MCP Servers for Browser Automation: Official Options
Three browser servers are published by the companies that build them: Playwright MCP from Microsoft, Chrome DevTools MCP from the Chrome DevTools team, and the Browserbase server from Browserbase. Playwright and Chrome DevTools run on your computer and start with npx. Browserbase runs in the cloud at https://mcp.browserbase.com/mcp, or you can run it yourself. Browser servers give an AI app a browser it can click through, so only connect one you'd trust with the sites you open.
What you'll have at the end
You'll know which official browser server fits your job, whether it should run locally or in the cloud, and how it differs from WebMCP. Each server below links to the source it was checked against.
Playwright MCP and Chrome DevTools MCP are local servers. They start a browser on your machine and the AI app talks to them over stdio, the same way it talks to any local server. Browserbase has a hosted version, so your AI app connects to a URL and the browser runs on Browserbase's infrastructure. For the difference between the two transports, read local vs remote MCP servers.
Before you start
- Node.js 22.19 or later, for
npx. The Playwright and Chrome DevTools READMEs both start their servers withnpx. - Google Chrome installed, if you use Chrome DevTools MCP. Its README lists it as a requirement.
- A Browserbase account and API key, if you use the Browserbase server.
Local and cloud browser servers
The two local servers drive a browser on the machine where the AI app runs. The Browserbase server can work either way: its hosted version runs the browser for you, and the self-hosted version runs locally and calls Browserbase's cloud browsers.
| Server | Publisher | Where the browser runs | Connection | Sign-in | Read-only mode | Risk note |
|---|---|---|---|---|---|---|
| Playwright MCP | Microsoft | Your computer | Local, stdio (npx @playwright/mcp@latest) | The README describes no sign-in for the server. You sign in to sites inside the browser it opens. | Not documented. Use --allowed-origins, --allowed-hosts and --blocked-origins to limit where it can go. | The browser can open any site the model asks for, so restrict it with the origin flags. |
| Chrome DevTools MCP | Chrome DevTools team | Your computer, in Chrome | Local, stdio (npx -y chrome-devtools-mcp@latest) | The README describes no sign-in for the server. | Not documented. The README offers --slim for basic tasks, which exposes fewer tools. | The README says to avoid sharing sensitive or personal information you don't want the client to see. |
| Browserbase MCP | Browserbase | Browserbase's cloud, when hosted. Your computer, when self-hosted. | Hosted: remote at https://mcp.browserbase.com/mcp. Self-hosted: stdio with npx @browserbasehq/mcp. | Browserbase API credentials, set in the MCP configuration | Not documented | Pages you open in a hosted session run on Browserbase's infrastructure, not yours. |
The Playwright server starts a headed browser window by default. Add --headless to run without a window. The Chrome DevTools README also lists --headless and --isolated among its server options. Check its configuration guide for what --isolated does before you rely on it.
Start a local server
This config starts Playwright MCP with the origin flags set. Replace example.com with the sites you want the browser to reach:
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": ["-y", "@playwright/mcp@latest", "--allowed-origins", "example.com", "--headless"]
}
}
}
The file's location depends on your AI app. The Cursor and Claude Code guides show where each one reads its config. For Chrome DevTools MCP, use the same structure with "command": "npx" and "args": ["-y", "chrome-devtools-mcp@latest", "--slim", "--headless"].
Connect to the hosted Browserbase server
Clients that accept a remote server take the URL https://mcp.browserbase.com/mcp. Browserbase's MCP introduction lists the steps: choose hosted Streamable HTTP or local stdio, then set up your Browserbase API credentials in the MCP configuration. The hosted option is the one Browserbase recommends for setup.
Browser servers and WebMCP solve different problems
A browser MCP server runs outside the page. It starts or attaches to a browser, then uses the browser the way a person would: it reads the page, clicks, types and takes screenshots.
WebMCP runs inside the page. A website publishes its own tools to agents, and an agent can call those tools directly instead of working out which buttons to click. The site decides what each tool does.
| Browser MCP server | WebMCP | |
|---|---|---|
| Who builds it | The browser or automation vendor | The website |
| Where it runs | Beside the browser, on your machine or in the cloud | Inside the page |
| What it can do | Anything a user can do in the browser, by driving the page | Only the tools the site has registered |
| Typical use | Testing, scraping, filling forms on sites you don't control | A site offering its own tools to visiting agents |
Use a browser server when you need to drive a site that has no tools of its own. Use WebMCP when you're building the site and want to publish specific actions for agents. For the full comparison, read WebMCP vs MCP, and for the proposal itself, what is WebMCP.
A browser server can read whatever the browser shows, including pages where you're signed in. Before you connect one, use a browser profile or a hosted session without accounts you care about, and limit the sites it can open with the origin flags where the server supports them.
Search the MCPBundles directory
The MCPBundles directory indexes MCP servers from many publishers, including browser servers. Search it when you want a server with setup steps for your app, rather than reading each vendor's README first.
- Open www.mcpbundles.com and type a keyword in the search box. For browser servers, try
https://www.mcpbundles.com/?q=browser, orhttps://www.mcpbundles.com/?q=playwrightfor one tool by name. - Open a listing and check the publisher and the sign-in it asks for. The directory lists third-party products, and MCPBundles isn't affiliated with them unless the listing says so.
- Choose your AI app in the setup tabs on the listing and follow the steps shown there.
Before you connect any server, read evaluate MCP servers and the MCP security best practices guide.
FAQ
Which official MCP server should I use for browser automation?
It depends on where the browser should run. Playwright MCP from Microsoft and Chrome DevTools MCP run on your computer. The Browserbase server runs in the cloud when you use its hosted URL. Each one is published by the team behind the browser tool, so start with its README.
Is a browser MCP server the same as WebMCP?
No. A browser MCP server drives a browser from outside the page. WebMCP lets a website register its own tools inside the page, so an agent can call them without clicking through the UI. The WebMCP vs MCP guide compares the two.
Can I stop Playwright MCP from visiting every site?
Yes. The Playwright README documents the --allowed-origins, --allowed-hosts and --blocked-origins flags. List only the sites you need in the start command, and check the README for the current format before you rely on them.
Do I need to sign in to a local browser server?
The server itself doesn't sign in to anything. If you need to reach a site that requires an account, you sign in to that site inside the browser the server opens. Hosted Browserbase sessions use your Browserbase API credentials instead.
- WebMCP vs MCPWhere each one runs, who calls it, and when to use which or both.
- MCP server examplesOfficial and reference servers grouped by what they do.
- Evaluate MCP serversA checklist to run before you connect a server.
- Local vs remote MCP serversstdio, Streamable HTTP, and who can reach each.
- How to find MCP serversWhere to look beyond the official examples.