The Best MCP Servers for Browser Automation (2026)

For tests in a clean, isolated browser, use Playwright MCP: it is Microsoft's official server and runs Chrome, Firefox or WebKit. To debug performance, network requests or console errors, use Chrome DevTools MCP: it records real DevTools traces. To let a coding agent work in your own logged-in Chrome, use Browser Harness: it attaches to the Chrome you already use. In our matched eval of four tools, run as a CLI and skill on a cloud browser, it tied the others on score and used the fewest input tokens. To hand whole web tasks to cloud browsers, many at once, use the Browser Use Cloud MCP server: one tool call runs a full browser agent. If you already pay for a cloud browser vendor, use its server: Browserbase, Kernel, Steel or Hyperbrowser.

TL;DR

What an MCP Server for Browser Automation Does

The Model Context Protocol (MCP) lets an AI app like Claude Code, Cursor or Claude Desktop call tools from another program. A browser MCP server turns browser actions into those tools: open a page, click, type, read the page, take a screenshot.

The biggest difference between servers is which browser they drive:

• Their own browser. The server starts a fresh Chrome or Chromium with its own profile. Clean and repeatable, but you are logged out of everything.
• Your Chrome. The server attaches to the Chrome you already use, with your tabs, cookies and logins.
• A cloud browser. The server drives a browser on a vendor's machines. You can run many at once and keep your own machine out of it, but you pay per browser-hour.

MCP is not the only way to give a coding agent a browser. Playwright's own README says coding agents may benefit from a CLI with skills instead, because CLI calls are more token-efficient: the agent does not load large tool schemas and accessibility trees into context. Several tools below ship both.

Choosing an MCP Server

Four questions narrow it down.

• Which browser? Logged-in work on your own accounts needs your Chrome. Tests want an isolated profile. Parallel or unattended work wants cloud browsers.
• Low-level tools or a whole agent? Most servers expose clicks and page reads, so your agent does every step. Some, like the Browser Use Cloud server and Browserbase, take an instruction in plain words and run it, which keeps raw page snapshots out of your agent's context.
• Context cost. Every page snapshot your agent reads costs input tokens. On long tasks this is most of the bill.
• Cost and license. Local servers are free and open source. Cloud servers bill for browser time.

Our Matched Eval

We ran four of these tools on 100 public tasks from BU Bench V2.1, with the same agent (Codex CLI with GPT-6 Luna), the same Browser Use cloud browser and the same judge. Only the browser tool changed. Playwright MCP and Chrome DevTools MCP ran as MCP servers. Browser Harness ran as its CLI and skill, not as browser-harness-mcp. The data file does not record whether agent-browser ran as its CLI or its MCP server.

On score, no tool won. Paired differences, Browser Harness minus each tool: Playwright MCP +0.4 points (95% interval -5.7 to +6.6), Chrome DevTools MCP -0.6 (-6.6 to +5.6), agent-browser +4.0 (-1.8 to +10.0). Every interval includes zero. Browser Harness used the fewest input tokens and cost the least per task. By median time Playwright MCP was fastest, then Chrome DevTools MCP, Browser Harness and agent-browser.

Limits: one run per task, 100 of the benchmark's 200 tasks, one agent (Codex) and one model, and a fresh cloud browser for every task, not your logged-in Chrome. Method, versions and every per-task score are in the data file.

The MCP Servers

1. Playwright MCP

The best choice for tests, isolated sessions and cross-browser checks.

Playwright MCP is Microsoft's MCP server for Playwright. It reads pages through Playwright's accessibility tree, not screenshots, so it works without a vision model.

By default it starts its own browser with a persistent profile. --isolated keeps the profile in memory, so every session starts clean. To use your own Chrome or Edge, install the Playwright Extension and pass --extension. To drive a remote browser, pass --cdp-endpoint. --browser picks chrome, firefox, webkit or msedge.

Pick it over Browser Harness for test runs that must start clean, for cross-browser checks, or if your team already runs Playwright. On our eval it scored the same and finished slightly faster. Playwright MCP vs Browser Harness.

• Controls: its own browser; your Chrome through an extension; a remote browser over CDP
• License: Apache-2.0
• Cost: free

claude mcp add playwright npx @playwright/mcp@latest

2. Chrome DevTools MCP

The best choice for debugging performance, network requests and console errors.

Chrome DevTools MCP comes from the Chrome DevTools team. It records performance traces, inspects network requests, and reads console messages with source-mapped stack traces. It automates Chrome with Puppeteer and waits for each action's result. --slim cuts it to basic browser tasks.

By default it starts Chrome with a dedicated profile that is kept between runs; --isolated uses a temporary one. --autoConnect connects to a Chrome you started, with access to every open window of its default profile. It officially supports Chrome and Chrome for Testing only.

Google collects usage statistics by default; --no-usage-statistics turns that off. Performance tools may send trace URLs to the Chrome UX Report API; --no-performance-crux turns that off.

Pick it over Browser Harness when you are profiling or debugging your own app. On our eval it scored 61.0, the highest of the four, though within noise. Chrome DevTools MCP vs Browser Harness.

• Controls: its own Chrome profile, or your running Chrome with --autoConnect
• License: Apache-2.0
• Cost: free

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

3. Browser Harness (browser-harness-mcp)

The best choice for coding agents that need your logged-in Chrome, at the lowest token cost.

Browser Harness is our open-source tool for letting a coding agent use your real Chrome. Its MCP server, browser-harness-mcp, connects to your local Chrome over CDP and exposes 23 tools, from browser_click and browser_screenshot to browser_js and raw browser_cdp. Every tool returns JSON.

Most coding agents do not need the MCP server. The setup prompt installs Browser Harness as a CLI and skill with no config to edit, and that is the setup we measured. Use the MCP server for clients that only speak MCP.

On our eval, run as CLI and skill, it tied Playwright MCP and Chrome DevTools MCP on score with half and a third of their median input tokens, and was slower than both. The Browser Harness CLI can also drive Browser Use cloud browsers at $0.02 per browser-hour; the MCP server's docs cover only your local Chrome.

• Controls: your local Chrome; the CLI can also drive Browser Use cloud browsers
• License: MIT
• Cost: free; cloud browsers through the CLI $0.02 per browser-hour

claude mcp add browser-harness uvx --from 'browser-harness[mcp]' browser-harness-mcp

4. Browser Use Cloud MCP

The best choice for handing whole web tasks to cloud browsers, many at once.

The Browser Use Cloud MCP server is hosted at https://api.browser-use.com/v3/mcp. It does not expose clicks. run_session takes a task in plain words and runs a full browser agent on a Browser Use cloud browser, then get_session returns status, output, cost and a live URL. MCP sessions run on API V3, which the docs say trades accuracy for speed and lower cost compared with V4. send_task sends a follow-up to a kept-alive session, and profile_id reuses a saved login. Sessions use a US proxy by default.

Because the agent runs on our side, your coding agent reads the result, not every page snapshot. New projects can run 10 browsers at once. See /mcp for every client config.

• Controls: Browser Use cloud browsers, driven by a hosted agent
• License: hosted service (the open-source browser-use library is MIT)
• Cost: pay as you go: browsers $0.02 per browser-hour, model tokens at 1.2x provider rates, and proxy traffic at $5/GB, with the US proxy on by default (pricing)

claude mcp add -t http -H "Authorization: Bearer YOUR_API_KEY" browser-use https://api.browser-use.com/v3/mcp

The legacy x-browser-use-api-key header also works.

5. Browser Use Open-Source MCP

A good choice for a local agent with low-level tools, using your own model key.

The open-source browser-use library can run as a local MCP server. It launches a local Chrome or Chromium and exposes direct tools for navigation, clicks, typing, tabs, content extraction and screenshots. One more tool, retry_with_browser_use_agent, hands the task to a full browser-use agent when direct control fails. The server needs your own OpenAI or Anthropic API key.

• Controls: a local Chrome or Chromium it launches
• License: MIT
• Cost: free; you pay your model provider

claude mcp add browser-use -- uvx --from 'browser-use[cli]' browser-use --mcp

6. agent-browser

A good choice if you want a fast CLI with an MCP mode and a small default tool set.

agent-browser from Vercel Labs is a native Rust CLI for browser automation. agent-browser mcp starts a stdio MCP server. Its default core profile keeps context small; --tools all exposes every command.

It runs headless Chrome by default, downloaded from Chrome for Testing. --auto-connect attaches to a running Chrome, and --profile reuses a Chrome profile's logins. In our matched eval it scored 56.3 against Browser Harness's 60.4; that gap is not statistically significant. agent-browser vs Browser Harness.

• Controls: its own headless Chrome; your Chrome with --auto-connect or --profile
• License: Apache-2.0
• Cost: free

npm install -g agent-browser && agent-browser install

Then add {"command": "agent-browser", "args": ["mcp"]} to your MCP client config.

7. Browserbase MCP

A good choice for Browserbase customers who want natural-language actions.

The Browserbase MCP server is hosted at https://mcp.browserbase.com/mcp and built on Stagehand. Its tools take plain words: act ("click the login button"), observe and extract, plus navigate and tools to start and end a session. The default model is Gemini 2.5 Flash Lite; other models need your own key.

Browserbase recommends the hosted endpoint. For a local STDIO server, its docs offer the @browserbasehq/mcp npm package (Apache-2.0, version 3.0.0 from March 2026), which still drives Browserbase cloud browsers. The older self-hostable reference repo is archived.

• Controls: Browserbase cloud browsers
• License: hosted service; the npm package is Apache-2.0
• Cost: $0.10 to $0.12 per browser-hour after included hours; Free, $20 or $99 per month (pricing)

claude mcp add --transport http browserbase "https://mcp.browserbase.com/mcp?browserbaseApiKey=YOUR_BROWSERBASE_API_KEY"

8. Kernel MCP

A good choice for agents that write Playwright code against cloud browsers.

The Kernel MCP server is open source and hosted at https://mcp.onkernel.com/mcp, with OAuth sign-in. It manages Kernel cloud browsers and runs Playwright code against them, plus a persistent JavaScript REPL with raw CDP. It also deploys and monitors Kernel apps.

• Controls: Kernel cloud browsers
• License: MIT
• Cost: $0.06 to $0.48 per browser-hour, headless to headful; Free, $30 or $200 per month plus usage (pricing)

claude mcp add --transport http kernel https://mcp.onkernel.com/mcp

Then run /mcp once in Claude Code to sign in.

9. Steel MCP

A good choice when a person has to step in for logins and CAPTCHAs.

The Steel MCP server drives Steel cloud Chrome. steel_scrape reads a page without starting a billed session. For interactive work it reads the page as an accessibility tree and acts by element reference. steel_session_handoff gives the browser to you for a login, CAPTCHA, payment or local file, then you hand it back.

It also runs against a self-hosted steel-browser with STEEL_LOCAL=true, one session at a time. Version 3.0.0 ships from GitHub releases and is not on npm.

• Controls: Steel cloud Chrome, or self-hosted steel-browser
• License: MIT
• Cost: $0.08 to $0.10 per browser-hour; free tier with $30 one-time credits, then $250 per month (pricing)

claude mcp add steel -e STEEL_API_KEY=your-steel-api-key -- npx -y github:steel-dev/steel-mcp-server

10. Hyperbrowser MCP

A good choice for scraping, crawling and extraction from an MCP client.

The Hyperbrowser MCP server is built around whole jobs, not clicks: scrape a page, crawl linked pages, extract structured JSON, or search with Bing. It also runs hosted browser agents, including the Browser Use agent, OpenAI's computer-use agent and Claude computer use. Its last npm release, 1.0.25, was in April 2025.

• Controls: Hyperbrowser cloud browsers
• License: MIT
• Cost: $0.10 per browser-hour, billed in credits; Free plan with 5,000 credits, paid plans up to $100 per month (pricing)

claude mcp add hyperbrowser -e HYPERBROWSER_API_KEY=YOUR_API_KEY -- npx -y hyperbrowser-mcp

11. Browser MCP

A quick way to try MCP in your own Chrome tab, if you accept slow maintenance.

Browser MCP pairs a local MCP server with a Chrome extension. You click Connect in the extension, and the agent works in that tab, logged in as you. It was adapted from Playwright MCP. Its last npm release, 0.1.3, was in April 2025, and its repo says the code cannot yet be built on its own.

• Controls: your Chrome, one connected tab
• License: Apache-2.0
• Cost: free; no account or API key in its setup docs

{ "mcpServers": { "browsermcp": { "command": "npx", "args": ["@browsermcp/mcp@latest"] } } }

Comparison Table

Cloud prices are per browser-hour as listed by each vendor, checked October 15, 2026, before plan fees, included hours and proxy traffic. Everything else was checked against each project's repo or docs on October 17, 2026.

Conclusion

Playwright MCP and Chrome DevTools MCP are the safe defaults: official, free, and as accurate as anything we measured. Pick Playwright MCP for clean test sessions and Chrome DevTools MCP for debugging. For a coding agent working in your own logged-in Chrome, use Browser Harness; in our eval it matched them on score with fewer tokens. For whole tasks on cloud browsers, use the Browser Use Cloud MCP server, or your current cloud vendor's server.

Try one real task in your top two choices and compare the result, the tokens used and the time taken before you commit.