WebMCP：让网站与AI智能体实现对话

谷歌与微软在 W3C Web Machine Learning Community Group 联合提出了一项名为 WebMCP 的 Web 标准草案，旨在让网页通过 JavaScript API 主动向 AI 智能体声明可调用的结构化工具，以取代传统的屏幕抓取方式1。目前该草案仍为社区组草案，尚未进入正式的 W3C 标准轨道，但已在 Chrome 浏览器中开启试验，用户可通过 chrome://flags/#enable-webmcp-testing 标志进行体验1。

在技术细节上，WebMCP 的核心机制包括工具发现（Discovery）、基于 JSON Schema 定义的模式（Schemas）以及状态（State）共享，其 API 入口已从 navigator.modelContext 迁移至 document.modelContext1。该标准的安全模型要求工具仅在源隔离（origin-isolated）的文档中注册，默认的 toolsPermissions 策略为 self，且敏感操作可强制要求弹出用户确认对话框1。为展示其应用潜力，文章作者部署了名为 Career Copilot 的实验项目，演示了智能体如何通过 WebMCP 进行工具的链式调用1。

Google (Chrome) and Microsoft (Edge) have jointly introduced WebMCP, a Web standard draft developed within the W3C Web Machine Learning Community Group 1. This initiative aims to enable web pages to proactively declare callable, structured tools to AI agents through JavaScript APIs, offering a modern alternative to traditional screen scraping methods 1. Although currently classified as a Community Group draft and not yet on the formal W3C standard track, the technology is already undergoing experimental testing in the Chrome browser via the chrome://flags/#enable-webmcp-testing flag 1.

The core mechanisms of WebMCP encompass tool discovery, state sharing, and schemas defined by JSON Schema, and the API entry point has been migrated from navigator.modelContext to document.modelContext 1. To ensure security, the standard emphasizes operating within the user's logged-in browser tabs and mandates that tools be registered exclusively within origin-isolated documents, setting the default toolsPermissions policy to self 1. Furthermore, the framework allows for mandatory user confirmation dialogs when sensitive operations are triggered 1.

To demonstrate the practical capabilities of this new standard, the author deployed an experimental project named Career Copilot 1. This job-seeking assistant illustrates how an AI agent can execute chained tool calls to accomplish complex tasks while maintaining human oversight by requiring approval at critical steps 1.