,1463],{},"等 5-10 分钟看 Sparkpage 输出 → 编辑 → 分享链接",[137,1465,1466],{},"AI Sheets 上传 CSV → 自动出图 → 集成进 Sparkpage",[137,1468,1469],{},"Call For Me 试一次小餐厅预订（自有号码 + 同意）",[137,1471,1472],{},"Plus $25 \u002F 月体验完整 → 决定 Pro",[16,1474,898],{"id":898},[296,1476,1477,1494],{},[299,1478,1479],{},[302,1480,1481,1483,1485,1488,1491],{},[305,1482,907],{},[305,1484,452],{},[305,1486,1487],{},"Flowith",[305,1489,1490],{},"Perplexity Pro",[305,1492,1493],{},"ChatGPT Plus",[315,1495,1496,1511,1526,1540,1554,1569,1584,1600,1616],{},[302,1497,1498,1500,1503,1506,1508],{},[320,1499,1288],{},[320,1501,1502],{},"✅ 多 Agent",[320,1504,1505],{},"✅ Neo 自治",[320,1507,964],{},[320,1509,1510],{},"Operator",[302,1512,1513,1516,1519,1522,1524],{},[320,1514,1515],{},"可视化输出",[320,1517,1518],{},"✅ Sparkpage",[320,1520,1521],{},"✅ 画布",[320,1523,964],{},[320,1525,964],{},[302,1527,1528,1531,1534,1536,1538],{},[320,1529,1530],{},"真实电话调用",[320,1532,1533],{},"✅ Call For Me",[320,1535,964],{},[320,1537,964],{},[320,1539,964],{},[302,1541,1542,1545,1548,1550,1552],{},[320,1543,1544],{},"一键建站",[320,1546,1547],{},"✅ AI Sites",[320,1549,964],{},[320,1551,964],{},[320,1553,964],{},[302,1555,1556,1559,1562,1564,1566],{},[320,1557,1558],{},"视频 \u002F 音频生成",[320,1560,1561],{},"✅ Video \u002F Pods",[320,1563,967],{},[320,1565,964],{},[320,1567,1568],{},"Sora（独立）",[302,1570,1571,1574,1577,1579,1582],{},[320,1572,1573],{},"引用透明",[320,1575,1576],{},"中",[320,1578,964],{},[320,1580,1581],{},"✅ 顶级",[320,1583,967],{},[302,1585,1586,1589,1592,1595,1598],{},[320,1587,1588],{},"起价",[320,1590,1591],{},"$25\u002F月（Plus）",[320,1593,1594],{},"$19.90\u002F月",[320,1596,1597],{},"$20\u002F月",[320,1599,1597],{},[302,1601,1602,1605,1608,1611,1613],{},[320,1603,1604],{},"最贵",[320,1606,1607],{},"$249\u002F月",[320,1609,1610],{},"$499.90\u002F月",[320,1612,1597],{},[320,1614,1615],{},"$200\u002F月",[302,1617,1618,1620,1623,1626,1629],{},[320,1619,1033],{},[320,1621,1622],{},"一站式 Super Agent",[320,1624,1625],{},"画布深度",[320,1627,1628],{},"搜索引用",[320,1630,1631],{},"通用",[16,1633,1048],{"id":1048},[134,1635,1636,1642,1648,1654,1660,1666,1672,1678,1684],{},[137,1637,1638,1641],{},[24,1639,1640],{},"Plus 先用一个月","：评估实际 credit 消耗再决定升 Pro",[137,1643,1644,1647],{},[24,1645,1646],{},"复杂任务前估算 credit","：AI Video \u002F Pods \u002F Sites 烧得快",[137,1649,1650,1653],{},[24,1651,1652],{},"Sparkpage 引用要核实","：学术场景不要直接用，把链接打开看原文",[137,1655,1656,1659],{},[24,1657,1658],{},"Call For Me 合规","：欧盟 \u002F 加州录音同意法规要遵守",[137,1661,1662,1665],{},[24,1663,1664],{},"生成内容人工润色","：AI 味需要花 10-20% 时间打磨",[137,1667,1668,1671],{},[24,1669,1670],{},"地域限制","：在受限国家用 VPN + 海外卡，企业合规要走法务",[137,1673,1674,1677],{},[24,1675,1676],{},"AI Sites 不可深度编辑","：要灵活定制走 Webflow \u002F Framer + AI 工具",[137,1679,1680,1683],{},[24,1681,1682],{},"Pro 不要直接上","：Plus 重度可能仍够用，$200 差价省下来",[137,1685,1686,1689],{},[24,1687,1688],{},"Extra credits 性价比低","：$10 \u002F 5k credits 不如直接升 Pro",[16,1691,1102],{"id":1101},[134,1693,1694,1697,1700,1703,1706,1709,1712,1715],{},[137,1695,1696],{},"✅ 营销 \u002F 销售 \u002F 内容创作者 + 要全栈输出",[137,1698,1699],{},"✅ Plus $25 重度多模型不限 chat 用户",[137,1701,1702],{},"✅ 要真实电话调用 \u002F 一键建站 \u002F 短视频",[137,1704,1705],{},"✅ 中级专业用户（自由职业 \u002F 小团队）",[137,1707,1708],{},"❌ 纯研究问答（Perplexity 便宜 + 引用强）",[137,1710,1711],{},"❌ 学术严谨场景",[137,1713,1714],{},"❌ 预算极紧",[137,1716,1717],{},"❌ 要画布 + 多线程深度（用 Flowith）",[16,1719,488],{"id":488},[134,1721,1722,1728,1732],{},[137,1723,1724],{},[33,1725,1727],{"href":1726},"\u002Ftools\u002Fagent\u002Fgeneral\u002Fflowith","Flowith 评测",[137,1729,1730],{},[33,1731,1138],{"href":1137},[137,1733,1734],{},[33,1735,1737],{"href":1736},"\u002Ftools\u002Fagent\u002Fgeneral\u002Falice","Alice 评测",[16,1739,1153],{"id":1153},[1155,1741,1742,1749,1756,1763],{},[137,1743,1744,1745],{},"Genspark Pricing 官方 ",[33,1746,1747],{"href":1747,"rel":1748},"https:\u002F\u002Fgenspark.ai\u002Fpricing",[37],[137,1750,1751,1752],{},"WebCraft — Genspark 2026 Review（Plus \u002F Pro \u002F credit 消耗）",[33,1753,1754],{"href":1754,"rel":1755},"https:\u002F\u002Fwebscraft.org\u002Fblog\u002Fgenspark-ai-oglyad-superagent-yakiy-avtonomno-stvoryuye-sayti-prezentatsiyi?lang=en",[37],[137,1757,1758,1759],{},"Rimo — Genspark 2026 + ARR $250M 数据 ",[33,1760,1761],{"href":1761,"rel":1762},"https:\u002F\u002Frimo.app\u002Fen\u002Fblogs\u002Fgenspark-ai_en-US",[37],[137,1764,1765,1766],{},"Lindy — Genspark Features 2026 Tested ",[33,1767,1768],{"href":1768,"rel":1769},"https:\u002F\u002Fwww.lindy.ai\u002Fblog\u002Fgenspark-ai-features",[37],{"title":520,"searchDepth":521,"depth":521,"links":1771},[1772,1773,1774,1775,1776,1777,1778,1779,1780,1781],{"id":569,"depth":524,"text":570},{"id":579,"depth":524,"text":579},{"id":310,"depth":524,"text":310},{"id":1391,"depth":524,"text":1392},{"id":1452,"depth":524,"text":1452},{"id":898,"depth":524,"text":898},{"id":1048,"depth":524,"text":1048},{"id":1101,"depth":524,"text":1102},{"id":488,"depth":524,"text":488},{"id":1153,"depth":524,"text":1153},"general","\u002Fimg\u002Ftools\u002Fgenspark.webp","Genspark 真实评测：2024 年创立于 Palo Alto，CEO Eric Jing（前 Microsoft Bing \u002F 小冰），团队来自 Microsoft + Google。差异点：『多 Agent 架构』而非单 LLM + Super Agent 自治执行（研究 \u002F 内容 \u002F 调用 \u002F 建站 \u002F 视频 \u002F 数据分析）+ Sparkpage 可视化输出 + Call For Me 真实电话调度。2026-04 ARR $250M（12 个月内达成）。Free 100-200 daily credits \u002F Plus $25 月度 12k credits \u002F Pro $249 月度 125k credits。",[1786,1789,1792,1795],{"q":1787,"a":1788},"credit 消耗如何？","Simple Sparkpage \u002F 研究 30-80 credits；AI Slides（10-15 张 + 图表）250-450；AI Sites（完整 landing）600-1200；AI Video（30-60s）800-2000；Call For Me（2-4 分钟）400-900；AI Pods（10 分钟 podcast）1000-1800。Plus 12k credits 大约一个月做 10-15 个严肃任务（不重度跑视频 \u002F 通话）。",{"q":1790,"a":1791},"Call For Me 真的能打电话？","对，调用真实电话 API（背后用 Bland AI \u002F Vapi 类供应商），帮你预订餐厅 \u002F 跟进客户 \u002F 查信息。需要明确隐私 + 合规：欧盟 \u002F 加州的呼叫录音 \u002F 同意法规要遵守，敏感场景慎用。",{"q":1793,"a":1794},"多 Agent 架构和单 LLM 区别？","Genspark 不依赖单一 LLM，而是按任务类型路由到不同 agent + 模型组合（搜索 agent \u002F 研究 agent \u002F 内容 agent \u002F 调用 agent 等）。优势：每类任务用最适合的工具 + 大型搜索 + 验证流水线；劣势：黑盒程度高 + credit 消耗模型让用量不可预测。",{"q":1796,"a":1797},"和 Flowith \u002F ChatGPT \u002F Perplexity 怎么选？","Genspark 强在『全栈 Super Agent + 真实调用（电话 \u002F 搜索 \u002F 建站）+ Sparkpage 可视化』。Flowith 强在『画布 + 长任务自治 + Knowledge Garden』。Perplexity 强在『搜索 + 引用透明 + 价格便宜』。ChatGPT 强在『通用 + 生态成熟』。一站式 Super Agent → Genspark；画布多线程 → Flowith；搜索问答 → Perplexity。",[1217,1216,1799],"multi",{},"\u002Ftools\u002Fagent\u002Fgeneral\u002Fgenspark",[1224,1803,1222],"ios",[1805,1808,1812],{"plan":1364,"price":325,"features":1806,"notes":1807},"100-200 daily credits + 基础模型 + Sparkpage \u002F 研究","试水",{"plan":1370,"price":1809,"features":1810,"notes":1811},"$25\u002F月","12,000 月度 credits + 顶级模型不限聊天（o3-Pro \u002F Claude \u002F Gemini）+ Slides \u002F 研究","年付 $240",{"plan":1376,"price":1607,"features":1813,"notes":1814},"125,000 月度 credits + 优先速度 + 早期功能 + AI Sites \u002F Video \u002F Call","年付 $2388","Free (100-200 daily) \u002F Plus $25·月 (12k credits) \u002F Pro $249·月 (125k credits) \u002F 年付有折扣",[1817],"onboarding\u002Fsuper-agent-workflow",{"power":799,"ux":792,"price":521,"cn_support":521,"stability":792},{"title":452,"description":1784},[1821,1823,1825,1827],{"name":1822,"url":1747,"accessed":1246},"Genspark 官网 + Pricing",{"name":1824,"url":1754,"accessed":1246},"WebCraft — Genspark 2026 Review (Plus \u002F Pro \u002F Use Cases)",{"name":1826,"url":1761,"accessed":1246},"Rimo — Genspark 2026 + ARR $250M",{"name":1828,"url":1768,"accessed":1246},"Lindy — Genspark Features Tested 2026","tools\u002Fagent\u002Fgeneral\u002Fgenspark","Palo Alto 出品的 AI Super Agent——多 Agent 架构 + Sparkpage \u002F Slides \u002F Sites \u002F Video \u002F Call For Me 全栈",[1832,1833,1834,1835,1836],"super-agent","multi-agent","sparkpage","call-agent","genspark","Super Agent 类目里目前最完整的一站式产品——研究 \u002F 内容 \u002F 建站 \u002F 视频 \u002F 电话调度全包。Plus $25 \u002F 月在重度场景非常划算。要纯研究问答用 Perplexity \u002F ChatGPT 更便宜。","https:\u002F\u002Fgenspark.ai","4z6BpXPUsDg0c4H6t-MNdIES6Gq2hbI5zP2Qe-fl4uU",{"id":1841,"title":60,"alternatives":1842,"api_compatible":1846,"body":1847,"category":1782,"chinese_friendly":792,"cover":2734,"description":2735,"domestic":1201,"extension":539,"faq":564,"free":1201,"github":564,"languages":2736,"meta":2737,"models":2738,"navigation":541,"notSuitable":2743,"opensource":1201,"path":2748,"pillar":1220,"platforms":2749,"priceTable":2753,"pricing":2768,"published":2769,"relatedPlaybooks":2770,"relatedReviews":2772,"score":2774,"self_host":1201,"seo":2775,"seoTitle":564,"slug":546,"sources":2776,"stem":2792,"suitable":2793,"tagline":2799,"tags":2800,"updated":1246,"verdict":2806,"website":2224,"__hash__":2807},"tools\u002Ftools\u002Fagent\u002Fgeneral\u002Fmanus.md",[547,1843,1844,1845],"agent\u002Fgeneral\u002Felicit","agent\u002Fgeneral\u002Fdevin","agent\u002Fplatform\u002Fdify",[],{"type":13,"value":1848,"toc":2714},[1849,1851,1884,1889,1892,1956,1959,1963,1971,1976,1983,1987,1995,2061,2066,2070,2077,2081,2089,2094,2111,2114,2117,2120,2127,2130,2136,2190,2196,2202,2207,2213,2217,2260,2263,2301,2304,2476,2481,2498,2503,2523,2526,2592,2594,2597,2614,2617,2638,2640,2688,2690,2707],[16,1850,570],{"id":569},[48,1852,1854,1870],{"className":1853},[51,52,53],[20,1855,1856,1859,1860,1863,1864,1869],{},[24,1857,1858],{},"一句话："," Butterfly Effect（中国创办、新加坡注册）2025-03-06 首发的通用 AI Agent，",[24,1861,1862],{},"邀请码一度被炒到 ¥5 万-10 万","（",[33,1865,1868],{"href":1866,"rel":1867},"https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FManus_%28AI_agent%29",[37],"据维基百科引用 China Daily 报道","）。多 sub-agent 并行架构——浏览、数据分析、代码执行、写作各自专门子 agent，主 orchestrator 自动路由到 Claude \u002F Qwen \u002F 自研模型，结果合成交付。",[20,1871,1872,1873,1863,1876,1880,1881,61],{},"2025-12 传被 ",[24,1874,1875],{},"Meta 以约 20-30 亿美元收购",[33,1877,1879],{"href":1866,"rel":1878},[37],"Reuters \u002F AP 报道","），目前仍独立运营。最大价值是 ",[24,1882,1883],{},"deep research 类任务的引用密度和质量明显超过 ChatGPT Deep Research",[669,1885,1886],{},[20,1887,1888],{},"来源说明：本文基于 manus.im 官方页面、Wikipedia \"Manus (AI agent)\" 条目、第三方评测（Pick Right \u002F HyzenPro \u002F Info-Tech Research Group \u002F 36Kr \u002F 世界经济论坛企业页）综合整理。Meta 收购案细节、产品路线图仍在变化中，请以最新官方公告为准。",[16,1890,1891],{"id":1891},"背景与公司",[134,1893,1894,1905,1911,1917,1923,1934,1944,1950],{},[137,1895,1896,1899,1900,1904],{},[24,1897,1898],{},"公司","：Butterfly Effect Pte. Ltd.（",[33,1901,1903],{"href":1866,"rel":1902},[37],"蝴蝶效应","），创始人 Xiao Hong（季逸超 Ji Yichao 为 Manus 联合创始人 + 首席科学家）",[137,1906,1907,1910],{},[24,1908,1909],{},"公司位置","：办公室在北京 + 武汉 + 新加坡，目标市场北美 \u002F 日本 \u002F 韩国（不主打中国大陆）",[137,1912,1913,1916],{},[24,1914,1915],{},"前作","：Monica，2023 年发布的浏览器扩展 AI 助手",[137,1918,1919,1922],{},[24,1920,1921],{},"历史融资","：2024 年字节跳动曾出价 ~3000 万美元收购被拒（据 36Kr）",[137,1924,1925,1928,1929,1933],{},[24,1926,1927],{},"Manus 启动","：2024-10 立项，灵感来自 ",[33,1930,1932],{"href":1931},"\u002Fcoding\u002Fide\u002Fcursor.html","Cursor","；名字来自 MIT 拉丁校训 \"Mens et Manus\"（手与脑）",[137,1935,1936,1939,1940],{},[24,1937,1938],{},"首发数据","：2025-03-06 邀请制 beta，7 天内 200 万人候补，",[33,1941,1943],{"href":1866,"rel":1942},[37],"Demo 视频 20 小时百万播放",[137,1945,1946,1949],{},[24,1947,1948],{},"营收","：2025-08 ARR 约 9000 万美元 → 2025-12 升至 1.25 亿美元",[137,1951,1952,1955],{},[24,1953,1954],{},"Meta 收购","：2025-12-29 宣布，估值 20-30 亿美元，目前仍独立运营，但中国大陆已被屏蔽访问 + 关闭中文社交账号",[16,1957,1958],{"id":1958},"核心特性",[123,1960,1962],{"id":1961},"多-sub-agent-并行架构最大差异化","多 sub-agent 并行架构（最大差异化）",[20,1964,1965,1970],{},[33,1966,1969],{"href":1967,"rel":1968},"https:\u002F\u002Fpick-right.com\u002Ftools\u002Fmanus-ai",[37],"Pick Right 2026-04 评测"," 描述的架构：",[669,1972,1973],{},[20,1974,1975],{},"\"Where most general-purpose AI products use one model end-to-end, Manus runs multiple specialized sub-agents in parallel: one handles web browsing, one handles data analysis, one handles code execution, one handles synthesis and writing.\"",[20,1977,1978,1979,1982],{},"主 orchestrator 给每一步选最合适的模型（Claude \u002F Qwen \u002F 自研），结果合成。这是 Manus 研究输出 ",[24,1980,1981],{},"引用密度更高、幻觉更少","的工程根源——分工 + 并行让每个 sub-agent 只做自己最擅长的部分。",[123,1984,1986],{"id":1985},"通用-agent-模式","通用 Agent 模式",[20,1988,1989,1994],{},[33,1990,1993],{"href":1991,"rel":1992},"https:\u002F\u002Fwww.infotech.com\u002Fresearch\u002Fassessing-manus-the-future-of-agentic-ai",[37],"Info-Tech 评测"," 列出的 GAIA 基准（General AI Assistants）：",[296,1996,1997,2007],{},[299,1998,1999],{},[302,2000,2001,2004],{},[305,2002,2003],{},"模型",[305,2005,2006],{},"GAIA 准确率",[315,2008,2009,2021,2029,2037,2045,2053],{},[302,2010,2011,2016],{},[320,2012,2013],{},[24,2014,2015],{},"Manus AI",[320,2017,2018],{},[24,2019,2020],{},">65%（SOTA）",[302,2022,2023,2026],{},[320,2024,2025],{},"H2O.ai (h2oGPTe)",[320,2027,2028],{},"65%",[302,2030,2031,2034],{},[320,2032,2033],{},"Google Langfun",[320,2035,2036],{},"49%",[302,2038,2039,2042],{},[320,2040,2041],{},"Microsoft o1",[320,2043,2044],{},"38%",[302,2046,2047,2050],{},[320,2048,2049],{},"OpenAI GPT-4o",[320,2051,2052],{},"32%",[302,2054,2055,2058],{},[320,2056,2057],{},"OpenAI GPT-4 + Plugins",[320,2059,2060],{},"15-30%",[669,2062,2063],{},[20,2064,2065],{},"注：基准数据需第三方验证；Manus 官方公布数据，请审慎参考。",[123,2067,2069],{"id":2068},"后台执行-长任务","后台执行 + 长任务",[20,2071,2072,2073,2076],{},"最特别的体验：任务下达后",[24,2074,2075],{},"可以关掉浏览器","，Manus 在云端继续跑（小时级），完成时通知。这种\"开着任务下班\"的模式是它 viral 的关键。",[123,2078,2080],{"id":2079},"web-app-builder","Web App Builder",[20,2082,2083,2084,2088],{},"直接生成完整网站和应用，内置数据库 + Stripe 支付 + SEO。但 ",[33,2085,2087],{"href":1967,"rel":2086},[37],"Pick Right 评测"," 直白警告：",[669,2090,2091],{},[20,2092,2093],{},"\"Promising but buggy enough that I wouldn't ship to production from it yet.\"",[20,2095,2096,2097,2101,2102,2101,2106,2110],{},"复杂场景下出 bug 多，目前不建议生产部署，",[33,2098,2100],{"href":2099},"\u002Fcoding\u002Fbuilder\u002Fbolt-new.html","Bolt.new"," \u002F ",[33,2103,2105],{"href":2104},"\u002Fcoding\u002Fbuilder\u002Flovable.html","Lovable",[33,2107,2109],{"href":2108},"\u002Fcoding\u002Fbuilder\u002Fv0.html","v0"," 仍是 production app building 的更稳选项。",[123,2112,2113],{"id":2113},"桌面应用",[20,2115,2116],{},"提供 desktop app，能读本地文件 + 集成你的机器，不仅限于浏览器内任务。",[123,2118,2119],{"id":2119},"多模型路由",[20,2121,2122,2123,2126],{},"Claude \u002F Qwen \u002F Manus 自研模型，",[24,2124,2125],{},"按任务步骤自动选","——这是 Manus 跟单一模型 agent 的根本差异。",[16,2128,2129],{"id":2129},"价格与运行成本",[20,2131,2132,2135],{},[33,2133,2087],{"href":1967,"rel":2134},[37]," 公开档位：",[296,2137,2138,2149],{},[299,2139,2140],{},[302,2141,2142,2145,2147],{},[305,2143,2144],{},"档位",[305,2146,310],{},[305,2148,313],{},[315,2150,2151,2160,2169,2179],{},[302,2152,2153,2155,2157],{},[320,2154,1364],{},[320,2156,325],{},[320,2158,2159],{},"每日有限 credits，够 1 个高强度任务\u002F天",[302,2161,2162,2164,2166],{},[320,2163,1376],{},[320,2165,1597],{},[320,2167,2168],{},"大多数付费用户落点；中等 credit + 多模型",[302,2170,2171,2173,2176],{},[320,2172,1370],{},[320,2174,2175],{},"$50\u002F月",[320,2177,2178],{},"更高 credit + 优先队列，10+ 任务\u002F周",[302,2180,2181,2184,2187],{},[320,2182,2183],{},"Pro+ \u002F Team",[320,2185,2186],{},"最高 $200\u002F月",[320,2188,2189],{},"最大 credit + 团队空间（功能仍有限）",[20,2191,2192,2195],{},[24,2193,2194],{},"credit 经济学","：每个动作消耗 credits（浏览、代码运行、模型调用都计费）。",[20,2197,2198,72],{},[33,2199,2201],{"href":1967,"rel":2200},[37],"Pick Right 真实使用反馈",[669,2203,2204],{},[20,2205,2206],{},"\"Credits run out faster than the pricing page suggests. Heavy users routinely buy credit packs on top of subscriptions.\"",[20,2208,2209,2212],{},[24,2210,2211],{},"预算建议","：先用 Free 跑 3-5 个真实任务评估消耗速度，再决定档位。",[16,2214,2216],{"id":2215},"上手-5-分钟","上手 5 分钟",[1155,2218,2219,2227,2230,2245,2248,2254,2257],{},[137,2220,2221,2222],{},"打开 ",[33,2223,2226],{"href":2224,"rel":2225},"https:\u002F\u002Fmanus.im",[37],"manus.im",[137,2228,2229],{},"账号注册（Google \u002F Apple OAuth 最快，国内手机号注册受限）",[137,2231,2232,2233,2236,2237],{},"给一个完整任务描述（",[24,2234,2235],{},"关键","：不要碎片化指令，给全场景）\n",[736,2238,2243],{"className":2239,"code":2241,"language":2242},[2240],"language-text","\"调研欧洲前 10 大 EV 充电网络（覆盖率、价格、可靠性、充电速度），\n产出带引用的 Markdown 对比表\"\n","text",[742,2244,2241],{"__ignoreMap":520},[137,2246,2247],{},"选模型路由（Auto 推荐）",[137,2249,2250,2251,2253],{},"提交后",[24,2252,2075],{},"——任务在云端跑",[137,2255,2256],{},"完成后邮件 \u002F 站内通知",[137,2258,2259],{},"看结果 \u002F 下载交付物（Markdown \u002F Excel \u002F Word \u002F 网页）",[16,2261,2262],{"id":2262},"国内使用注意事项",[1155,2264,2265,2275,2281,2287],{},[137,2266,2267,72,2270,2274],{},[24,2268,2269],{},"大陆访问被屏蔽",[33,2271,2273],{"href":1866,"rel":2272},[37],"Wikipedia 引用 36Kr 报道","，Butterfly Effect 已关闭中文社交账号、阻断中国大陆访问，原 Alibaba Qwen 合作版\"中文版 Manus\"已搁置",[137,2276,2277,2280],{},[24,2278,2279],{},"访问需稳定代理","：日韩 \u002F 美国节点",[137,2282,2283,2286],{},[24,2284,2285],{},"账号 \u002F 支付","：海外信用卡（Visa \u002F MasterCard），第三方代付方案有限",[137,2288,2289,2292,2293,2295,2296,2300],{},[24,2290,2291],{},"替代路径","：国内可考虑 ",[33,2294,452],{"href":451}," \u002F Devv \u002F ",[33,2297,2299],{"href":2298},"\u002Fagent\u002Fgeneral\u002Fflowith.html","元宝 Yuanbao"," \u002F 秘塔 Metaso 等",[16,2302,2303],{"id":2303},"与同类怎么选",[296,2305,2306,2330],{},[299,2307,2308],{},[302,2309,2310,2312,2314,2320,2324,2327],{},[305,2311,907],{},[305,2313,60],{},[305,2315,2316],{},[33,2317,2319],{"href":2318},"\u002Fcoding\u002Fagent\u002Fdevin.html","Devin",[305,2321,2322],{},[33,2323,452],{"href":451},[305,2325,2326],{},"ChatGPT Deep Research",[305,2328,2329],{},"GenAgent",[315,2331,2332,2351,2371,2390,2405,2421,2438,2456],{},[302,2333,2334,2337,2340,2343,2346,2349],{},[320,2335,2336],{},"核心定位",[320,2338,2339],{},"通用 Agent",[320,2341,2342],{},"AI 程序员",[320,2344,2345],{},"AI 搜索 + Agent",[320,2347,2348],{},"LLM Deep Research",[320,2350,1631],{},[302,2352,2353,2356,2359,2362,2365,2368],{},[320,2354,2355],{},"架构",[320,2357,2358],{},"多 sub-agent 并行",[320,2360,2361],{},"单 Agent + 沙盒",[320,2363,2364],{},"多模型",[320,2366,2367],{},"单模型",[320,2369,2370],{},"—",[302,2372,2373,2376,2379,2382,2385,2388],{},[320,2374,2375],{},"长任务",[320,2377,2378],{},"★★★★★ 小时级",[320,2380,2381],{},"★★★★★",[320,2383,2384],{},"★★★☆☆",[320,2386,2387],{},"★★★★☆",[320,2389,2384],{},[302,2391,2392,2395,2397,2399,2401,2403],{},[320,2393,2394],{},"引用密度",[320,2396,2381],{},[320,2398,2384],{},[320,2400,2387],{},[320,2402,2384],{},[320,2404,2384],{},[302,2406,2407,2410,2413,2415,2417,2419],{},[320,2408,2409],{},"App Builder",[320,2411,2412],{},"⚠️ 有但 buggy",[320,2414,953],{},[320,2416,953],{},[320,2418,953],{},[320,2420,953],{},[302,2422,2423,2426,2429,2432,2434,2436],{},[320,2424,2425],{},"中文",[320,2427,2428],{},"⚠️ 大陆屏蔽",[320,2430,2431],{},"⚠️",[320,2433,2381],{},[320,2435,2431],{},[320,2437,2387],{},[302,2439,2440,2442,2445,2448,2451,2454],{},[320,2441,310],{},[320,2443,2444],{},"$20-$200",[320,2446,2447],{},"$500\u002F月",[320,2449,2450],{},"$24.99\u002F月",[320,2452,2453],{},"随 ChatGPT Plus",[320,2455,2370],{},[302,2457,2458,2461,2464,2467,2470,2473],{},[320,2459,2460],{},"适合场景",[320,2462,2463],{},"research \u002F 数据分析",[320,2465,2466],{},"写代码 \u002F 修 bug",[320,2468,2469],{},"信息检索",[320,2471,2472],{},"单次深度研究",[320,2474,2475],{},"综合",[20,2477,2478,72],{},[24,2479,2480],{},"选 Manus 如果你",[134,2482,2483,2489,2492,2495],{},[137,2484,2485,2486],{},"重视 research \u002F 数据分析任务的 ",[24,2487,2488],{},"引用密度和结构化输出",[137,2490,2491],{},"想试\"开任务下班、明早看结果\"的工作流",[137,2493,2494],{},"海外 \u002F 能解决账号网络问题",[137,2496,2497],{},"预算 $20-$50\u002F月，重度用户",[20,2499,2500,72],{},[24,2501,2502],{},"别选 Manus 如果你",[134,2504,2505,2514,2517,2520],{},[137,2506,2507,2508,2101,2510,2513],{},"国内裸用（",[33,2509,452],{"href":451},[33,2511,2512],{"href":2298},"Yuanbao"," 更顺）",[137,2515,2516],{},"想生产部署 App（Web App Builder bug 多）",[137,2518,2519],{},"团队协作场景（功能不完善）",[137,2521,2522],{},"预算 \u003C $20\u002F月（Free 档够评估，付费档不一定划算）",[16,2524,2525],{"id":2525},"避坑清单",[134,2527,2528,2534,2544,2550,2556,2568,2574,2580],{},[137,2529,2530,2533],{},[24,2531,2532],{},"大陆访问已被官方屏蔽","：2025 年起阻断中国大陆访问，原\"中文版 Manus\"项目搁置",[137,2535,2536,72,2539,2543],{},[24,2537,2538],{},"credit 烧得比官方页面暗示的快",[33,2540,2542],{"href":1967,"rel":2541},[37],"Pick Right 2026 评测"," 真实反馈，重度用户经常额外买 credit pack",[137,2545,2546,2549],{},[24,2547,2548],{},"任务一启动无法控预算","：开始跑后只能 cancel 或等结束，credits 会一直消耗",[137,2551,2552,2555],{},[24,2553,2554],{},"Web App Builder 别上生产","：demo 漂亮，复杂场景翻车，Bolt.new \u002F Lovable \u002F v0 仍是生产部署更稳选项",[137,2557,2558,2561,2562,2567],{},[24,2559,2560],{},"每次任务从零开始","：没有持久 workspace，不像 ",[33,2563,2566],{"href":2564,"rel":2565},"https:\u002F\u002Fclaude.com",[37],"Claude Projects"," \u002F ChatGPT Custom GPTs 能跨会话记忆",[137,2569,2570,2573],{},[24,2571,2572],{},"没有 HubSpot \u002F Salesforce \u002F Notion \u002F Slack 集成","：拿到结果后要自己手动搬到工具栈",[137,2575,2576,2579],{},[24,2577,2578],{},"团队协作能力差","：单用户产品为主，无共享 workspace \u002F 评论 \u002F 审计",[137,2581,2582,2585,2586,2591],{},[24,2583,2584],{},"Meta 收购的不确定性","：Wikipedia 引用 ",[33,2587,2590],{"href":2588,"rel":2589},"https:\u002F\u002Fwww.nytimes.com\u002F2026\u002F03\u002F17\u002Ftechnology\u002Fchina-scrutiny-meta-manus.html",[37],"纽时 2026-03"," 报道，中国审查 Meta 收购案，长期路线图待观察",[16,2593,1102],{"id":1101},[20,2595,2596],{},"✅ 适合：",[134,2598,2599,2602,2605,2608,2611],{},[137,2600,2601],{},"委托多步骤 research（\"调研 X，产出 Markdown 对比表\"）",[137,2603,2604],{},"CSV \u002F Excel 重的数据分析任务",[137,2606,2607],{},"长任务 + 不想盯着看（小时级）",[137,2609,2610],{},"单兵作战的咨询顾问 \u002F 分析师 \u002F 创业者",[137,2612,2613],{},"对\"通用 Agent 体感天花板\"感兴趣的测试者",[20,2615,2616],{},"❌ 不适合：",[134,2618,2619,2622,2625,2628,2631],{},[137,2620,2621],{},"生产级应用部署",[137,2623,2624],{},"团队协作工作流",[137,2626,2627],{},"预算极敏感（free 档非常受限）",[137,2629,2630],{},"大陆稳定访问需求",[137,2632,2633,2634,2637],{},"需要深度持久上下文（用 ",[33,2635,2566],{"href":2564,"rel":2636},[37]," \u002F ChatGPT Custom GPT 等）",[16,2639,488],{"id":488},[134,2641,2642,2653,2662,2677],{},[137,2643,2644,2645,2101,2647,2649,2650,2652],{},"同类对比：",[33,2646,2319],{"href":2318},[33,2648,452],{"href":451}," \u002F Elicit \u002F ",[33,2651,2512],{"href":2298}," \u002F Metaso",[137,2654,2655,2656,2101,2658,2661],{},"概念：",[33,2657,553],{"href":517},[33,2659,2660],{"href":517},"Multi-Agent"," \u002F Computer Use \u002F Deep Research",[137,2663,2664,2665,2101,2669,2101,2673],{},"模型：",[33,2666,2668],{"href":2667},"\u002Fmodels\u002Fclaude-sonnet-4.html","Claude Sonnet 4",[33,2670,2672],{"href":2671},"\u002Fmodels\u002Fclaude-opus-4.html","Claude Opus 4",[33,2674,2676],{"href":2675},"\u002Fmodels\u002Fqwen-3.html","Qwen3",[137,2678,2679,2680,2101,2684],{},"进阶：",[33,2681,2683],{"href":2682},"\u002Fwiki\u002Fcontext-engineering.html","Context Engineering",[33,2685,2687],{"href":2686},"\u002Fwiki\u002Fprompt-engineering.html","Prompt Engineering",[16,2689,1153],{"id":1153},[134,2691,2692,2698,2701,2704],{},[137,2693,2694,2695],{},"官网：",[33,2696,2224],{"href":2224,"rel":2697},[37],[137,2699,2700],{},"Wikipedia：\"Manus (AI agent)\" 条目",[137,2702,2703],{},"第三方评测：pick-right.com \u002F hyzenpro.com \u002F infotech.com \u002F weforum.org",[137,2705,2706],{},"媒体报道：Reuters \u002F AP \u002F China Daily \u002F 36Kr \u002F NYT",[20,2708,2709,2710,2713],{},"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；如发现价格 \u002F 功能 \u002F 公司状态与最新官方信息不一致，请通过 ",[33,2711,2712],{"href":2712},"\u002Fsubmit"," 反馈。",{"title":520,"searchDepth":521,"depth":521,"links":2715},[2716,2717,2718,2726,2727,2728,2729,2730,2731,2732,2733],{"id":569,"depth":524,"text":570},{"id":1891,"depth":524,"text":1891},{"id":1958,"depth":524,"text":1958,"children":2719},[2720,2721,2722,2723,2724,2725],{"id":1961,"depth":521,"text":1962},{"id":1985,"depth":521,"text":1986},{"id":2068,"depth":521,"text":2069},{"id":2079,"depth":521,"text":2080},{"id":2113,"depth":521,"text":2113},{"id":2119,"depth":521,"text":2119},{"id":2129,"depth":524,"text":2129},{"id":2215,"depth":524,"text":2216},{"id":2262,"depth":524,"text":2262},{"id":2303,"depth":524,"text":2303},{"id":2525,"depth":524,"text":2525},{"id":1101,"depth":524,"text":1102},{"id":488,"depth":524,"text":488},{"id":1153,"depth":524,"text":1153},"\u002Fimg\u002Ftools\u002Fmanus.webp","Manus 真实评测：Butterfly Effect（蝴蝶效应）出品的通用 AI Agent，多 sub-agent 并行架构 + 多模型路由（Claude \u002F Qwen \u002F 自研）。2025-03 首发即引爆，2025-12 传被 Meta 以约 20 亿美元收购。AIHO 编辑部基于官方文档与多份评测整理。",[1216,1217],{},[2739,2740,2741,2742],"claude-sonnet-4","claude-opus-4","qwen-max","manus-internal",[2744,2745,2746,2747],"生产级 Web App 部署（Web App Builder 仍有 bug）","团队协作场景（功能未完善）","需要持久工作空间 \u002F 跨会话上下文","对成本极敏感（credits 烧得快）","\u002Ftools\u002Fagent\u002Fgeneral\u002Fmanus",[1224,2750,2751,2752],"windows","macos","linux",[2754,2757,2761,2764],{"plan":1364,"price":325,"limit":2755,"cn_pay":2370,"note":2756},"每日有限 credits，足够每天 1 个 demanding 任务","试水\u002F评估",{"plan":1376,"price":1597,"limit":2758,"cn_pay":2759,"note":2760},"更高 credit + 更长任务时长 + 多模型路由","⚠️ 需海外卡","个人主力档",{"plan":1370,"price":2175,"limit":2762,"cn_pay":2431,"note":2763},"更大 credit 池 + 优先队列","10+ 任务\u002F周",{"plan":2765,"price":1615,"limit":2766,"cn_pay":2431,"note":2767},"Pro+\u002FTeam","最大 credit + 团队工作空间（功能受限）","团队功能仍有限","Free \u002F Pro $20\u002Fmo \u002F Plus $50\u002Fmo \u002F Pro+ Team 最高 $200\u002Fmo","2026-06-18",[2771],"onboarding\u002Fmanus-getting-started",[2773],"manus-deep-review",{"power":799,"ux":799,"price":521,"cn_support":792,"stability":521},{"title":60,"description":2735},[2777,2779,2782,2784,2786,2789],{"title":2778,"url":2224},"Manus 官网",{"title":2780,"url":2781},"Manus 维基百科","https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FManus_(AI_agent)",{"title":2783,"url":1967},"Manus AI Review 2026 (Pick Right)",{"title":2785,"url":1991},"Info-Tech Manus Assessment",{"title":2787,"url":2788},"WEF Butterfly Effect Page","https:\u002F\u002Fwww.weforum.org\u002Forganizations\u002Fbutterfly-effect",{"title":2790,"url":2791},"HyzenPro 2026 Review","https:\u002F\u002Fhyzenpro.com\u002Fblog\u002Fmanus-ai-review","tools\u002Fagent\u002Fgeneral\u002Fmanus",[2794,2795,2796,2797,2798],"需要委托多步骤研究 \u002F 报告（带引用密度的 deep research）","数据分析（CSV \u002F Excel 重的任务）","需要","调研类工作（市场调研、竞品对比、文献综述）","对 Computer Use 类 Agent 实践有兴趣的人","通用 Agent 体感天花板，自主完成复杂多步骤任务",[2801,2802,2803,1833,2804,2805],"general-agent","autonomous","computer-use","butterfly-effect","monica","通用 Agent 体感天花板代表。多 sub-agent 并行 + 自动模型路由让 research \u002F 数据分析输出明显比 ChatGPT Deep Research 引用密度更高。慢、credit 烧得快、Web App Builder 仍有 bug，但任务跑通时确实让人惊艳。","AaSPUCWb2nxYWHxSC_BBzBy3OJG69bjyadYk9bISuwc",1783173060685]