和 Meta Muse、OpenAI dots 比
Muse 和 dots 是厂商替你托管的助理：agent 跑在厂商的云电脑上，用厂商自己的模型。Agent J 不托管 agent。它让你在手机上用你自己的 agent，那个 agent 跑在你自己的电脑或你自己的云服务器（AWS / VPS）上，数据都在你自己的机器上。
先说它们强在哪。上手不用碰终端，装个 App 就能用。agent 在厂商的云电脑上全天在线，你的电脑关了也照样干活。现成的连接器多，语音和入口也比我们多。要办个人杂事、又不想管电脑，选它们更省事。
我们适合另一种人：已经在用 Claude Code、Codex 或 OpenCode，想让 agent 在自己的机器上干活，在手机上点头批准，在一个后台管所有席位。
怎么读这张表。对手两列截至 2026-10-02，Agent J 一列更新于 2026-10-03。对手的每条都标了出处编号，对应页面底部的官方页面；官方没写的，我们写「未公布」。第三方评测和媒体转述的说法，我们没有采用。Agent J 一列只写 Beta 今天已经上线的东西，还没做的标「路线图」。我们自己的安全说法各做到哪一步，以数据安全页为准。
逐项对比
手机上可以左右滑动看全表。
| 维度 | Agent J（Beta） | Meta Muse | OpenAI dots |
|---|---|---|---|
| 是什么 | 手机遥控你自己的 agent，加上管席位的账号后台 | Meta 托管的个人 AI agent M1 | 放在 ChatGPT 里、全天在线的个人 agent D1 |
| agent 在哪里跑 | 你自己的电脑，或你自己的云服务器（AWS / VPS）。Linux、macOS；Windows 用 WSL2。两种都支持，数据都在你自己的机器上 | Meta 云上每人一台专属电脑 M1；Mac 版还能操作你电脑上的文件和 App M5 | 每个 dot 一台 OpenAI 的云电脑；连你自己的电脑是可选的，默认关 D2 |
| 模型 | 你自己登录的 Claude Code、Codex 或 OpenCode（OpenCode 可以配智谱、DeepSeek 等国产模型），费用在你和模型厂商之间；换 agent 不用换产品 | Muse Spark M1 | GPT-6 Astra D1 |
| 你的设备关机后 | 放在家里或办公室的电脑关了，agent 就停。放在云服务器上，它就和服务器一样全天在线 | 照常干活，它在 Meta 的云电脑上 M1 | 照常干活，全天候 D1 |
| 谁看得到内容 | 你和 agent 的聊天是加密的，我们看不到；你选的模型厂商看得到它处理的那部分。细节见数据安全页 | 内容在 Meta 的云上；对话和云电脑里的数据不给 Meta 广告系统。整台云电脑用只有你持有的密钥加密的 Confidential VM 已经宣布，截至 2026-10-02 还没上线 M1 | 内容在 OpenAI 那里，存储和传输时加密；即使关了模型改进，在安全事件等有限情况下仍可能有人工查看 D3 |
| 拿你的数据训练模型吗 | 我们拿不到对话内容，没法用来训练；模型厂商怎么用，看你在它那边的账号设置 | 可以选择退出 M1 | 个人方案默认会用，可在设置里关掉；Business / Enterprise / Edu 默认不用 D3 D6 |
| 审批 | 需要你点头的环节，它会在手机上弹窗问你，每次的决定都记在你的电脑上。花钱、删除、对外发送、改密码、改价格这五类，每次都要你在手机上单独批准，你自己的设置也放宽不了 | 发邮件、购物等敏感动作先问你；同一台云电脑上另有一个 Sentinel agent，任何动作出网都要它放行；操作记录完整可查 M1 M5 | 自带默认规则，可用 Custom Rules 调整。改密码、转账要你自己接手；永久删除数据、装软件可能每次都要批准。独立的 Auto-review 在动作执行前检查，Custom Rules 关不掉它 D3 |
| 记忆 | 用 Claude Code / Codex / OpenCode 自己的记忆，就是你电脑上的文件。手机上的「记忆」页能逐条看、逐条删（7 天内可以撤销）；「记录」页列出每轮对话、每个请求和你的决定，只存在你的电脑上 | 会记住要紧的事，你可以让它忘掉某件具体的事 M1 | 不能逐条查看或删除 dot 自己的记忆，只能删掉整个 dot 重来 D3 |
| 暂停 | 手机上一键「全部停下」：中断正在做的事、拒绝所有等批准的请求、暂停全部定时任务；电脑上也可以运行 agentj stop | 未公布 | 可以在 dot 的资料页暂停；监控发现安全问题时，系统也会暂停或停止它 D1 D2 |
| 连接器 | 你的 agent 本来能用的 MCP 和命令行工具都能用，我们不拦；我们自己没有连接器目录 | 首发几十家合作方加 Shopify 商品目录，之后陆续增加；支持自定义连接器 M2 M3 | 经插件生态连接 4,000 多个 App D1 |
| 入口 | 手机网页版，可以添加到主屏幕，锁屏推送不带内容。原生 App：路线图 | iOS、Android、网页、Mac、WhatsApp M1 M5；AI 眼镜未来几个月上线 M3 | ChatGPT 的桌面、网页和手机端，Slack、Teams；短信只对美国部分 Pro 用户小范围测试；创建 dot 只能在桌面端 D1 D2 |
| 语音 | 目前只有文字 | 有语音模式，边聊边在后台干活 M3 | 可以给它打语音电话；发布时它不能主动打给你 D1 D2 |
| 工作流 | 在手机上说「帮我设计工作流」，你电脑上的 agent 会一题一题问你，把工作流写进它的工作文件夹，问答不发给我们。五份跨境电商起步模板（经营日报、广告复盘、Listing、客服回复草稿、选品）装好是关着的，要不要按时运行由你打开 | 给它一个目标，它自己拆步骤；小企业版带一批经营用的技能和连接器 M2 | 定时任务和 Custom Rules；给企业的 specialist dots 还在试点 D1 D2 |
| 团队 | 一个账号、按席位付费、可以把席位交给同事；新手机只能在那台电脑上批准 | 有小企业版 M2；Meta Enterprise Platform 已宣布，团队管理功能未公布 M6 | Business Premium 可用；Enterprise 由管理员开启 beta D2 |
| 价格 | $20 / 席位 / 月，或 $200 / 席位 / 年；模型费用另付给模型厂商 | 免费（有用量上限）；Power $20 / 月；Maximum $100 / 月 M4 | 第一个 dot 含在 Pro（$100 / $200 / $500 三档）或 Business Premium（$125 / 席位 / 月，年付 $100）里 D2 D4 D5 |
| 地区 | Claude Code / Codex 能用的地方。中国大陆：用 OpenCode 加你自己注册的国产模型；我们的网站和转发服务器在海外，从大陆访问可能慢 | 美国和加拿大 M2 | Pro 不含欧洲经济区、瑞士和英国；Business Premium 覆盖 ChatGPT 支持的所有地区 D2 |
| 更适合谁 | 已经在用 Claude Code / Codex / OpenCode，想把日常工作交给 agent、数据留在自己机器上的个人和团队 | 想零门槛处理个人杂事的美国、加拿大用户 | 已经是 ChatGPT Pro 或企业客户，想要一个全天在后台做研究、办事的助理 |
差别在哪
它们强的地方
- 不用任何电脑或终端，装好就能用。
- 云电脑全天在线，你的设备关机也不耽误。
- 连接器、语音、入口都比我们多得多。
- 审批之外还有独立的检查：Muse 有 Sentinel，dots 有 Auto-review。
我们的不同
- agent 跑在你自己的电脑或云服务器上；你和它的聊天是加密的，我们看不到。
- 不绑模型：用你已经付费的 Claude Code、Codex，或者配国产模型的 OpenCode。
- 适合团队：一人一个席位，可以发给同事，每次批准都有记录。
- 记忆就是你电脑上的文件，可以逐条看、改、删。
Agent J, Meta Muse and OpenAI dots
Muse and dots are assistants the vendor hosts for you: the agent runs on the vendor's cloud computer, on the vendor's own model. Agent J hosts no agent. It puts your own agent in your pocket, and that agent runs on your own computer or your own cloud server (AWS or any VPS). Either way, the data stays on machines you control.
What they do better: no terminal, nothing to set up; the agent stays online in the vendor's cloud when your laptop is closed; far more connectors, voice and places to reach it. For personal errands with zero setup, they are the easier choice. We are for people already using Claude Code, Codex or OpenCode who want an agent working on their own machines, with approvals on the phone and every seat managed in one place.
Muse and dots: as of 2026-10-02 (the Agent J column updated 2026-10-03), each item footnoted to the vendor's own page; where the vendor says nothing, we say "not published". We did not use reviews or press summaries. Agent J: only what is live in the Beta today; the rest is marked "roadmap". How far each of our security claims has got is on the data security page.
| Item | Agent J (Beta) | Meta Muse | OpenAI dots |
|---|---|---|---|
| What it is | A phone remote for your own agent, plus an account dashboard for seats | A personal AI agent hosted by Meta M1 | An always-on personal agent inside ChatGPT D1 |
| Where the agent runs | Your own computer or your own cloud server (AWS / VPS). Linux, macOS; Windows via WSL2. Both are supported; the data stays on machines you control | A dedicated computer per person in Meta's cloud M1; the Mac app can also work with files and apps on your Mac M5 | An OpenAI cloud computer per dot; access to your own computer is optional and off by default D2 |
| Model | The Claude Code, Codex or OpenCode you are signed in to (OpenCode can use Chinese models such as Zhipu or DeepSeek); you pay the model vendor directly | Muse Spark M1 | GPT-6 Astra D1 |
| When your device is off | Turn off the computer at home or in the office and the agent stops. On a cloud server it stays online as long as the server does | Keeps working in Meta's cloud M1 | Keeps working, around the clock D1 |
| Who can see the content | Your chats with the agent are encrypted and we can't see them; the model vendor you chose sees what it processes. Details on the data security page | Content lives in Meta's cloud; conversations and VM data are not shared with Meta's ad systems. Confidential VM (the whole VM encrypted with a key only you hold) is announced, not live as of 2026-10-02 M1 | Content is with OpenAI, encrypted at rest and in transit; human review may occur in limited cases such as safety, even with model improvement off D3 |
| Training on your data | We never receive conversation content, so we cannot train on it; the model vendor follows your account settings | You can opt out M1 | On by default for personal plans, can be turned off; off by default for Business / Enterprise / Edu D3 D6 |
| Approvals | Whenever it needs your OK, it pops up a question on your phone, and every answer is logged on your computer. Five kinds of action (spending, deleting, sending out, changing passwords, changing prices) always need your approval on the phone, one by one; your own settings can't loosen that | Asks before sensitive actions such as sending email or buying; a separate Sentinel agent must approve anything that reaches the internet; full audit trail M1 M5 | Built-in default rules plus Custom Rules. Changing a password or moving money is handed back to you; permanently deleting data or installing software may need approval each time. A separate Auto-review checks actions before they run, and Custom Rules cannot turn it off D3 |
| Memory | Claude Code / Codex / OpenCode memory, stored as plain files on your computer. The Memory page on the phone lists each item and deletes it (undo within 7 days); the Activity page lists every turn, request and decision, kept only on your computer | Remembers what matters; you can tell it to forget specific things M1 | You cannot view or delete individual dot memories; the only way is to delete the dot D3 |
| Pause | One tap on the phone stops everything: ends the current work, denies every pending approval, pauses all scheduled tasks; agentj stop does the same on the computer | Not published | Pause from the dot's profile; safety monitoring can also pause or stop it D1 D2 |
| Connectors | Whatever MCP servers and command-line tools your agent already uses; we don't block them. No connector catalog of our own | Dozens of partners at launch plus the Shopify catalog, more since; custom connectors M2 M3 | Over 4,000 apps through plugins D1 |
| Where you reach it | Phone web app, can be added to the home screen; lock-screen push carries no content. Native apps: roadmap | iOS, Android, web, Mac, WhatsApp M1 M5; AI glasses in the coming months M3 | ChatGPT on desktop, web and mobile; Slack, Teams; texting is a limited beta for some US Pro users; dots are created on desktop only D1 D2 |
| Voice | Text only today | Voice mode; keeps working in the background while you talk M3 | You can call it; at launch it cannot call you D1 D2 |
| Workflows | Say "help me design a workflow" on the phone and the agent on your computer interviews you, then writes the workflows into its working folder; the answers never reach us. Five cross-border e-commerce starter templates (daily report, ads review, listings, draft buyer replies, product research) install switched off, and you decide which run on schedule | Give it a goal and it works out the steps; the small-business edition adds business skills and connectors M2 | Scheduled tasks and Custom Rules; specialist dots for companies are in pilots D1 D2 |
| Teams | One account, paid per seat, seats can be handed to teammates; new phones are approved only on the computer itself | Small-business edition M2; Meta Enterprise Platform announced, team admin features not published M6 | Business Premium; Enterprise beta when an admin turns it on D2 |
| Price | $20 per seat per month, or $200 per seat per year; model costs paid to the model vendor | Free with usage limits; Power $20 / month; Maximum $100 / month M4 | First dot included in Pro ($100 / $200 / $500 tiers) or Business Premium ($125 per seat per month, $100 billed annually) D2 D4 D5 |
| Regions | Wherever Claude Code / Codex are available. Mainland China: OpenCode with a Chinese model you sign up for; our website and relay are outside China, so access can be slow | US and Canada M2 | Pro excludes the EEA, Switzerland and the UK; Business Premium in all ChatGPT regions D2 |
| Best for | People and teams already using Claude Code / Codex / OpenCode who want to hand everyday work to an agent and keep the data on their own machines | US and Canadian users who want help with personal errands and no setup | ChatGPT Pro or business customers who want an assistant researching and getting things done in the background all day |
出处 · Sources
全部是厂商自己的页面，查阅于 2026-10-02。
- M1 Meta, Introducing Muse (2026-09-08): about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
- M2 Meta, Muse for Small Business (2026-09-29): about.fb.com/news/2026/09/introducing-muse-small-business/
- M3 Meta, The Biggest News From Connect 2026 (2026-09-24): about.fb.com/news/2026/09/the-biggest-news-from-connect-2026/
- M4 Meta Help Center, About Muse subscriptions: meta.com/help/subscriptions/1021145227643680/
- M5 Meta AI, Muse: ai.meta.com/muse/
- M6 Meta, Launching Meta Enterprise Platform (2026-09-28): about.fb.com/news/2026/09/launching-meta-enterprise-platform/
- D1 OpenAI, Introducing dots (2026-09-29): openai.com/index/introducing-dots/
- D2 OpenAI Help Center, Getting started with your dot: help.openai.com/en/articles/20001530
- D3 OpenAI Help Center, Dots privacy, security, and safety FAQs: help.openai.com/en/articles/20001529
- D4 OpenAI Help Center, About ChatGPT Pro tiers: help.openai.com/en/articles/9793128
- D5 OpenAI, Premium seats are coming to ChatGPT Business: openai.com/index/premium-seats-chatgpt-business/
- D6 ChatGPT pricing (plan comparison, "Content is used to train our models"): openai.com/chatgpt/pricing/
发现哪一条过时或写错了，请写信到下面的地址，我们核对后改。