好友与播客

使用端到端加密聊天、经隐私遮盖的分享卡、按关系档位回复的管家，以及区分发言人的当面播客录音。

在 OpenMousse app 的「对话」页顶部将开关切到「好友」一侧，即可管理联系人、发起私聊、发送分享卡或录制播客；点 + 按钮可添加好友，点「我的名片 Agent」可打开名片 Agent 设置。

OpenMousse 支持你与互为好友的联系人直接沟通、分享你与 AI 的讨论片段、通过各自的管家核对时间，以及当面共同录制播客并整理纪要。所有好友互动均建立在相互邀请、端到端加密与严格的权限隔离之上。

核心功能

•端到端加密聊天：好友关系仅在双方相互邀请后建立，支持已读回执、表情回应、引用回复、编辑与收回消息。每条消息均附带签名并端到端加密，直接发往好友的服务器；若某一方没有公网地址，消息会经由 OpenMousse 中继转发。中继仅能看到收件地址、数据大小与发送时间，看不到消息内容，也不保存任何消息。
•隐私遮盖分享卡：你可以将自己与 AI 的一段讨论做成分享卡发给好友。在发出前，系统会先逐条识别并遮盖其中的隐私信息，逐一询问你是将其模糊成「一个地方」「一位朋友」这类泛称，还是照实保留。
•管家（对外代表你） 即将推出：由当前的「名片 Agent」升级而来。好友或好友的管家可以向你的管家询问你的忙闲与近况（例如「周四晚上有空吗」）。你可将好友设为亲近、朋友、同学三档关系，并分别设定每一档可读取的内容范围：日程（给详情、只给忙闲或不给）、近况、分享过的内容、学习笔记、住址。管家使用不带任何工具的模型调用进行回复，回复中绝不出现未放出的住址、电话、邮箱、身体数字与家人名字。遇到约时间、要东西等涉及时间、金钱或承诺的事，管家不能自己答应，一律回答「这个得（你的名字）本人定，我去问一下。」并生成确认卡交给你；每一句回复发出前还要经 Doorman 复查，不妥即扣下交给你定。
•当面录制播客：和朋友当面讨论时，只需将一部手机放在中间即可开启播客录音。系统能自动区分每个人的声音（录完后由你核对、认人），将原话按人分好并整理成一份纪要；从中提炼出的关于你的观点与决定可写进世界树。

安全与隔离边界

•好友 agent 的消息只当资料：好友的 agent 发来的话，均由不能使用任何工具的模型调用处理，无法让你的 Agent 动用任何工具；若其中夹带指令，会在「今天 → 待处理」中生成安全提醒。
•管家看不到核心隐私：管家完全看不到世界树，也看不到任何健康与身体数据；陌生人的询问默认一律不理。

不在这里的东西

•没有好友动态：系统不设朋友圈动态流。
•播客纪要不会群发给参与者：播客录完整理出的原话与纪要仅留在发起人自己的 app 里，不会自动发给参与录音的朋友。

另见

•先问你（收件箱）：了解管家确认卡、Doorman 扣留与安全提醒的处理方式。
•数据与安全：了解中继服务器、转写服务与外来信息隔离的详细说明。

最后更新于 2026-10-19这一页有帮助吗？

For a complete index of the docs, see llms.txt. Append .md to any URL for the Markdown version; this page's Markdown is at friends.md.

Features

Friends and Podcasts

Use end-to-end encrypted chat, privacy-redacted share cards, relationship-tiered Butler replies, and speaker-diarized in-person podcasts.

Flip the switch at the top of Chats in the OpenMousse app to Friends to manage contacts, start direct chats, send share cards, or record podcasts; tap the + button to add a friend, and tap My card agent to open your card agent settings.

OpenMousse lets you communicate directly with mutually invited contacts, share excerpts of your discussions with AI, coordinate schedules through each other's Butlers, and record in-person podcasts together with generated minutes. All friend interactions are built on mutual invitations, end-to-end encryption, and strict permission isolation.

Core Features

•End-to-end encrypted chat: A friend connection is established only after both sides invite each other. Chat supports read receipts, emoji reactions, quoted replies, editing, and unsending messages. Every message is signed, end-to-end encrypted, and sent directly to your friend's server; if either side does not have a public address, messages are forwarded through the OpenMousse relay. The relay can see only the recipient address, payload size, and timestamp; it cannot read message content and stores no messages.
•Privacy-redacted share cards: You can turn an excerpt of your discussion with AI into a share card and send it to a friend. Before sending, the system identifies and masks private details one by one, asking for each item whether to generalize it into a generic phrase such as "a place" or "a friend," or keep it as is.
•The Butler (representing you externally) Coming soon: An upgrade of today's card agent. Friends or their Butlers can ask your Butler about your availability and recent updates (for example, "Are you free Thursday evening?"). You can assign each friend to one of three relationship tiers—close, friend, or classmate—and configure the scope of content each tier is allowed to read: schedule (full details, free/busy only, or none), recent updates, previously shared content, study notes, and home address. The Butler replies using a model call with no tools attached, and its replies never include unshared home addresses, phone numbers, email addresses, body metrics, or family members' names. When a request involves scheduling, asking for items, or anything involving time, money, or commitments, the Butler cannot agree on your behalf; it always replies, "That's up to (your name) to decide—I'll go ask," and generates an approval card for you. Every reply is also reviewed by Doorman before being sent; anything inappropriate is held back for your decision.
•In-person podcast recording: When discussing with friends in person, place a single phone in the middle to start a podcast recording. The system automatically separates each speaker's voice (you verify and assign speakers after recording), groups the transcript by speaker, and compiles meeting minutes; distilled opinions and decisions about you can be written to the memory tree.

Security and Isolation Boundaries

•Messages from a friend's agent are treated strictly as reference material: Messages sent by a friend's agent are processed by a model call that cannot use any tools, so they cannot cause your agents to invoke any tools; if a message contains embedded instructions, a security alert is generated in Today → To do.
•The Butler cannot see core private data: The Butler has no access to the memory tree and no access to any health or body data; inquiries from strangers are ignored by default.

What Is Not Here

•No friends feed: The system does not have a social feed or timeline.
•Podcast minutes are not broadcast to participants: The transcript and minutes compiled after a podcast recording stay only in the initiator's own app and are not automatically sent to the friends who participated in the recording.