欢迎使用 OpenMousse

了解 OpenMousse 的核心组件、数据归属与运行边界。

主对话健身饮食学习与办公代办

这周精力不错，把落下的训练补上

会议纪要 · 健身 / 饮食 / 学习与办公周二、周四晚上各加一次 40 分钟力量训练；训练日晚餐多一份蛋白质。查看会议确认

今天待处理● 代办请求「提交并付款」，等你批准日程19:30 力量训练 · 40 分钟报告睡眠报告 · 恢复分 86

主对话里的会议纪要，和手机上的「今天」

OpenMousse 是开源的个人 AI agent 系统（AGPL-3.0）。你的对话、日常数据、工作与学习资料以及好友间的讨论均保存在只属于你的服务器上，从中提炼的事实、偏好、决定与成熟的观点沉淀为长期记忆（世界树）。多个功能 Agent 在此基础上，按你原有的工作方式负责安排、提醒与代办。

系统运行在你的自有服务器上，也可以运行在我们即将开放的托管专属云服务器上（可在官网首页加入等候名单）。客户端提供 iPhone（TestFlight 邀请）、Android（安装包）与网页版；电脑端（网页工作台与 Mac 客户端）即将推出。

你会得到什么

•可查阅、可修改的世界树。 每一条关于你的事实、偏好、决定与观点都存为一片带出处的叶子，标明来自哪一次对话、哪一天。你可以随时查看、修改或删除，也可以将其存为 Markdown 笔记放入自己的笔记库。
•跨 AI 应用共用的长期记忆。 Claude、ChatGPT、Gemini、Notion 与 Claude Code 可通过各自的专属地址接入世界树，回答前先调取你的背景，并将新信息写回世界树，不再各存一份。
•职责分明的功能 Agent 与会议。 主对话是总入口，负责把事务转交给健身、饮食、睡眠、代办、学习与办公等功能 Agent，或召集它们开短会并出具纪要；每个功能 Agent 拥有独立的记忆、工具与看板（支持表格、看板、日历、时间线、图表五种视图）。
•先放下想法的 Zen。 刚读完的文章或一闪而过的念头可以先存入 Zen，只存不发，AI 不会介入；准备好之后再挑出相关内容点「讨论」一起推敲，点「完成」整理并「保存到库」，成熟的观点写入世界树。
•撤不回的操作先问你。 付款、对外发送、提交、登录等操作一律先停下来，生成锁定参数的确认卡放入收件箱与今天页的「待处理」，等你批准后才执行。
•隔离的代办沙箱与 Doorman。 代办 Agent 在隔离的沙箱中打开网页、比价、填表与预订；密码由 Doorman 保管，代办只拿到占位符，请求离开服务器的那一刻才换成真值并发往绑定的网站。
•有边界的好友互动与管家。 好友消息经签名并端到端加密；对外分享的讨论先逐条遮盖隐私；管家按亲近、朋友、同学三档关系回答忙闲与近况，看不到世界树与健康数据，涉及时间、金钱、承诺的事一律生成确认卡交给你定。
•长期辅助思考的赛博 即将推出。赛博拥有独立的性格、记忆与模型，在特定方向上长期辅助你思考；它不直接访问你的数据，查资料或改动数据均交由其装载的功能 Agent 处理。

提示：OpenMousse 的边界与承诺

OpenMousse 不会在未经你确认的情况下对外发信、付款或承诺时间。我们不投放广告，不出售数据，不用你的数据训练模型。每一轮对话发往你所选模型服务（包括备用模型）的内容，仅限该轮消息与回答所需的记忆；其余记忆与文件均留在你的服务器上。

开发者与架构

OpenMousse 采用自研的 Python 内核（agent harness）即将推出管理模型调用、上下文与工具权限分档。如果你希望了解发言方权限回合、通过兼容 OpenAI 的 API 或 ACP 协议调用系统，或查看源码结构，请参阅「内核（agent harness）」与 GitHub 仓库（github.com/openmousse/openmousse）。

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 welcome.md.

Overview

Welcome to OpenMousse

Learn about OpenMousse's core components, data ownership model, and operating boundaries.

Main chatFitnessMealsStudy & WorkErrands

Energy's good this week—let's catch up on the workouts I missed

Meeting minutes · Fitness / Meals / Study & WorkAdd a 40-min strength session on Tue and Thu evenings; extra serving of protein at dinner on training days.OpenConfirm

TodayTo do● Errands wants to "Submit & pay"—awaiting your approvalSchedule19:30 Strength · 40 minReportsSleep report · Recovery 86

Meeting minutes in main chat and Today on a phone

OpenMousse is an open-source personal AI agent system (AGPL-3.0). Your conversations, daily data, study and work materials, and discussions with friends stay on a server that belongs to you. Facts, preferences, decisions, and distilled opinions from those interactions settle into long-term memory (the memory tree). Built on that foundation, functional agents handle scheduling, reminders, and errands around the way you already work.

The system runs on your own server or on a dedicated cloud server hosted by OpenMousse (opening soon; join the waitlist on the home page). Clients are available for iPhone (via TestFlight invitation), Android (APK installer), and the web app; desktop clients (a web workspace and a Mac app) are coming soon.

What You Get

•An inspectable, editable memory tree. Every fact, preference, decision, and opinion about you is stored as a provenance-tracked leaf recording which conversation and which day it came from. You can view, edit, or delete any leaf at any time, or store leaves as Markdown notes in your own vault.
•Long-term memory shared across AI apps. Claude, ChatGPT, Gemini, Notion, and Claude Code connect to the memory tree through dedicated URLs, retrieving your background before answering and writing new information back to the memory tree instead of keeping separate copies.
•Functional agents with clear boundaries and short meetings. Main chat is the single entry point that routes tasks to functional agents such as Fitness, Diet, Sleep, Errands, and Study & Work, or calls them into a short meeting and produces minutes; each functional agent has its own memory, tools, and board (supporting table, kanban, calendar, timeline, and chart views).
•Zen for setting thoughts down first. Articles you just read or fleeting thoughts go into Zen as notes without sending—no AI steps in; when you are ready, pick related items to Discuss together, tap Done to organize them, and Save to the vault while distilled opinions are written to the memory tree.
•Approval before irreversible actions. Payments, outbound messages, submissions, and logins always pause and generate a parameter-locked approval card in the inbox and under To do in Today, executing only after you approve.
•An isolated Errands sandbox and Doorman. The Errands agent opens web pages, compares prices, fills forms, and books inside an isolated sandbox; Doorman holds your passwords and gives Errands only placeholders, swapping in the real values only at the moment a request leaves the server for its bound domain.
•Bounded friend interactions and the Butler. Friend messages are signed and end-to-end encrypted; shared discussions have private details redacted item by item first; the Butler answers availability and status questions across three relationship tiers (close, friend, acquaintance), cannot see the memory tree or health data, and turns anything involving time, money, or commitments into an approval card for you to decide.
•Cybers for long-term thinking Coming soon. Each Cyber has its own personality, memory, and model to assist your thinking in a specific direction over the long term; a Cyber never accesses your data directly and delegates any data lookup or change to the functional agents mounted on it.

Note: OpenMousse's Boundaries and Commitments

OpenMousse never sends external messages, makes payments, or commits your time without your confirmation. OpenMousse serves no ads, sells no data, and never trains models on your data. On each turn of a conversation, only that turn's message and the memories needed to answer are sent to your chosen model provider (including any fallback model); all other memories and files remain on your server.

Developers and Architecture

OpenMousse uses an in-house Python engine (agent harness) Coming soon to manage model calls, context, and tiered tool permissions. To learn about speaker permission turns, calling the system via the OpenAI-compatible API or ACP protocol, or the source layout, see Engine (Agent Harness) and the GitHub repository (github.com/openmousse/openmousse).

Last updated 2026-10-19Was this page helpful?

这一页给 agent 读的版本：在网址后加 .md 就是下面这份纯文本，和给人看的那页由同一份内容生成。Agent-readable version of this page: append .md to the URL to get the plain text below, generated from the same source as the human view.