跳到正文Skip to content
OpenMousse
完整的文档索引见llms.txt。在网址后加.md可以拿到 Markdown 版本；本页的 Markdown 版本是roadmap.zh-CN.md。
概览

路线图

查看正在开发与即将上线的功能模块。

本页列出 OpenMousse 正在合入主分支、已在内部实例运行或已完成方案设计的功能模块。各模块按当前所处阶段分组列出，测试完成后依次进入公开仓库。

提示：关于发布节奏

本页仅反映工程进度与已确定的产品方案，不设固定发布日期。已上线的功能请参阅左侧目录中的「功能」与「指南」各页。

正在合入与内部运行

以下模块已在开发分支实现，或在内部实例中稳定运行，正在整理合入公开主分支（main）：

  • •自研 Python 内核（agent harness） 即将推出：负责模型调用、上下文管理与安全规则执行。包含按发言方划分的三类权限回合（本人、赛博、好友的 agent 与陌生人）、工具执行三档（只读直接执行、撤得回附带改动编号并支持回复「撤销」恢复、撤不回生成锁定参数的确认卡），以及主模型不可用时自动切换备用模型并注明。
  • •程序调用接口 即将推出：在内核中提供兼容 OpenAI 的 API 与 ACP 协议，支持代码编辑器及其他外部程序将 OpenMousse 作为模型直接调用。
  • •世界树每周修剪 即将推出：每周日自动排查重复或相近的世界树叶子，提议合并或删除，并交由你逐条确认后生效。
  • •Agent 编写的隔离页面 即将推出：运行在隔离环境中，无法读取登录凭证与其他系统数据，也无法向外部发送数据；首次打开时列出所需读取的数据项，经你同意后仅作只读访问，代码更新后需重新确认。

方案已定，正在开发

以下模块已完成交互与架构方案设计，正在推进客户端或服务端开发：

  • •电脑端与 Android 健康接入：推出电脑端（网页工作台与 Mac 客户端）；在 Android 客户端支持接入 Health Connect 读取睡眠、心率与训练数据。
  • •托管专属云服务器：提供官方托管的专属云服务器选项，当前可在官网首页加入等候名单。
  • •记忆页与年轮视图：客户端底栏第四格更新为「记忆」，世界树移入其中。新增年轮视图：外圈为今天，向内为过往日期；每个点代表一条新记忆，停在环外的琥珀点代表等待你确认的条目。
  • •赛博（长期思考辅助）：拥有独立性格、记忆与模型的 Agent，在特定方向上长期辅助思考。资料页为一张包含「身份、记忆、装载」三项的证件；赛博不直接访问你的数据，查询资料与执行操作均交由其装载的功能 Agent 完成，修改数据或写入世界树须经确认卡批准。
  • •管家（由「名片 Agent」升级）：对外代表你，按亲近、朋友、同学三档关系回答好友及好友管家的询问；仅读取对应档位开放的信息，看不到世界树与健康数据，涉及时间、金钱与承诺的事项一律交由你本人决定，且每句对外回复均经 Doorman 复查。
  • •交互式新建 Agent 流程：支持「一句话描述 → 系统补问 → 预填草稿 → 试聊满意后再正式建立」的创建方式。

另见

  • •欢迎使用 OpenMousse：了解当前已上线的核心组件与运行边界。
  • •内核（agent harness）：了解内核的权限回合与工具分档设计。
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 roadmap.md.
Overview

Roadmap

View features currently in development and upcoming releases.

This page lists OpenMousse modules that are being merged into the main branch, running on internal instances, or have finalized designs in active development. Modules are grouped by current stage and land in the public repository once testing completes.

Note: Release Cadence

This page reflects current engineering progress and finalized product designs without fixed release dates. For features that are already live, see the pages under Features and Guides in the sidebar.

Merging and Running Internally

The following modules are implemented on development branches or running steadily on internal instances and are being prepared for merge into the public main branch:

  • •In-house Python engine (agent harness) Coming soon: Manages model calls, context, and security rule enforcement. Includes three speaker permission turns (you, a Cyber, and a friend's agent or stranger), three tool execution tiers (read-only runs immediately; reversible runs and returns a change ID that can be rolled back by replying Undo; irreversible pauses and creates a parameter-locked approval card), and automatic fallback to a backup model with a note when the primary model is unavailable.
  • •Programmatic API Coming soon: Exposes an OpenAI-compatible API and ACP protocol in the engine so code editors and other external programs can call OpenMousse directly as a model.
  • •Weekly memory tree pruning Coming soon: Checks the memory tree every Sunday for duplicate or near-duplicate leaves, proposes merges or deletions, and applies them item by item after you confirm.
  • •Isolated agent-authored pages Coming soon: Run inside an isolated environment that cannot read login credentials or other system data and cannot send data externally; opening a page for the first time lists the data items it wants to read and grants read-only access only after you approve, requiring re-approval whenever its code changes.

Design Finalized, in Development

The following modules have finalized interaction and architecture designs and are under active client or server development:

  • •Desktop clients and Android Health Connect: Launching desktop clients (a web workspace and a Mac app), and adding Health Connect support on Android to read sleep, heart rate, and workout data.
  • •Hosted dedicated cloud servers: Offering an official hosted cloud server option; you can join the waitlist on the home page now.
  • •Memory tab and tree-ring view: Updating the fourth tab in the bottom bar to Memory and moving the memory tree inside it. Adds a tree-ring view where the outer ring is today and inner rings are past dates; each dot is a new memory, and amber dots resting outside the ring are entries waiting for your confirmation.
  • •Cybers (long-term thinking companions): Agents with their own personality, memory, and model that assist your thinking in a specific direction over the long term. Each Cyber's profile page is a badge with three sections (Identity, Memory, and Mounts); a Cyber never accesses your data directly, delegates all lookups and actions to the functional agents mounted on it, and requires approval via an approval card before modifying data or writing to the memory tree.
  • •The Butler (upgraded from Card agent): Represents you externally and answers questions from friends and friends' Butlers across three relationship tiers (close, friend, acquaintance); reads only the information opened to that tier, cannot see the memory tree or health data, hands anything involving time, money, or commitments to you in person, and has every outbound reply reviewed by Doorman.
  • •Interactive agent creation flow: Supports creating a functional agent through a "one-line description → system follow-up questions → pre-filled draft → test chat before finalizing" flow.

See Also

  • •Welcome to OpenMousse: Learn about the core components and operating boundaries that are live today.
  • •Engine (Agent Harness): Learn about speaker permission turns and three-tier tool execution in the engine.
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.