Adopting the WebMCP Standard — Reducing AI Web Browsing Tokens by 90%

The era of AI agents having to capture screens one by one or aggressively scrape text while web browsing is coming to an end. This is because global standard bodies like the W3C and the FIDO Alliance have begun to fully adopt new browser standards, WebMCP and AP2, that treat AI agents as official web users. With the web ecosystem moving rapidly—notably with the core API structure changing in Chrome 150—let's take a light look at the changes coming to our lives and development environments.

From Screen Captures to Structured Data — How WebMCP is Changing Agent Browsing

Imagine an AI agent navigating the standard websites we use today. Until now, for an agent to browse the web, it had to capture screenshots to analyze them as images or force its way through complex HTML code. Because machines were struggling to interpret visual layouts designed for humans, it was both slow and expensive.

WebMCP is a standard designed to solve this inefficiency. It is a browser API specification that helps websites provide machine-readable tools and data directly to AI agents, instead of forcing them to use the complex visual layouts intended for humans. Simply put, it provides a 'digital menu' exclusively for AI at the browser level.

The impact is immediate: since there is no need to analyze entire screens as images, AI agents can save up to 90% on token usage. With an active public origin trial running from Chrome 149 through 156, and Microsoft Edge also actively providing support, the complex and costly landscape of agent browsing is set to fundamentally change in the near future.

From Navigator to Document — Security and Structural Changes in Chrome 150

Why did the standards related to web agents suddenly change in Chrome 150? The key lies in where the API used by AI is located within the browser.

In the initial draft, the browser global object navigator.modelContext was used. However, this approach posed risks: the tool state accessible to the agent could persist even after the user navigated to a different website, or sensitive data could bleed across different tabs. Therefore, the W3C standards group and the Google Chrome team moved it to document.modelContext to ensure it is strictly synchronized with the lifecycle of individual web pages. As a result, when a page is closed or refreshed, the tools are cleanly cleared, making security boundaries much more robust.

Security improvements don't stop there. To prevent attacks where malicious ads or external scripts overwrite an AI tool, the provideContext API, which injected tool sets in bulk, was discontinued. It was replaced by a registerTool method, where tools are verified and added one by one. Additionally, a cleanup mechanism using AbortSignal was introduced to ensure tool registrations are safely revoked when a user navigates away from a page.

Here is how you can safely register an AI agent tool using actual JavaScript in accordance with the Chrome 150 specification.

// Chrome 150+ WebMCP API 구현 예시
const controller = new AbortController();

try {
  await document.modelContext.registerTool({
    name: "get_product_stock",
    description: "상품 ID를 바탕으로 실시간 재고 상태를 조회합니다.",
    inputSchema: {
      type: "object",
      properties: {
        productId: { type: "string" }
      },
      required: ["productId"]
    },
    // 간접 프롬프트 주입 공격을 미연에 방지하기 위한 보안 설정
    untrustedContentHint: false,
    readOnlyHint: true
  }, {
    signal: controller.signal // 컴포넌트 소멸 또는 페이지 이탈 시 자동 등록 해제
  });
} catch (error) {
  console.error("WebMCP 도구 등록 실패:", error);
}

Developers can now safely expose tools to AI agents while marking data regions that might contain malicious instructions (untrustedContentHint), and control whether read-only tasks require an approval pop-up (readOnlyHint). A secure, agent-centric web ecosystem is finally being built from the ground up within our browsers.

The Agent with a Wallet — FIDO AP2 and the Era of Micro-payments

Can we go beyond just having agents browse the web, and enable them to pay for paid services or purchase goods on our behalf? To make this a reality, the Agent Payment Protocol (AP2) standard, designed by Google and donated to the FIDO Alliance, was introduced. AP2 supports various methods like credit cards, bank transfers, and stablecoins, helping agents conduct transactions safely in online stores.

The key here is not giving the agent full access to your wallet, but rather implementing safety measures that allow the agent to spend money only within the limits authorized by the user. It strictly distinguishes between 'intent delegation,' where users immediately approve payments only for items in their cart, and 'long-term autonomous payment' limits, where agents pay for items within pre-set constraints. This dual-security system prevents AI malfunctions or fraudulent payment incidents at the source.

Taking it a step further, a future economic ecosystem is being prepared where agents help each other and automatically settle micro-payments. With the combination of the Agent-to-Agent (A2A) communication protocol and the x402 micro-payment standard, an autonomous economic system where AIs perform minor tasks for each other and settle tiny payments in real-time is expected to become possible.

Angular 22 Support Begins — Building Agent-Friendly Webs

The development ecosystem is moving fast, with Angular 22 introducing experimental features that make it easy to integrate WebMCP tools into the framework. Microsoft Edge already supports this standard, and other major browsers are hurrying to follow suit.

The future of web development will extend beyond creating intuitive interfaces for humans to building machine-friendly data layers that AI agents can understand without error. As we prepare for a future where agents act as official users of the web, try applying WebMCP to your projects today!