The best MCP servers for browser automation in 2026

The right one depends on which browser your agent should work in. In your own signed-in Chrome: Tabhand, Playwright MCP’s extension or Playwriter. In a clean browser, to test and debug your app: Playwright MCP or Chrome DevTools MCP. In the cloud, for agents on servers: Browserbase, Kernel or Browser Use.

We make Tabhand, so weigh our pick with that in mind. Every fact about the others comes from their own docs and code, checked on October 27, 2026.

Quick picks

• Your signed-in Chrome, several agents at once, on a Mac: Tabhand.
• Your signed-in Chrome on Windows or Linux: Playwright MCP with its browser extension.
• Testing your own app in a fresh browser: Playwright MCP.
• Profiling and debugging a page: Chrome DevTools MCP.
• Hundreds of browsers on servers: Browserbase, Kernel or Browser Use Cloud.

In your own signed-in Chrome

These work in the browser you already use, so your agent is signed in wherever you are: no sign-in pages, no copied cookies.

Tabhand

A Chrome extension and an MCP server. Each agent opens tabs of its own in the background, in a tab group with its own color and cursor, and can’t reach your tabs or another agent’s. One click on the toolbar icon pauses them all. npx tabhand setup adds it to Claude Code, Codex, Cursor and 16 more agents.

• Good to know: macOS and Google Chrome only, for now. No debugging tools: no performance traces, no network log.
• Price: free, MIT license.

Playwright MCP, with its extension

Microsoft’s Playwright MCP normally opens a browser of its own. Add --extension and install the Playwright Extension, and it connects to your running Chrome or Edge instead. The first time, you choose which tab it gets, and you approve each connection unless you save a token. Several agents can connect at once, and each gets its own tab group.

• Good to know: works on Windows, Linux and macOS, in Chrome and Edge.
• Price: free, Apache 2.0.

Playwriter

An extension plus a CLI and MCP server that gives your agent the full Playwright API through a single tool. You turn it on per tab by clicking its icon until it turns green. Separate sessions can each get their own tab group.

• Good to know: its local relay sends anonymous usage stats unless you set PLAYWRITER_TELEMETRY=0.
• Price: free, MIT license.

Browser Use’s Browser Harness

Lets Claude Code, Codex or Cursor drive your running Chrome through remote debugging, which you switch on at chrome://inspect in Chrome 144 or later. Browser Use’s own guide says it “sees your whole Chrome, not a chosen tab or site”, and its CLI sends usage telemetry by default. How it compares.

• Price: free, MIT license.

Browser MCP

An extension and an MCP server with 100,000 users on the Chrome Web Store. The version that’s live, 1.x, works on one tab that you connect by clicking. Browser MCP 2.0, announced on October 26, gives each AI app tabs of its own in a named group; on October 27, npm and the Chrome Web Store still served 1.x.

• Good to know: Chrome, Edge and Brave.
• Price: free.

Claude in Chrome and OpenAI’s extension

Each AI company’s browser extension serves its own agent: Claude in Chrome needs a paid Claude plan, and OpenAI’s works from the ChatGPT app. Neither is an MCP server you can add to other agents. Claude in Chrome compared, OpenAI’s compared.

In a clean browser, for testing and debugging

These start a browser of their own, which is what you want for testing your app the way a new visitor sees it.

Playwright MCP

Microsoft’s browser MCP server, with about 39,000 GitHub stars. It drives Chrome, Firefox, WebKit or Edge through accessibility snapshots rather than screenshots. Its profile keeps logins between runs once you sign in there, and --isolated starts fresh each time. How it compares.

• Price: free, Apache 2.0.

Chrome DevTools MCP

From the Chrome DevTools team: performance traces with LCP, INP and CLS, Lighthouse audits, network requests and console messages. It starts Chrome with a profile of its own, or attaches to yours with --autoConnect. How it compares.

• Price: free, Apache 2.0.

agent-browser

Vercel Labs’ fast command-line browser for agents, with an MCP mode. It launches its own Chrome for Testing, can start from a copy of your Chrome profile, and gives parallel agents sessions of their own.

• Price: free, Apache 2.0.

In the cloud, for agents on servers

These rent your agent browsers in a data center. They start signed out, and they’re built for scale: many browsers at once, stealth modes and CAPTCHA solving.

• Browserbase. A hosted MCP server whose tools run through its Stagehand library. The free plan is one browser hour a month. How it compares.
• Kernel. A hosted MCP server for its cloud browsers, with managed sign-ins (three sites on the free plan) and usage billed by the second. How it compares.
• Browser Use Cloud. Browsers from $0.02 an hour, plus an agent of its own that takes a goal. How it compares.

Side by side

Try ours

If your agents work on your Mac in sites you’re signed in to, Tabhand takes one command:

npx tabhand setup

Checked on October 27, 2026 against each project’s own docs and repository: Playwright MCP and its extension, Playwriter, Browser Harness, Browser MCP and its changelog, Chrome DevTools MCP, agent-browser, Browserbase, Kernel and Browser Use. If something here is out of date, tell us.