Skip to content

Playwright MCP vs Claude in Chrome vs a Cloud Browser

Playwright MCP vs Claude in Chrome vs a cloud browser: pick by whose logins it uses, where it runs and who can step in, with measured token costs.

Janis Zesiger6 min readShare
Three browsers an AI agent can drive side by side: a separate browser on a laptop, the person's own signed-in browser, and a cloud session watched on a phone

Playwright MCP vs Claude in Chrome comes down to whose browser your AI agent drives: Playwright MCP starts a separate browser on your computer, Claude in Chrome uses your own signed-in Chrome. A cloud browser is the third way, a session on a provider's servers that you watch and take over from any device.

Playwright MCP vs Claude in Chrome vs a cloud browser: what is the difference?

MCP (Model Context Protocol) is the standard plug that lets AI apps such as Claude Code, Cursor or Codex use a tool, here a browser. Playwright is Microsoft's free tool for driving browsers from code.

Browser options for AI agents, as of November 2026
OptionWhose browserWhere it runsWho can step inCost
Playwright MCPIts own (yours with an add-on)Your computerYou, if it has a windowFree
Chrome DevTools MCPIts own (yours on request)Your computerYou, if it has a windowFree
Claude in ChromeYours, signed inYour computerYou, in your ChromePaid Claude plan
Cowork built-in browserClaude's, separateIn Claude DesktopYou, in the side panelPaid Claude plan
Hosted cloud browserA separate sessionProvider's serversYou, from any devicePlan or per hour

Playwright MCP or Claude in Chrome: which is the best browser MCP for you?

Most people end up with two: one tool for their own code and one browser for the rest of the web.

  • You test your own app, by hand or in automated tests, or need Firefox or WebKit (Safari's engine)Playwright MCP, or Playwright CLI in a coding agent
  • You debug slow pages, network requests or console errorsChrome DevTools MCP
  • A quick task in an account you are signed in to, while you watchClaude in Chrome
  • You work in Claude Desktop and want Claude kept away from your own tabsThe built-in browser in Claude Cowork
  • The agent runs in claude.ai, ChatGPT, a phone app, WSL or a server, and you want to step in at loginsA hosted browser MCP such as Browser.lol
  • You build a product that starts hundreds of browsers from codeBrowserbase, Steel, Browser Use Cloud or Kernel

Playwright MCP vs Playwright CLI vs Chrome DevTools MCP

Playwright comes in two forms; Google's server adds the debugging both lack.

Playwright MCP

For any MCP client

Reads each page as a text outline of its buttons and fields. Drives Chromium, Firefox and WebKit.

Playwright CLI

For coding agents

Short shell commands instead of MCP tools. Its README calls it "the best fit" for coding agents.

Chrome DevTools MCP

For debugging

Google's server adds performance traces, network requests and console errors. Chrome only.

DevTools MCP sends usage statistics to Google by default, and --no-usage-statistics turns that off.

Is Claude in Chrome safe to use?

Claude in Chrome is a Chrome extension, generally available since August 2026. It "shares your browser's login state", and you set how much it does alone with three modes, Manual, Auto or Skip, from approving each action to no checks at all.

3.8%

of prompt injections (hidden orders on a page) that reached Claude Opus 5 succeeded without extra safeguards

0%

succeeded against Opus 5 and Sonnet 5 with Anthropic's safeguards on

0.3%

succeeded against Fable 5 with the safeguards on

These are Anthropic's own tests, and a browser MCP server you add yourself gets none of those safeguards. Even so, Anthropic "strongly advise[s] against" using its browsers for financial accounts, medical information or other people's personal data.

Which cloud browser MCP is best for AI agents?

A hosted browser connects through a remote MCP server, a web address instead of a program. Browser.lol is ours; the other rows link to each vendor's pages, checked on 17 November 2026.

Hosted browser MCP servers, as of November 2026
ServiceConnect withWho picks the clicksYou take overPrice
Browser.lol (ours)Sign-in or API keyYour modelYes, and the agent can ask youPremium plan
BrowserbaseAPI key in the URLIts own model (Stagehand)Through the live view1 free hour, then from $20/month
SteelSign-in or API keyYour modelInteractive live view$30 one-time credit, $0.10/hour
Browser Use CloudAPI keyIts own model, whole tasksAfter the run stops$0.06/hour, no subscription
KernelSign-in or API keyYour modelWith a control link$5/month credit, plans from $30/month

Browser.lol is a cloud browser for AI agents made for one person rather than a fleet: add one address, sign in and click Allow. Its agent connection needs Premium, drives Chrome-family browsers only and offers no raw connection for your own Playwright code.

A Browser.lol live view of a session an AI agent is driving, with the Agent is working bar and a Take over button at the bottom of the page
The live view: the Take over button pauses the agent.

How many tokens does a browser MCP use?

Tokens are small pieces of words: everything the model reads for one step counts, and you pay per token. Two measurements from autumn 2026 (68 runs, MCP vs CLI) give real numbers.

4,568

tokens of tool descriptions in Playwright MCP 0.0.85

18,407

tokens for one snapshot of the Hacker News front page

205,600

tokens for one full snapshot of a long Wikipedia article

340

tokens to find one table cell by text search, instead of 43,445

Everyday tasks in the 68-run test used 90,000 to 200,000 tokens, about $0.16 to $0.22 each, mostly Claude Code's own instructions sent again with every step. So the number of steps drives the bill. Claude Code caps one MCP result at 25,000 tokens by default, so the full Wikipedia snapshot never reaches the model; other clients may pass it through.

Frequently asked questions

Can Playwright MCP use my logged-in Chrome?

Yes: with the --extension option and the Playwright Extension it drives your running Chrome or Edge, and with it every site you are signed in to. By default it keeps a profile of its own between runs, and --isolated starts clean each time.

Is Playwright MCP faster than Claude in Chrome?

Not clearly. In Ui.Vision's test Playwright's tools needed 1 to 2 seconds of tool time against 16 to 41 seconds per task for Claude in Chrome, but with the model's thinking an eight-field form took Playwright MCP 32 seconds and Claude in Chrome 21. In Dejo's benchmark Playwright MCP got 17 of 19 tasks right; nobody has published a comparable rate for Claude in Chrome.

Is it safe to open Chrome's remote debugging port for an agent?

Only with a separate profile: Google warns that "any application on your machine can connect to this port". Since Chrome 136 it no longer works on your default profile, because attackers used it to steal cookies.

Can I use two browser MCP servers at once?

Yes, Claude Code, Cursor and Codex run several side by side. Name the one you mean in your prompt so the agent does not pick the wrong browser.

Janis Zesiger

Founder of Browser.lol and Guard.ch

Janis Zesiger founded Browser.lol and Guard.ch and runs Zesiger.net in Switzerland. Janis builds the isolated session platform behind both products and writes about browser security, privacy and AI agents that use the web.

All articles