端云同境：本地的留在本地，助理照样干活Hybrid Context: local stays local, the assistant still does the work
员工电脑上的文件、命令、浏览器登录态，助理直接调用；内容不经服务器保存，只出站、不开端口，一个总闸随时拉。Files, commands and browser sessions on an employee's own computer, used directly by the assistant; content never stored on the server, outbound only, with a kill switch.
企业用 AI 助理，很快会撞上一堵墙：真正有用的东西，一半在平台上（订单、库存、知识库），另一半在员工自己的电脑上（一份供应商 Excel、一个登录过的网站、一条只有在本机才能跑的命令）。把后者搬到云上，既不现实，也不该。
端云同境（Hybrid Context）就是为这堵墙做的：让平台上的助理能用到员工电脑上的能力，而内容不离开那台电脑。
一条只出站的通道
员工电脑上跑一个很小的常驻程序，它主动连到平台，平台永远不反向连它，电脑上也不开任何端口。助理要调一个本地工具时，请求顺着这条通道下来，结果顺着同一条通道回去，在服务器内存里过一道就转给助理，不落库、不落盘。
四条规矩从第一天起就是硬的：
- 只出站，不监听。
- 服务器不留副本。工具结果不写任何持久化存储。
- 绑定到人。一台电脑配对到某个用户名下，只有他自己的助理能用它。
- 平台编排，本地执行。AI 决定调什么，真正的动作发生在本地。
装什么、开放什么，用户自己挑
通道装好只是一条空管道。要让助理能看哪个目录、跑哪些命令，用户在页面上从目录里挑「连接器」装上：装是守护进程做的，进度推回页面；装好不等于启用，启用前先在表单里填清楚授权目录、允许的命令。全程不用用户手打任何路径——上一版原型就栽在这里，产品不能让人填 /某个/venv/bin/python。
每个连接器在页面上只有有限几种状态，而且每种状态都写明下一步：未安装、安装中、已安装未启用、在跑、起不来（把原因原文摆出来）、已停用。状态由守护进程一处算出来推上来，页面不做第二套判断。
三层总闸
- 本地总闸：菜单栏或页面一按，所有调用一律拒绝，守护进程仍在线，重启后也还是关的，直到人工恢复。
- 单连接器闸：停掉某一个，别的照常。
- 服务端闸：撤销设备令牌。本地闸防「我现在不想让它读」，服务端闸防「这台电脑不在我手上了」。
拉闸时助理收到的是一句人话：「用户暂停了这台电脑，现在读不到任何本地内容」，并且明说不要重试。
它长什么样
一句话的事：「供应商列表.xlsx 里有哪些供应商，Fortbrain 里他们的采购价是多少？」助理在本机读表、在平台查价、合成一张小表回给你。表没有离开过你的电脑，价格没有离开过平台。
首页第 05 章有一个可以点的演示，走的就是这条路。
Companies that adopt AI assistants soon hit a wall: half of what is useful lives on the platform (orders, stock, knowledge bases) and the other half lives on the employee's own computer (a supplier spreadsheet, a website they are logged into, a command that only runs locally). Moving the second half to the cloud is neither realistic nor right.
Hybrid Context exists for that wall: the assistant on the platform can use what is on the employee's computer, and the content never leaves that computer.
An outbound-only channel
A small resident program on the employee's computer dials out to the platform. The platform never connects back, and nothing listens on the computer. When the assistant calls a local tool, the request travels down that channel and the result travels back up it, passing through server memory once on its way to the assistant: never written to a database, never written to disk.
Four rules have been hard from day one:
- Outbound only, nothing listens.
- The server keeps no copy. Tool results never touch persistent storage.
- Bound to a person. A computer is paired to one user, and only that user's assistant can use it.
- The platform orchestrates, the computer executes. The AI decides what to call; the action happens locally.
The user chooses what to install and what to open
A freshly installed bridge is an empty pipe. To let the assistant see a folder or run commands, the user picks "connectors" from a catalogue on the page. The daemon installs them and reports progress back; installed is not enabled, and enabling first means filling in a form: the allowed folder, the allowed commands. The user never types a path. The previous prototype failed exactly there: a product cannot ask people to type /some/venv/bin/python.
Each connector has only a handful of states on the page, and every state names the next step: not installed, installing, installed but off, running, failed (with the raw reason shown), stopped. The daemon computes the state in one place and pushes it; the page does not run a second judgement.
Three kill switches
- Local: one press in the menu bar or on the page refuses every call. The daemon stays online, and the switch survives a restart until a person lifts it.
- Per connector: stop one, the rest keep working.
- Server side: revoke the device token. The local switch covers "I don't want it reading right now"; the server switch covers "that computer is no longer in my hands".
When the switch is down the assistant gets a plain sentence: "The user paused this computer; no local content can be read", with an explicit "do not retry".
What it looks like
One sentence: "Which suppliers are in suppliers.xlsx, and what are their purchase prices in Fortbrain?" The assistant reads the sheet on your machine, looks up prices on the platform, and returns a small table. The sheet never left your computer; the prices never left the platform.
Chapter 05 on the homepage has a clickable demo that walks exactly this path.