There is a specific moment where teams start shopping for a cloud browser. Someone wired Playwright into an agent, it worked on a laptop, then production started failing for reasons unrelated to the code: Chromium eating 900MB a tab, /dev/shm at 64MB crashing renderers, a login repeated cold every run, and eventually a block page. The instinct is to rent all of it, which is often right. Sometimes it means buying an anti-blocking product to fix a memory problem.
I'm Ajay, I build PandaStack, and PandaStack is one of the seven below. Hard numbers are only for my own platform; everyone else gets described by what they're built for, because pricing here changes monthly and a confidently wrong figure is worse than none. Check current docs.
What a cloud browser actually sells you
The category is narrow. A vendor runs Chromium somewhere and gives you a WebSocket URL speaking the Chrome DevTools Protocol. You point Playwright or Puppeteer at it and drive a browser you never installed. On top they sell some mix of four things: session lifecycle, proxies and geolocation, anti-bot work, and observability.
The connect call is the same everywhere and differentiates nobody:
import { chromium } from "playwright";
import { writeFile } from "node:fs/promises";
// Every vendor in this post hands you a variant of this URL. The token in the
// query string is usually the only thing that changes between them.
const wsEndpoint = process.env.BROWSER_WS_ENDPOINT!; // wss://.../?token=...
const browser = await chromium.connectOverCDP(wsEndpoint);
// connectOverCDP attaches to an EXISTING browser, so it usually already has a
// context. Calling newContext() blindly is how people end up with a fresh
// cookie jar and a mysteriously logged-out session.
const context = browser.contexts()[0] ?? (await browser.newContext());
const page = await context.newPage();
await page.goto("https://example.com/login", { waitUntil: "domcontentloaded" });
await page.fill("#email", process.env.LOGIN_EMAIL!);
await page.fill("#password", process.env.LOGIN_PASSWORD!);
await page.click("button[type=submit]");
await page.waitForURL("**/dashboard");
// The valuable artefact is not the screenshot. It is this.
await writeFile("session.json", JSON.stringify(await context.storageState()));
await browser.close();Note the last two lines. Logging in is the expensive step: slow, rate-limited, sometimes needing a human. How a platform keeps that state changes your bill more than stealth.
Where Hyperbrowser sits
Hyperbrowser is one of the newer entrants: managed headless browser sessions with a CDP endpoint, plus agent-oriented APIs on top. That's a deliberately neutral description and I'm leaving it there — I don't track their tiers closely enough to summarise them honestly, and a stale comparison table is how these posts go wrong. Read their docs, then use the criteria below.
The criteria that actually decide this
Work out which of these is forcing your hand:
- Scraping-under-blocking versus an agent driving your own app. Identical in code, completely different problems. Extracting data from sites that don't want you to is an arms race: residential IPs, fingerprints, captchas. An agent clicking through your own product needs none of it.
- Session persistence. Ask what actually persists — cookies only, or the full profile including IndexedDB and service workers. That gap is where 'it worked in dev' incidents live.
- Concurrency and cost per session-hour. Browser work is bursty and mostly waiting, and a parked session costs the same as a working one. Multiply peak concurrent sessions by average session length; that is your bill, not the headline rate.
- Proxy and geo. Datacentre IPs, or residential in twelve countries? A hard gate that disqualifies most of this list if it's the latter.
- Observability. When an agent misbehaves at 3am, can you see what it saw? Homegrown setups underinvest here.
- Isolation. A browser rendering a page you didn't write executes untrusted code. Chromium's sandbox is good; the question is what's underneath — a shared kernel next to other tenants, or a hardware boundary.
The seven alternatives
1. Browserbase
The category-defining product and the default for a first browser-driven agent. Managed sessions over CDP, stealth and proxies handled, session recording and a live view. Pick it when you want the browser problem to stop being yours.
2. Browserless
The longest-running option, and the one with a real self-hosting story: a container you run in your own cloud account, or consume hosted. Pick it when 'inside our VPC' is a hard constraint and you still want a supported product.
3. Steel.dev
Open-source browser infrastructure aimed at AI agents, with a hosted offering alongside the code. The pitch is that agents need different primitives from test runners — session reuse, resumable state, structured extraction — and open source lets you read how sessions are managed.
4. Scrapybara
Broader than a browser: remote desktop instances for computer-use agents, where the browser is one application among several. If your agent needs a PDF viewer, a spreadsheet, or a desktop app with no web equivalent, a CDP-only vendor cannot help.
5. Bright Data's browser API
The heavyweight for scraping-under-blocking. Their business is the proxy network — residential and mobile IPs at a scale nobody else here operates — and the browser product is that network with a CDP endpoint attached. No infrastructure cleverness substitutes for IP diversity. Pick it when the adversary is the target site.
6. Playwright or Puppeteer on a VM you own
Dismissed too fast. Playwright ships a browser download command and a container image; running it on a VM or in CI is a couple of hours. For a nightly scrape of cooperative sites, renting browser infrastructure solves a problem you don't have. The costs arrive later: concurrency ceilings, zombie cleanup, upgrades, no session recording.
7. PandaStack
Mine, so here's the shape rather than a pitch. We ship a browser template: a Firecracker microVM running Ubuntu 24.04 with Node 24, Playwright and a real Chromium, plus the Python crawl4ai and trafilatura stack, in a 4 GiB guest. One hardware-isolated VM per session with its own kernel, and root inside it — your Playwright version, your apt packages, arbitrary ports. Creates restore at p50 179ms.
The different part is copy-on-write fork. Log in once, snapshot the whole machine — browser process, profile, disk, memory — then branch a fresh VM off it per job. Not a cookie jar replayed into a cold browser: the same machine, resumed.
from pandastack import Sandbox
# The browser template: Ubuntu 24.04, Node 24, Playwright with a real Chromium,
# plus the Python crawl4ai / trafilatura stack. 4 GiB guest, own kernel, root.
sbx = Sandbox.create(template="browser", ttl_seconds=900,
metadata={"job": "orders-export"})
SCRIPT = """
import asyncio
from playwright.async_api import async_playwright
async def main():
async with async_playwright() as p:
# Launched locally, inside this VM. No CDP over the public internet,
# no shared pool -- the browser and your code are on the same machine.
browser = await p.chromium.launch()
ctx = await browser.new_context(storage_state="/workspace/session.json")
page = await ctx.new_page()
await page.goto("https://example.com/orders", wait_until="domcontentloaded")
print(await page.title())
await page.screenshot(path="/workspace/orders.png", full_page=True)
await browser.close()
asyncio.run(main())
"""
try:
sbx.filesystem.write("/workspace/run.py", SCRIPT)
r = sbx.exec("cd /workspace && python3 run.py", timeout_seconds=180)
if r.exit_code != 0:
raise RuntimeError(r.stderr)
png = sbx.filesystem.read("/workspace/orders.png")
finally:
sbx.kill()# Log in ONCE, freeze the machine, then branch a VM per job from the frozen
# state. Each fork is copy-on-write: pages are shared until something writes.
base = Sandbox.create(template="browser", ttl_seconds=3600)
base.filesystem.write("/workspace/login.py", LOGIN_SCRIPT)
base.exec("python3 /workspace/login.py", timeout_seconds=300)
ready = base.snapshot() # browser profile, disk and memory, all of it
for account in accounts:
job = Sandbox.fork(ready) # already authenticated, no login round-trip
try:
job.exec(f"python3 /workspace/export.py {account}", timeout_seconds=300)
finally:
job.kill()Who should not pick PandaStack
Plainly: we do not run a managed residential proxy pool. We do not offer captcha solving. We do no fingerprint or stealth tuning — no patched Chromium, no TLS signature work, no evasion layer. Those are whole products with dedicated teams.
You can bring your own proxy — it's a real VM, so you set Chromium's proxy flags or the guest's routing — but you maintain that relationship. We fit when the browser is a tool your agent uses rather than a weapon: driving your own product, testing your own app, or a partner's admin panel that never shipped an API.
How to choose
- Answer the blocking question first. If target sites are adversarial, your shortlist is Bright Data or a stealth-first vendor.
- If you need a desktop and not just a browser, it's Scrapybara or a VM you control — a CDP-only product cannot open a PDF viewer.
- If compliance says page content cannot leave your infrastructure: self-hosted Browserless, self-hosted Steel, or your own VMs.
- If you're not fighting anyone, start with Browserbase or Hyperbrowser and revisit when the bill hurts.
- If logging in is your dominant cost, look hard at resumable state — full-machine snapshots, or documented profile persistence rather than cookies alone.
- Multiply peak concurrent sessions by average session length before signing.
The short version
Browserbase if you want it handled. Browserless if it must run in your account. Steel for open source with agent ergonomics. Scrapybara if the task is a desktop. Bright Data if you're being blocked. Your own Playwright if the volume is small and the sites are friendly. PandaStack if you want the browser inside a machine you control, hardware-isolated per session, with a logged-in state you can snapshot and branch — and you don't need proxies, captchas or stealth.
Two things matter on day one whichever you pick: know whether your workload is adversarial, and never build a system that logs in every run.