[{"data":1,"prerenderedAt":1302},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"compare-open-interpreter-vs-openmanus":8,"compare-a-open-interpreter":9,"compare-b-openmanus":564},{"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":527,"chinese_friendly":516,"cover":528,"description":529,"domestic":530,"extension":531,"faq":8,"free":530,"github":508,"languages":532,"lastVerified":534,"meta":535,"models":8,"navigation":536,"notSuitable":8,"opensource":536,"path":537,"pillar":538,"platforms":539,"priceTable":8,"pricing":543,"published":544,"relatedPlaybooks":8,"relatedReviews":8,"score":545,"self_host":530,"seo":548,"seoTitle":549,"slug":550,"sources":551,"stem":554,"suitable":8,"tagline":555,"tags":556,"updated":534,"verdict":562,"website":500,"__hash__":563},"tools\u002Ftools\u002Fagent\u002Fdesktop\u002Fopen-interpreter.md","Open Interpreter",[13,14,15],"agent\u002Fdesktop\u002Fclaude-desktop","agent\u002Fgeneral\u002Fopenmanus","agent\u002Fdesktop\u002Fopenclaw",{"type":17,"value":18,"toc":511},"minimark",[19,24,33,36,39,96,99,102,106,116,121,144,149,169,172,210,213,348,351,402,406,435,439,445,451,461,467,470,486,489,494],[20,21,23],"h2",{"id":22},"tldr","TL;DR",[25,26,27,28,32],"p",{},"Open Interpreter 是一个开源的本地代码执行 Agent（AGPL 协议），让 LLM 直接在你的终端里写代码、跑代码、读写文件、执行系统命令、操控浏览器。它不像 ChatGPT 那样在沙箱里跑代码——它在你真实的机器上执行，有完整的系统权限。支持 GPT-4 \u002F Claude \u002F 本地模型（Ollama \u002F LM Studio），Python 实现，",[29,30,31],"code",{},"pip install"," 即用。",[25,34,35],{},"适合：技术用户做本地自动化（批量文件处理、数据分析、系统管理）、探索性编程（让 AI 跑代码看结果再迭代）、原型开发。不适合：生产环境（稳定性不够）、非技术用户（安全风险高）、敏感数据环境（LLM 能读你所有文件）、商业闭源项目（AGPL 传染性协议）。",[20,37,38],{"id":38},"核心能力",[40,41,42,50,56,62,68,74,80,90],"ul",{},[43,44,45,49],"li",{},[46,47,48],"strong",{},"本地代码执行","：LLM 生成 Python \u002F JS \u002F Shell 代码，直接在你的机器上跑，看到真实输出",[43,51,52,55],{},[46,53,54],{},"文件系统操作","：读写文件、创建目录、搜索文件内容、批量重命名，完整的文件系统访问",[43,57,58,61],{},[46,59,60],{},"系统命令执行","：跑任意 shell 命令——安装软件、配置环境、管理进程、操作系统",[43,63,64,67],{},[46,65,66],{},"浏览器控制","：通过 Playwright \u002F Selenium 操控浏览器，自动化网页操作、数据抓取",[43,69,70,73],{},[46,71,72],{},"多模型支持","：OpenAI \u002F Anthropic \u002F Google \u002F Ollama \u002F LM Studio \u002F 任何 OpenAI 兼容端点",[43,75,76,79],{},[46,77,78],{},"交互式 REPL","：终端对话式交互，自然语言描述任务，Agent 自主分解 + 执行 + 反馈",[43,81,82,85,86,89],{},[46,83,84],{},"执行确认机制","：每步代码执行前可选确认（",[29,87,88],{},"--auto_run"," 跳过），防止误操作",[43,91,92,95],{},[46,93,94],{},"多语言执行","：不限于 Python——JS、Shell、R、Ruby 等按需切换",[20,97,98],{"id":98},"价格",[25,100,101],{},"完全免费，AGPL-3.0 开源。成本在于 LLM API 调用费（用 GPT-4 \u002F Claude）或本地模型硬件成本（用 Ollama 免费）。",[20,103,105],{"id":104},"体验与评测资料整理","体验与评测（资料整理）",[107,108,109],"blockquote",{},[25,110,111,112,115],{},"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",[46,113,114],{},"典型场景","：自动化文件处理、数据分析、本地脚本执行等任务（配云 API 或本地 Ollama 均可，硬件门槛低）。",[25,117,118],{},[46,119,120],{},"亮点：",[40,122,123,129,132,135,138,141],{},[43,124,125,128],{},[29,126,127],{},"interpreter"," 一行启动，配 API key 即用，对话式交互体验流畅",[43,130,131],{},"让它「把这个文件夹所有 CSV 合并去重，按日期排序，生成图表」——写 Python → 跑 → 看报错 → 改 → 再跑，全自主完成",[43,133,134],{},"文件操作强——搜索、批量重命名、格式转换，比手写脚本快",[43,136,137],{},"数据分析场景好用——让它读 Excel、清洗数据、画 matplotlib 图，迭代式探索很方便",[43,139,140],{},"配 Ollama 本地模型完全离线运行，隐私敏感任务不碰云",[43,142,143],{},"浏览器自动化——让它打开网页、填表、截图，比手动写 Playwright 脚本快",[25,145,146],{},[46,147,148],{},"踩坑：",[40,150,151,154,157,160,163,166],{},[43,152,153],{},"安全风险是最大问题——LLM 能跑任意命令，一次让它「清理临时文件」差点删了项目目录，幸好有确认机制拦住",[43,155,156],{},"稳定性一般——复杂任务偶尔死循环（反复跑失败代码）、MCP 连接断开、进程卡死",[43,158,159],{},"长任务上下文丢失——对话太长后 Agent 忘记前面的指令，需要手动提醒",[43,161,162],{},"本地模型效果有限——Qwen3 7B 能做简单任务，复杂逻辑推理要上 Claude \u002F GPT-4",[43,164,165],{},"AGPL 协议限制商业使用——如果你在公司用且分发，要注意开源传染性问题",[43,167,168],{},"Windows 路径问题——部分 Shell 命令在 Windows 有兼容性 bug，Mac \u002F Linux 更稳",[20,170,171],{"id":171},"上手",[173,174,175,181,192,198,201,207],"ol",{},[43,176,177,178],{},"安装：",[29,179,180],{},"pip install open-interpreter",[43,182,183,184,187,188,191],{},"配模型：",[29,185,186],{},"interpreter --model claude-3-5-sonnet","（或 ",[29,189,190],{},"--model ollama\u002Fqwen3-coder:7b"," 用本地）",[43,193,194,195,197],{},"启动：",[29,196,127],{}," 进入交互式 REPL",[43,199,200],{},"描述任务：「把 ~\u002FDownloads 所有图片转成 webp，压缩到 80% 质量」",[43,202,203,204,206],{},"确认执行——每步代码执行前会提示确认（开发阶段别开 ",[29,205,88],{},"）",[43,208,209],{},"进阶：配自定义工具 \u002F MCP server 扩展 Agent 能力",[20,211,212],{"id":212},"对比",[214,215,216,237],"table",{},[217,218,219],"thead",{},[220,221,222,226,228,231,234],"tr",{},[223,224,225],"th",{},"维度",[223,227,11],{},[223,229,230],{},"Claude Desktop",[223,232,233],{},"OpenManus",[223,235,236],{},"Devika",[238,239,240,257,274,289,303,317,333],"tbody",{},[220,241,242,246,249,252,255],{},[243,244,245],"td",{},"形态",[243,247,248],{},"CLI",[243,250,251],{},"Desktop App",[243,253,254],{},"Web",[243,256,254],{},[220,258,259,262,265,268,271],{},[243,260,261],{},"开源",[243,263,264],{},"✅ AGPL",[243,266,267],{},"❌",[243,269,270],{},"✅ Apache 2.0",[243,272,273],{},"✅ MIT",[220,275,276,279,282,285,287],{},[243,277,278],{},"代码执行",[243,280,281],{},"✅ 本地真实",[243,283,284],{},"✅ 沙箱",[243,286,284],{},[243,288,284],{},[220,290,291,293,296,299,301],{},[243,292,66],{},[243,294,295],{},"✅",[243,297,298],{},"✅ Computer Use",[243,300,295],{},[243,302,295],{},[220,304,305,308,311,313,315],{},[243,306,307],{},"本地模型",[243,309,310],{},"✅ Ollama",[243,312,267],{},[243,314,295],{},[243,316,295],{},[220,318,319,322,325,328,331],{},[243,320,321],{},"安全性",[243,323,324],{},"⚠️ 低（真实执行）",[243,326,327],{},"✅ 高（沙箱）",[243,329,330],{},"✅ 高",[243,332,330],{},[220,334,335,338,341,344,346],{},[243,336,337],{},"稳定性",[243,339,340],{},"中",[243,342,343],{},"高",[243,345,340],{},[243,347,340],{},[20,349,350],{"id":350},"避坑",[40,352,353,362,372,378,384,390,396],{},[43,354,355,358,359,361],{},[46,356,357],{},"一定要开确认机制","：默认每步执行前确认，别图省事开 ",[29,360,88],{},"，LLM 误删文件的代价远大于点几下确认",[43,363,364,367,368,371],{},[46,365,366],{},"别在生产服务器跑","：Open Interpreter 有完整系统权限，LLM 幻觉可能跑出 ",[29,369,370],{},"rm -rf"," \u002F 覆盖配置文件等灾难操作",[43,373,374,377],{},[46,375,376],{},"敏感数据隔离","：LLM 能读你所有文件，处理含密钥 \u002F 密码 \u002F 用户数据的环境要慎重，建议在 Docker 容器里跑",[43,379,380,383],{},[46,381,382],{},"长任务分段","：对话超过 20-30 轮后上下文退化，复杂任务拆成小段，每段确认结果再继续",[43,385,386,389],{},[46,387,388],{},"本地模型别期望太高","：7B\u002F14B 模型能做文件操作和简单脚本，复杂逻辑推理和调试要上 Claude \u002F GPT-4",[43,391,392,395],{},[46,393,394],{},"AGPL 协议注意","：商业项目分发时 AGPL 有传染性，内部使用没问题，分发 \u002F SaaS 化要咨询法务",[43,397,398,401],{},[46,399,400],{},"Windows 用 WSL","：原生 Windows 有路径 \u002F 权限 \u002F Shell 兼容问题，WSL 下体验一致",[20,403,405],{"id":404},"适合-不适合","适合 \u002F 不适合",[40,407,408,411,414,417,420,423,426,429,432],{},[43,409,410],{},"✅ 技术用户做本地自动化（文件批处理、数据分析、系统管理）",[43,412,413],{},"✅ 探索性编程（让 AI 跑代码看结果再迭代）",[43,415,416],{},"✅ 原型开发 \u002F 快速脚本生成",[43,418,419],{},"✅ 想用本地模型（Ollama）做隐私敏感的自动化",[43,421,422],{},"✅ 学习 \u002F 研究本地 Agent 架构",[43,424,425],{},"❌ 生产环境（稳定性不够 + 安全风险高）",[43,427,428],{},"❌ 非技术用户（误操作风险大，确认机制需要判断力）",[43,430,431],{},"❌ 敏感数据环境（LLM 能访问所有文件）",[43,433,434],{},"❌ 商业闭源项目分发（AGPL 传染性协议）",[20,436,438],{"id":437},"faq","FAQ",[25,440,441,444],{},[46,442,443],{},"Q: Open Interpreter 安全吗？","\nA: 本身不安全——它在你真实机器上执行任意代码，LLM 幻觉可能导致误删文件 \u002F 执行危险命令。缓解措施：开确认机制、在 Docker 容器跑、别给 root 权限、敏感数据隔离。生产环境不推荐。",[25,446,447,450],{},[46,448,449],{},"Q: 和 Claude Desktop 的 Computer Use 区别？","\nA: Claude Desktop 在沙箱虚拟机里操作，安全隔离好。Open Interpreter 在你真实机器上执行，能力更强但风险更高。追求安全选 Claude Desktop，追求能力和免费选 Open Interpreter。",[25,452,453,456,457,460],{},[46,454,455],{},"Q: 能用本地模型吗？","\nA: 可以。",[29,458,459],{},"interpreter --model ollama\u002Fqwen3-coder:7b"," 即用 Ollama 本地模型，完全不碰云 API。简单任务（文件操作、基础脚本）够用，复杂推理和调试建议上 Claude \u002F GPT-4。",[25,462,463,466],{},[46,464,465],{},"Q: 和 AutoGPT \u002F BabyAGI 区别？","\nA: AutoGPT \u002F BabyAGI 偏「自主规划 + 工具调用」的通用 Agent，工具链固定。Open Interpreter 偏「代码执行」——它通过写代码来完成任务，灵活性更高（任何能写代码实现的都能做），但安全风险也更高。",[20,468,469],{"id":469},"相关阅读",[25,471,472,477,478,477,482],{},[473,474,476],"a",{"href":475},"\u002Fcoding\u002Flocal\u002Fjan.html","Jan"," · ",[473,479,481],{"href":480},"\u002Fcoding\u002Flocal\u002Fgpt4all.html","GPT4All",[473,483,485],{"href":484},"\u002Fcoding\u002Flocal\u002Fvllm.html","vLLM",[20,487,488],{"id":488},"来源",[107,490,491],{},[25,492,493],{},"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。",[40,495,496,504],{},[43,497,498],{},[473,499,503],{"href":500,"rel":501},"https:\u002F\u002Fopeninterpreter.com",[502],"nofollow","官网",[43,505,506],{},[473,507,510],{"href":508,"rel":509},"https:\u002F\u002Fgithub.com\u002FOpenInterpreter\u002Fopen-interpreter",[502],"GitHub",{"title":512,"searchDepth":513,"depth":513,"links":514},"",3,[515,517,518,519,520,521,522,523,524,525,526],{"id":22,"depth":516,"text":23},2,{"id":38,"depth":516,"text":38},{"id":98,"depth":516,"text":98},{"id":104,"depth":516,"text":105},{"id":171,"depth":516,"text":171},{"id":212,"depth":516,"text":212},{"id":350,"depth":516,"text":350},{"id":404,"depth":516,"text":405},{"id":437,"depth":516,"text":438},{"id":469,"depth":516,"text":469},{"id":488,"depth":516,"text":488},"desktop","\u002Fimg\u002Ftools\u002Fopen-interpreter.webp","Open Interpreter 真实评测：开源本地代码执行 Agent（AGPL 协议），让 LLM 在终端直接跑 Python\u002FJS\u002FShell 代码，操控浏览器、读写文件、执行系统命令。支持 GPT-4\u002FClaude\u002F本地模型，适合需要 LLM 实时操作本地环境的开发者。",false,"md",[533],"en","2026-07-30",{},true,"\u002Ftools\u002Fagent\u002Fdesktop\u002Fopen-interpreter","agent",[540,541,542],"windows","macos","linux","Free \u002F 开源（AGPL）","2026-07-05",{"power":546,"ux":513,"price":547,"cn_support":516,"stability":513},4,5,{"title":11,"description":529},"Open Interpreter - 本地代码执行 Agent 评测 | AIHO","agent\u002Fdesktop\u002Fopen-interpreter",[552,553],{"title":503,"url":500},{"title":510,"url":508},"tools\u002Fagent\u002Fdesktop\u002Fopen-interpreter","本地代码执行 Agent，让 LLM 在你的机器上跑代码",[557,558,559,560,561],"desktop-agent","opensource","code-execution","local","python","开创性的本地代码执行 Agent，让 LLM 直接操控你的电脑跑代码、读写文件、操作浏览器，适合技术用户做自动化和探索；安全风险高、稳定性一般，不适合生产环境或非技术用户。","L4JM70girxbUKgBD0dOIoVvQD-2eCVHnpHH7ZtdCUYo",{"id":565,"title":233,"alternatives":566,"api_compatible":8,"body":570,"category":1244,"chinese_friendly":546,"cover":1245,"description":1246,"domestic":530,"extension":531,"faq":1247,"free":530,"github":8,"languages":1260,"lastVerified":8,"meta":1263,"models":8,"navigation":536,"notSuitable":8,"opensource":536,"path":1264,"pillar":538,"platforms":1265,"priceTable":1267,"pricing":1272,"published":1273,"relatedPlaybooks":1274,"relatedReviews":1276,"score":1279,"self_host":536,"seo":1280,"seoTitle":1281,"slug":14,"sources":1282,"stem":1292,"suitable":8,"tagline":1293,"tags":1294,"updated":1285,"verdict":1300,"website":1205,"__hash__":1301},"tools\u002Ftools\u002Fagent\u002Fgeneral\u002Fopenmanus.md",[567,568,569],"agent\u002Fplatform\u002Flangflow","agent\u002Fplatform\u002Fn8n","agent\u002Fprotocol\u002Fcomposio",{"type":17,"value":571,"toc":1232},[572,574,581,584,586,663,665,682,686,690,716,720,746,748,901,904,912,914,1084,1086,1146,1148,1174,1176,1196,1198,1228],[20,573,23],{"id":22},[25,575,576,577,580],{},"OpenManus 是 MetaGPT 核心团队 2025-03 推出的开源 Manus 替代方案，52,000+ GitHub stars。差异点：MIT 协议 + Python 模块化架构 + 多 agent orchestration（",[29,578,579],{},"run_flow.py","）+ Playwright 浏览器自动化 + MCP 工具协议支持 + DataAnalysis 内置模式 + OpenManus-RL 强化学习分支 + 自定义工具基类（BaseTool）+ 多模型（GPT-4o \u002F Claude \u002F Qwen VL Plus）。零邀请码、零订阅、零供应商绑定。",[25,582,583],{},"适合：想自托管复刻 Manus 体验的开发者；研究 \u002F 学术 \u002F 教育用通用 agent 实现学习；隐私敏感 + 不愿数据上 Manus 商业云；预算紧（只付 LLM API）；中国大陆开发者（搭配 Qwen \u002F DeepSeek 本地化）。不适合：非开发者 \u002F 不会折腾 Python + Playwright；要 GUI \u002F 上手即用；生产级稳定（项目演进快，文档滞后）。",[20,585,38],{"id":38},[40,587,588,597,603,609,615,621,627,633,639,645,651,657],{},[43,589,590,593,594,596],{},[46,591,592],{},"多 agent orchestration","：",[29,595,579],{}," 编排多个专门 agent 协作",[43,598,599,602],{},[46,600,601],{},"Playwright 浏览器自动化","：截图 + DOM 操作 + 表单填写 + 信息抓取",[43,604,605,608],{},[46,606,607],{},"MCP 协议支持","：可调用 MCP server（filesystem \u002F GitHub \u002F Postgres）",[43,610,611,614],{},[46,612,613],{},"DataAnalysis 模式","：内置 CSV \u002F 数据分析 agent",[43,616,617,620],{},[46,618,619],{},"OpenManus-RL","：强化学习微调分支",[43,622,623,626],{},[46,624,625],{},"BaseTool 自定义工具","：Python 继承基类快速添加新工具",[43,628,629,632],{},[46,630,631],{},"多模态","：文本 + 视觉输入 + 浏览器截图回环",[43,634,635,638],{},[46,636,637],{},"多 LLM provider","：GPT-4o \u002F Claude 3.5 \u002F Qwen VL Plus \u002F DeepSeek \u002F Gemini",[43,640,641,644],{},[46,642,643],{},"核心 agent 引擎","：reasoning + planning + execution 三阶段",[43,646,647,650],{},[46,648,649],{},"Web UI 监控","：实时看 AI thinking process",[43,652,653,656],{},[46,654,655],{},"任务可视化","：步骤拆解 + 执行树",[43,658,659,662],{},[46,660,661],{},"MIT 协议","：个人 + 商用全免费",[20,664,98],{"id":98},[40,666,667,673,676,679],{},[43,668,669,672],{},[46,670,671],{},"Free \u002F OSS","：$0；MIT 协议",[43,674,675],{},"真实成本 = LLM API（GPT-4o ~$5\u002FM input + $15\u002FM output \u002F Claude \u002F Qwen 等）",[43,677,678],{},"一次中等任务（10-20 步）API 费用 $0.05-0.5",[43,680,681],{},"本地 Qwen2.5 32B \u002F DeepSeek 走 vLLM \u002F Ollama 路径 $0",[20,683,685],{"id":684},"实测开发者-自托管-研究场景","实测（开发者 \u002F 自托管 \u002F 研究场景）",[25,687,688],{},[46,689,120],{},[40,691,692,695,698,701,704,707,710,713],{},[43,693,694],{},"52k stars 印证社区认同度 + 活跃度",[43,696,697],{},"MetaGPT 团队背景保证架构质量",[43,699,700],{},"多 agent 协作场景比单 agent 实现稳得多",[43,702,703],{},"Playwright 浏览器自动化非常完整",[43,705,706],{},"MCP 协议接入打通 Claude \u002F Cursor 生态",[43,708,709],{},"DataAnalysis 内置 agent 模式开箱即用",[43,711,712],{},"中文支持自然（Qwen VL Plus 接入）",[43,714,715],{},"自托管 + 数据本地，隐私 \u002F 合规友好",[25,717,718],{},[46,719,148],{},[40,721,722,725,728,731,734,737,740,743],{},[43,723,724],{},"项目演进快，breaking change 偶发（pin commit 跑生产）",[43,726,727],{},"文档滞后新功能 1-2 个月",[43,729,730],{},"无官方 GUI，监控 UI 在做但不完整",[43,732,733],{},"需要 Python 3.12+ + Playwright 依赖（首次安装 chromium 慢）",[43,735,736],{},"LLM API 配置非平凡（多 provider \u002F key \u002F 模型选择）",[43,738,739],{},"浏览器任务遇到 CAPTCHA \u002F 反爬偶尔卡死",[43,741,742],{},"中文 prompt 效果依赖底层模型",[43,744,745],{},"生产部署需自己加监控 \u002F 错误恢复 \u002F 重试",[20,747,171],{"id":171},[749,750,754],"pre",{"className":751,"code":752,"language":753,"meta":512,"style":512},"language-bash shiki shiki-themes github-light github-dark","git clone https:\u002F\u002Fgithub.com\u002FFoundationAgents\u002FOpenManus.git\ncd OpenManus\npython3.12 -m venv .venv && source .venv\u002Fbin\u002Factivate  # 或 .venv\\Scripts\\activate\npip install -r requirements.txt\nplaywright install chromium\n\n# 配 LLM API\ncp config\u002Fconfig.example.toml config\u002Fconfig.toml\n# 编辑 config.toml 填 OPENAI \u002F ANTHROPIC \u002F DEEPSEEK key\n\n# 单 agent 模式\npython main.py\n\n# 多 agent 模式\npython run_flow.py\n","bash",[29,755,756,772,781,809,823,833,839,845,857,863,868,874,882,887,893],{"__ignoreMap":512},[757,758,761,765,769],"span",{"class":759,"line":760},"line",1,[757,762,764],{"class":763},"sScJk","git",[757,766,768],{"class":767},"sZZnC"," clone",[757,770,771],{"class":767}," https:\u002F\u002Fgithub.com\u002FFoundationAgents\u002FOpenManus.git\n",[757,773,774,778],{"class":759,"line":516},[757,775,777],{"class":776},"sj4cs","cd",[757,779,780],{"class":767}," OpenManus\n",[757,782,783,786,789,792,795,799,802,805],{"class":759,"line":513},[757,784,785],{"class":763},"python3.12",[757,787,788],{"class":776}," -m",[757,790,791],{"class":767}," venv",[757,793,794],{"class":767}," .venv",[757,796,798],{"class":797},"sVt8B"," && ",[757,800,801],{"class":776},"source",[757,803,804],{"class":767}," .venv\u002Fbin\u002Factivate",[757,806,808],{"class":807},"sJ8bj","  # 或 .venv\\Scripts\\activate\n",[757,810,811,814,817,820],{"class":759,"line":546},[757,812,813],{"class":763},"pip",[757,815,816],{"class":767}," install",[757,818,819],{"class":776}," -r",[757,821,822],{"class":767}," requirements.txt\n",[757,824,825,828,830],{"class":759,"line":547},[757,826,827],{"class":763},"playwright",[757,829,816],{"class":767},[757,831,832],{"class":767}," chromium\n",[757,834,836],{"class":759,"line":835},6,[757,837,838],{"emptyLinePlaceholder":536},"\n",[757,840,842],{"class":759,"line":841},7,[757,843,844],{"class":807},"# 配 LLM API\n",[757,846,848,851,854],{"class":759,"line":847},8,[757,849,850],{"class":763},"cp",[757,852,853],{"class":767}," config\u002Fconfig.example.toml",[757,855,856],{"class":767}," config\u002Fconfig.toml\n",[757,858,860],{"class":759,"line":859},9,[757,861,862],{"class":807},"# 编辑 config.toml 填 OPENAI \u002F ANTHROPIC \u002F DEEPSEEK key\n",[757,864,866],{"class":759,"line":865},10,[757,867,838],{"emptyLinePlaceholder":536},[757,869,871],{"class":759,"line":870},11,[757,872,873],{"class":807},"# 单 agent 模式\n",[757,875,877,879],{"class":759,"line":876},12,[757,878,561],{"class":763},[757,880,881],{"class":767}," main.py\n",[757,883,885],{"class":759,"line":884},13,[757,886,838],{"emptyLinePlaceholder":536},[757,888,890],{"class":759,"line":889},14,[757,891,892],{"class":807},"# 多 agent 模式\n",[757,894,896,898],{"class":759,"line":895},15,[757,897,561],{"class":763},[757,899,900],{"class":767}," run_flow.py\n",[25,902,903],{},"试任务示例：",[749,905,910],{"className":906,"code":908,"language":909},[907],"language-text","> 帮我做一份『2026 开源 AI Agent 框架』竞品对比，含表格 + 引用 + 趋势分析，输出为 markdown 文件\n","text",[29,911,908],{"__ignoreMap":512},[20,913,212],{"id":212},[214,915,916,933],{},[217,917,918],{},[220,919,920,922,924,927,930],{},[223,921,225],{},[223,923,233],{},[223,925,926],{},"LangChain",[223,928,929],{},"AutoGPT",[223,931,932],{},"CrewAI",[238,934,935,951,967,980,996,1009,1023,1037,1054,1067],{},[220,936,937,939,942,945,948],{},[243,938,245],{},[243,940,941],{},"现成 agent 实现",[243,943,944],{},"building blocks",[243,946,947],{},"早期通用 agent",[243,949,950],{},"多 agent 框架",[220,952,953,956,959,962,965],{},[243,954,955],{},"浏览器自动化",[243,957,958],{},"✅ Playwright",[243,960,961],{},"–",[243,963,964],{},"部分",[243,966,961],{},[220,968,969,972,974,976,978],{},[243,970,971],{},"MCP",[243,973,295],{},[243,975,964],{},[243,977,961],{},[243,979,961],{},[220,981,982,985,988,991,993],{},[243,983,984],{},"多 agent",[243,986,987],{},"✅ orchestration",[243,989,990],{},"需自搭",[243,992,267],{},[243,994,995],{},"✅ 旗舰",[220,997,998,1001,1003,1005,1007],{},[243,999,1000],{},"DataAnalysis 内置",[243,1002,295],{},[243,1004,961],{},[243,1006,961],{},[243,1008,961],{},[220,1010,1011,1014,1017,1019,1021],{},[243,1012,1013],{},"RL 微调",[243,1015,1016],{},"✅ OpenManus-RL",[243,1018,961],{},[243,1020,961],{},[243,1022,961],{},[220,1024,1025,1028,1031,1033,1035],{},[243,1026,1027],{},"协议",[243,1029,1030],{},"MIT",[243,1032,1030],{},[243,1034,1030],{},[243,1036,1030],{},[220,1038,1039,1042,1045,1048,1051],{},[243,1040,1041],{},"Stars",[243,1043,1044],{},"52k+",[243,1046,1047],{},"100k+",[243,1049,1050],{},"170k+",[243,1052,1053],{},"30k+",[220,1055,1056,1058,1060,1063,1065],{},[243,1057,171],{},[243,1059,340],{},[243,1061,1062],{},"难",[243,1064,340],{},[243,1066,340],{},[220,1068,1069,1072,1075,1078,1081],{},[243,1070,1071],{},"适合",[243,1073,1074],{},"自托管 Manus 复刻",[243,1076,1077],{},"底层 framework",[243,1079,1080],{},"学习经典",[243,1082,1083],{},"多 agent 协作",[20,1085,350],{"id":350},[40,1087,1088,1094,1104,1110,1116,1122,1128,1134,1140],{},[43,1089,1090,1093],{},[46,1091,1092],{},"pin commit 用生产","：项目演进快，main 分支偶尔 break",[43,1095,1096,1099,1100,1103],{},[46,1097,1098],{},"Playwright 依赖大","：首次 ",[29,1101,1102],{},"playwright install"," 下 chromium 慢，国内走镜像",[43,1105,1106,1109],{},[46,1107,1108],{},"LLM 选择","：日常用 GPT-4o-mini \u002F DeepSeek 省钱，复杂任务切 GPT-4o \u002F Claude Opus",[43,1111,1112,1115],{},[46,1113,1114],{},"本地化中文","：Qwen2.5 VL 32B + vLLM 部署可全本地 + 零成本",[43,1117,1118,1121],{},[46,1119,1120],{},"监控自加","：生产部署要加 prometheus + 错误重试 + 任务超时",[43,1123,1124,1127],{},[46,1125,1126],{},"浏览器反爬","：CAPTCHA 场景搭配 2captcha \u002F human-in-loop",[43,1129,1130,1133],{},[46,1131,1132],{},"OpenManus-RL 分支","：研究场景才需要，普通用户主仓库就够",[43,1135,1136,1139],{},[46,1137,1138],{},"MCP server","：信任来源很重要，能访问的目录 \u002F 工具要审慎",[43,1141,1142,1145],{},[46,1143,1144],{},"多 agent runaway","：复杂任务设 max_steps 防止失控烧 token",[20,1147,405],{"id":404},[40,1149,1150,1153,1156,1159,1162,1165,1168,1171],{},[43,1151,1152],{},"✅ 开发者 + 想自托管 Manus 风格 agent",[43,1154,1155],{},"✅ 研究 \u002F 学术 \u002F 教育用通用 agent 学习",[43,1157,1158],{},"✅ 隐私敏感 + 不愿数据上商业云",[43,1160,1161],{},"✅ 中国大陆开发者（Qwen \u002F DeepSeek 本地化）",[43,1163,1164],{},"❌ 非开发者 \u002F 不会 Python + Playwright",[43,1166,1167],{},"❌ 要 GUI \u002F 上手即用",[43,1169,1170],{},"❌ 生产级稳定（文档滞后 + 演进快）",[43,1172,1173],{},"❌ 团队协作 + 共享 workspace（用 Flowith \u002F Genspark Team）",[20,1175,469],{"id":469},[40,1177,1178,1184,1190],{},[43,1179,1180],{},[473,1181,1183],{"href":1182},"\u002Ftools\u002Fagent\u002Fplatform\u002Flangflow","Langflow 评测",[43,1185,1186],{},[473,1187,1189],{"href":1188},"\u002Ftools\u002Fagent\u002Fplatform\u002Fn8n","n8n 评测",[43,1191,1192],{},[473,1193,1195],{"href":1194},"\u002Ftools\u002Fagent\u002Fprotocol\u002Fcomposio","Composio 评测",[20,1197,488],{"id":488},[173,1199,1200,1207,1214,1221],{},[43,1201,1202,1203],{},"OpenManus GitHub 主仓库 + Foundation Agents 组织 ",[473,1204,1205],{"href":1205,"rel":1206},"https:\u002F\u002Fgithub.com\u002FFoundationAgents\u002FOpenManus",[502],[43,1208,1209,1210],{},"Foundation Agents — OpenManus 项目介绍 ",[473,1211,1212],{"href":1212,"rel":1213},"https:\u002F\u002Ffoundationagents.org\u002Fprojects\u002Fopenmanus\u002F",[502],[43,1215,1216,1217],{},"Toolsverse — OpenManus 评测 + 52k stars ",[473,1218,1219],{"href":1219,"rel":1220},"https:\u002F\u002Fthetoolsverse.com\u002Ftools\u002Fopenmanus",[502],[43,1222,1223,1224],{},"SoloSoft.dev — OpenManus 2026 Framework 综述 ",[473,1225,1226],{"href":1226,"rel":1227},"https:\u002F\u002Fwww.solosoft.dev\u002Fpost\u002Fopenmanus-agent-framework-2026\u002F",[502],[1229,1230,1231],"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 .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":512,"searchDepth":513,"depth":513,"links":1233},[1234,1235,1236,1237,1238,1239,1240,1241,1242,1243],{"id":22,"depth":516,"text":23},{"id":38,"depth":516,"text":38},{"id":98,"depth":516,"text":98},{"id":684,"depth":516,"text":685},{"id":171,"depth":516,"text":171},{"id":212,"depth":516,"text":212},{"id":350,"depth":516,"text":350},{"id":404,"depth":516,"text":405},{"id":469,"depth":516,"text":469},{"id":488,"depth":516,"text":488},"general","\u002Fimg\u002Ftools\u002Fopenmanus.webp","OpenManus 真实评测：MetaGPT 核心成员 Xinbin Liang \u002F Jinyu Xiang \u002F Zhaoyang Yu \u002F Jiayi Zhang \u002F Sirui Hong 在 2025-03 推出的开源 Manus 替代方案，52,000+ GitHub stars。MIT 协议 + Python 模块化架构 + 多 agent orchestration（run_flow.py）+ Playwright 浏览器自动化 + MCP 工具支持 + DataAnalysis 内置模式 + OpenManus-RL 强化学习微调分支 + 自定义工具基类（BaseTool）。零邀请码、零订阅、零供应商绑定，唯一成本 = LLM API。",[1248,1251,1254,1257],{"q":1249,"a":1250},"OpenManus 和 Manus 是什么关系？","Manus 是商业 \u002F 邀请制的通用 AI agent 产品。OpenManus 是 MetaGPT 核心团队 2025-03 推出的开源复刻版，目标是『让所有人不靠邀请码就能用上类 Manus 能力』。功能覆盖：研究 \u002F 浏览器 \u002F 数据分析 \u002F 文件操作 \u002F 多步 reasoning。不是 Manus 官方出品。",{"q":1252,"a":1253},"和 LangChain \u002F AutoGPT \u002F CrewAI 怎么定位？","OpenManus 不是 framework，更像『可直接跑的通用 agent 实现』。LangChain 是 building block 框架；AutoGPT 是早期通用 agent；CrewAI 是多 agent 协作 framework。要『拉下来配 API 就能跑 Manus 风格任务』→ OpenManus；要『从底层搭自己的 agent』→ LangChain \u002F CrewAI；要『历史经典 + 学习』→ AutoGPT。OpenManus 内部用 LangChain-like 模块，可视为『现成实现』。",{"q":1255,"a":1256},"OpenManus-RL 是什么？","OpenManus 项目下的强化学习分支，提供 RL-based 微调方法优化 agent 性能。对研究 \u002F 高定制场景有价值，普通用户主仓库已经够用。",{"q":1258,"a":1259},"上手门槛？","需要 Python 3.12+ + 熟悉终端 + 自配 LLM API。无 GUI（虽然 web 监控界面在做）。documentation 偶尔滞后。非开发者建议先试 GUI 工具（Flowith \u002F Genspark），开发者 \u002F 研究者 + 想自托管 + 隐私敏感 → OpenManus。",[533,1261,1262],"zh","multi",{},"\u002Ftools\u002Fagent\u002Fgeneral\u002Fopenmanus",[542,541,540,1266],"docker",[1268],{"plan":671,"price":1269,"features":1270,"notes":1271},"$0","MIT 协议 + 全部功能 + 自托管 + 多 agent + 浏览器 + MCP + DataAnalysis + OpenManus-RL","LLM API 自付","MIT 完全免费开源 \u002F 用户自付 LLM API（GPT-4o \u002F Claude 3.5 \u002F Qwen VL Plus 任选）","2026-06-19",[1275],"onboarding\u002Fopen-source-general-agent",[1277,1278],"openhuman-deep-review","openmanus-deep-review",{"power":547,"ux":513,"price":547,"cn_support":546,"stability":513},{"title":233,"description":1246},"OpenManus 评测 2026：MetaGPT 开源 Manus 替代，52k+ Stars",[1283,1286,1288,1290],{"name":1284,"url":1205,"accessed":1285},"OpenManus GitHub（FoundationAgents 组织）","2026-06-24",{"name":1287,"url":1212,"accessed":1285},"Foundation Agents — OpenManus 项目介绍",{"name":1289,"url":1219,"accessed":1285},"Toolsverse — OpenManus 评测 + 52k stars",{"name":1291,"url":1226,"accessed":1285},"SoloSoft.dev — OpenManus 2026 Framework 综述","tools\u002Fagent\u002Fgeneral\u002Fopenmanus","MetaGPT 团队开源版 Manus——52k+ stars \u002F MIT \u002F 多 agent + 浏览器自动化 + MCP + DataAnalysis",[558,1295,1296,1297,1298,1299],"multi-agent","browser-automation","mcp","metagpt","openmanus","想自托管复刻 Manus 全能 agent 体验 + 不愿等邀请码的开发者首选——浏览器 + 数据分析 + MCP 工具栈一站全。要 GUI \u002F 上手即用 \u002F 生产级稳定建议 Genspark \u002F Flowith 付费版。","27xZ_sl8oK4xtfoZj-H7oK3WtsdrfyluCPcf_wJzlRM",1785428440448]