cdpfleet Pre-launch Get early access

cdpfleet / CDP & Puppeteer

Chrome DevTools Protocol

Puppeteer, raw CDP, remote browsers.

Add "cdp": true to a launch and the response carries a cdpUrl next to wsUrl: a browser-level DevTools WebSocket for Puppeteer, Playwright's connectOverCDP in any version, or your own CDP client. The proxy, timeouts, threads and billing are the same as any other session.

CDP docsBuild the code
  • 6browsers with a cdpUrl
  • anyPuppeteer or Playwright version
  • 0browsers to install locally

Which browsers, and how to connect

  • cdp: true works on Chrome, Edge, Brave, Opera, Yandex and Chromium. Patchright, Rebrowser, Camoufox, Firefox and WebKit speak Playwright's own protocol only.
  • Puppeteer: puppeteer.connect({ browserWSEndpoint: cdpUrl, headers: { "x-api-key": KEY } }) with puppeteer-core — the browser runs on cdpfleet, nothing to download.
  • Playwright connectOverCDP doesn't need a client of the server's exact version, unlike connect(). You get the Chromium-level API (CDP & Puppeteer).
  • One URL that launches and connects: a Connect URL works for Playwright, Playwright Test and CDP clients that can only take a WebSocket address.
  • Recordings: a headless session is recorded when it has cdp or live_view (headless-recording recipe).
import puppeteer from 'puppeteer-core';
const KEY = process.env.CDPFLEET_API_KEY;
const res = await fetch('https://starter.cdpfleet.com/chrome/session', {
  method: 'POST',
  headers: { 'x-api-key': KEY, 'content-type': 'application/json' },
  body: JSON.stringify({ headless: 'new', cdp: true }),
});
const { cdpUrl } = await res.json();
const browser = await puppeteer.connect({ browserWSEndpoint: cdpUrl, headers: { 'x-api-key': KEY } });
const page = await browser.newPage();
await page.goto('https://example.com');
await browser.close(); // ends the session

What CDP adds

  • Core Web Vitals in lab runs: LCP, CLS and TTFB come back from one evaluate with buffered observers, from all five languages; through a proxy, TTFB includes reaching the server via the proxy (case 55).
  • A slow phone on a fast server: one newCDPSession call reaches DevTools' CPU throttling — the same workload ran 2.7–4.0× slower at 4× and 4.1–7.3× slower at 6× (case 56). No cdp: true needed for that: Playwright's newCDPSession works over the normal connection.
  • The cookie jar, read and edited: filter by URL, change a value or expiry, delete one cookie, and carry a login to another session as about 150 bytes of JSON (case 57).
  • Block telemetry, keep the page: news front pages sent dozens of analytics requests per load and kept reporting after the scraper left; blocking them changed no titles or headings (case 59).

What a page can tell

  • The classic CDP tell is gone: the Error.stack getter trick fired in none of six setups — not over CDP, not with Playwright's protocol (case 43).
  • What remains is the browser, not the client: the Error.message getter fires on every Chromium build, whether or not anything is listening, and logging a large object costs measurably more while a client receives the console (case 43).
  • Headless itself is the bigger signal on real protected sites: what gets through where and what a page can read about your browser.

Read next

Web scraping →AI agents →Bot protection →Proxies →