Build a Cloudflare WebMCP Gateway: Turn Any Website Into an AI Agent Tool in 2026

Why WebMCP Needs a Gateway

Cloudflare's WebMCP, launched during Agents Week 2026, enables any Cloudflare-proxied site to expose MCP tools. But 80% of websites are not on Cloudflare. This gateway server bridges the gap: it detects WebMCP-enabled sites and uses them directly, while falling back to Playwright browser automation for everything else. The result is a single MCP server that makes ANY website agent-accessible.

Server Implementation (src/webmcp-gateway.ts)

// src/webmcp-gateway.ts
import { FastMCP } from 'fastmcp';
import { z } from 'zod';
import { chromium, Browser } from 'playwright';
import httpx from 'undici';

const server = new FastMCP({
  name: 'webmcp-gateway',
  version: '1.0.0',
});

let browser: Browser;
async function getBrowser() {
  if (!browser) browser = await chromium.launch({ headless: true });
  return browser;
}

// Tool 1: Browse any page and extract content
server.tool(
  'browse_page',
  'Navigate to a URL and extract structured content',
  {
    url: z.string().url().describe('Target URL'),
    extract_type: z.enum(['full', 'text_only', 'links', 'headlines']).default('full'),
    wait_for: z.string().optional().describe('CSS selector to wait for'),
  },
  async ({ url, extract_type, wait_for }) => {
    // Check for native WebMCP first
    const hasWebMCP = await checkWebMCP(url);
    if (hasWebMCP) {
      return await webmcpExtract(url, extract_type);
    }

    // Fallback to Playwright
    const page = await (await getBrowser()).newPage();
    try {
      await page.goto(url, { waitUntil: 'networkidle', timeout: 30000 });
      if (wait_for) await page.waitForSelector(wait_for, { timeout: 10000 });

      let content: any;
      switch (extract_type) {
        case 'text_only':
          content = await page.evaluate(() => document.body.innerText);
          break;
        case 'links':
          content = await page.evaluate(() =>
            Array.from(document.querySelectorAll('a[href]')).map(a => ({
              text: a.textContent?.trim(), href: a.href
            })).filter(l => l.text && l.href)
          );
          break;
        case 'headlines':
          content = await page.evaluate(() =>
            Array.from(document.querySelectorAll('h1,h2,h3')).map(h => ({
              level: h.tagName, text: h.textContent?.trim()
            }))
          );
          break;
        default:
          content = await page.evaluate(() => ({
            title: document.title,
            text: document.body.innerText.slice(0, 5000),
            links: Array.from(document.querySelectorAll('a[href]')).slice(0, 20).map(a => ({
              text: a.textContent?.trim(), href: a.href
            })),
            images: Array.from(document.querySelectorAll('img[src]')).slice(0, 10).map(img => ({
              src: img.src, alt: img.alt
            }))
          }));
      }

      return { content: [{ type: 'text', text: JSON.stringify({ url, content, method: 'playwright' }, null, 2) }] };
    } finally {
      await page.close();
    }
  }
);

// Tool 2: Click an element on a page
server.tool(
  'click_element',
  'Click an element on a web page by selector',
  {
    url: z.string().url(),
    selector: z.string().describe('CSS selector for the element'),
    wait_after: z.number().optional().default(2000).describe('ms to wait after click'),
  },
  async ({ url, selector, wait_after }) => {
    const page = await (await getBrowser()).newPage();
    try {
      await page.goto(url, { waitUntil: 'networkidle', timeout: 30000 });
      await page.click(selector);
      await page.waitForTimeout(wait_after || 2000);

      return {
        content: [{
          type: 'text',
          text: JSON.stringify({
            success: true,
            new_url: page.url(),
            new_title: await page.title(),
            content_preview: (await page.evaluate(() => document.body.innerText)).slice(0, 2000)
          }, null, 2)
        }]
      };
    } finally {
      await page.close();
    }
  }
);

// Tool 3: Fill a form
server.tool(
  'fill_form',
  'Fill form fields on a web page',
  {
    url: z.string().url(),
    fields: z.record(z.string()).describe('Map of CSS selector -> value'),
    submit_selector: z.string().optional().describe('Submit button selector'),
  },
  async ({ url, fields, submit_selector }) => {
    const page = await (await getBrowser()).newPage();
    try {
      await page.goto(url, { waitUntil: 'networkidle', timeout: 30000 });
      for (const [selector, value] of Object.entries(fields)) {
        await page.fill(selector, value);
      }
      if (submit_selector) {
        await page.click(submit_selector);
        await page.waitForTimeout(3000);
      }
      return {
        content: [{
          type: 'text',
          text: JSON.stringify({
            filled: Object.keys(fields),
            submitted: !!submit_selector,
            final_url: page.url()
          }, null, 2)
        }]
      };
    } finally {
      await page.close();
    }
  }
);

// Tool 4: Take screenshot
server.tool(
  'screenshot',
  'Capture a screenshot of a web page',
  {
    url: z.string().url(),
    full_page: z.boolean().default(false),
  },
  async ({ url, full_page }) => {
    const page = await (await getBrowser()).newPage();
    try {
      await page.goto(url, { waitUntil: 'networkidle', timeout: 30000 });
      const screenshot = await page.screenshot({
        path: `/tmp/screenshot_${Date.now()}.png`,
        fullPage: full_page,
      });
      return {
        content: [{ type: 'text', text: `Screenshot saved to /tmp/screenshot_${Date.now()}.png` }]
      };
    } finally {
      await page.close();
    }
  }
);

async function checkWebMCP(url: string): Promise<boolean> {
  try {
    const resp = await httpx.fetch(`${url}/.well-known/mcp.json`, {
      method: 'HEAD',
      signal: AbortSignal.timeout(3000)
    });
    return resp.status === 200;
  } catch { return false; }
}

async function webmcpExtract(url: string, type: string) {
  const resp = await httpx.fetch(`${url}/mcp/tools/call`, {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ tool: 'extract_page', arguments: { url, format: type } })
  });
  const data = await resp.json();
  return { content: [{ type: 'text', text: JSON.stringify({ ...data, method: 'webmcp' }, null, 2) }] };
}

server.start({ transport: 'stdio' });

Performance Benchmarks

Production Reality Check

Rate-limit handling: Playwright consumes 50-80MB per browser context. Run a pool of 5 contexts with connection recycling. For Cloudflare WebMCP sites, rate limits are 100 req/min. Memory management: Close browser pages immediately after extraction. A leaked page consumes 80MB and never garbage collects. Use try/finally blocks. Failure recovery: If Playwright crashes, restart the browser pool automatically. Implement a health check that tests browser responsiveness every 60 seconds.

By <a href="https://x.com/deeepakbagada" rel="nofollow noopener noreferrer">Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last tested: August 2026 with FastMCP 3.14, Playwright 1.52, Node v22, and Cloudflare WebMCP preview.

Related Architecture & Implementation Resources

• Browse complementary servers and client connectors in the Daily AI World MCP Directory.
• Integrate this tool into multi-agent pipelines with our AI Workflows Blueprints.
• Review frontier LLM capabilities and token metrics on Latest AI News.