Agent in Chrome does not collect, store, or transmit any personal data to us. We operate no servers. There is no account, no analytics, no telemetry, no cookies, and no advertising.
Agent in Chrome lets an AI agent running on your own computer read and operate browser tabs that you explicitly hand to it. It does not attach to any tab automatically — attaching always follows an explicit instruction, and Chrome displays its own banner on any tab while it is attached.
This is the part that matters, so we state it precisely rather than generally.
The extension makes no network requests of its own. It communicates only with a native messaging host installed on your own computer. There is no server operated by us for it to talk to.
However, the purpose of the product is to give page content to an AI agent, so page content does travel. The path is:
Every hop but the last is on your own computer. The last hop is whichever AI client you chose to connect — for example Claude Desktop, Claude Code, or Cursor. If that client sends data to a cloud model provider, page content leaves your machine at that point, governed by that provider's privacy policy rather than by this one. We neither receive that data nor are technically able to.
Values of password fields, credit-card numbers, CVV codes, and one-time
verification codes — detected by the input's type and
autocomplete attributes — are never included in page-reading
results. They are reported only as "filled (N characters)", so the
agent knows a field is populated without learning its contents.
Local operation logs redact values by field name.
None of this is transmitted anywhere. Removing the extension and running the uninstaller deletes it.
debugger — to read page structure and simulate real user input
through the Chrome DevTools Protocol. This is what lets the extension work
correctly on pages that ordinary content scripts cannot handle reliably,
including Shadow DOM and cross-origin iframes.<all_urls> — you decide which tabs to hand over. The
extension cannot know in advance which sites those will be, so it cannot
request a narrower list.nativeMessaging — to talk to the local host.tabs, tabGroups — to open its own tabs and keep them
in a clearly labelled, separately coloured tab group, so you can always see
at a glance which pages the agent is touching.scripting — to read page structure from inside the page.storage, alarms — settings, and keeping the service
worker alive.We do not sell or share your data, because we never receive any.
Questions about this policy: liang@liangai.org
Agent in Chrome 不收集、不存储、也不向我们传输任何个人数据。我们没有运营任何服务器， 没有账号体系、没有统计埋点、没有 cookie、没有广告。
Agent in Chrome 让一个运行在你自己电脑上的 AI agent 读取和操作你显式交给它的标签页。 它不会自动接管任何标签页——接管必须来自一条明确的指令，而且接管期间 Chrome 会在该标签页 顶部显示它自己的调试提示条。
这一节是关键，所以写得具体，而不是泛泛而谈。
扩展自身不发起任何网络请求，它只和安装在你本机的 native messaging host 通信。 我们没有任何服务器可供它连接。
但这个产品的用途本身就是把页面内容交给 AI agent，所以页面内容确实会流动。路径是：
除最后一跳外，全部发生在你自己的电脑上。最后一跳是你自行选择连接的 AI 客户端—— 例如 Claude Desktop、Claude Code 或 Cursor。 如果那个客户端把数据发送给云端模型服务，页面内容就在这一步离开了你的设备， 此后受该服务的隐私政策约束，而不是本政策。这些数据我们既收不到，技术上也无从收到。
密码框、信用卡号、CVV、一次性验证码的值——依据输入框的 type 与
autocomplete 属性识别——不会出现在任何读取页面的结果里，
只会被报告为「已填写（N 字符）」：agent 知道这个框填过了，但读不到内容。
本机的操作留痕也会按字段名打码。
这些都不会被传输到任何地方。移除扩展并运行卸载程序即可删除。
debugger — 通过 Chrome DevTools Protocol 读取页面结构、模拟真实用户输入。
这是它能在普通 content script 处理不可靠的页面上正常工作的原因，包括 Shadow DOM
和跨源 iframe。<all_urls> — 交出哪些标签页由你决定，扩展无法预先知道那会是哪些站点，
因此无法申请更窄的范围。nativeMessaging — 与本机 host 通信。tabs、tabGroups — 开自己的标签页，并收进一个有明确标签、
独立配色的标签组，让你一眼就能看出 agent 在动哪些页面。scripting — 在页面内部读取页面结构。storage、alarms — 保存设置，以及维持 service worker 存活。我们不出售、不共享你的数据，因为我们从未收到过任何数据。
关于本政策的问题：liang@liangai.org