浏览器自动化最佳 MCP 服务器选择 (2026)

Every browser MCP server does one of three jobs: it lets your coding agent click through pages itself, it gives your agent browsers on someone else's machines, or it takes a goal and runs a whole browser agent for you. Figure out which job you have before you compare servers, because the best server for one job is usually wrong for the other two. My picks: Browser Harness for step-by-step work in your own Chrome, the Browser Harness CLI on our cloud browsers for step-by-step work in the cloud (Kernel if your client only speaks MCP, Steel if a person has to step in for logins), and the Browser Use Cloud MCP server when you want to hand off the whole task and only read the answer.

Disclosure: we make Browser Use and Browser Harness, so they come first. Competitor facts come from each project's own repo, docs and pricing page, checked October 18, 2026.

The Three Kinds

A browser MCP server gives Claude Code, Cursor or Codex browser tools. The servers split on two things: who decides each click, and where the browser runs.

• Your agent drives every step, on your machine. Clicks, typing, page reads, screenshots. Your coding agent plans each action, and every page it reads lands in its context. Most control, most tokens. All free: Browser Harness, Playwright MCP, Chrome DevTools MCP, agent-browser and Browser MCP.
• Cloud browsers. The browser runs on a vendor's machines: many at once, proxies and CAPTCHA handling, nothing on your laptop, and none of your local logins. You pay per browser-hour. Browser Use Cloud, Browserbase, Kernel, Steel and Hyperbrowser. Their tools come in different shapes: Kernel and Steel give your agent raw steps, Browserbase takes one step at a time in plain words, and Browser Use Cloud and Hyperbrowser take whole jobs.
• Give it a goal. You send a task in plain words, a second agent with its own model does the clicking, and your coding agent only reads the result. The Browser Use Cloud MCP server works this way. The Browser Use open-source server has a full-agent tool next to its step tools, Hyperbrowser has hosted agents, and agent-browser has an opt-in chat tool.

Browser Use Cloud and Hyperbrowser sit in two kinds at once: cloud browsers, and whole tasks.

Four Questions Before You Pick

• Which browser? Logged-in work on your own accounts needs your Chrome. Tests want a clean, isolated profile. Parallel or unattended work wants cloud browsers.
• Low-level tools or a whole agent? Step tools give your agent full control and make it read every page. Task tools keep raw pages out of its context, but you trust the other agent's work.
• Context cost. Every page snapshot your agent reads costs input tokens, and on long tasks that is most of the bill. Playwright's own README says coding agents "might benefit from using the CLI+SKILLS instead", because CLI calls "avoid loading large tool schemas and verbose accessibility trees into the model context". That's the reason Browser Harness defaults to a CLI and a skill.
• Cost and license. Local servers are free and open source, and your model tokens are the bill. Cloud servers add browser time, proxy traffic and sometimes a monthly plan.

My Pick for Each Kind

Your Agent Drives, on Your Machine: Browser Harness

Browser Harness is our open-source tool for letting a coding agent use your real Chrome, logged in as you. It talks to Chrome directly over the Chrome DevTools Protocol (CDP), with no Playwright and no extension. A background daemon holds one connection to your Chrome, and the tab it controls gets a horse emoji in its title. Its skill has the agent open its own tab first, so it doesn't take over the one you're looking at. There's no snapshot tool: the agent reads the page through JavaScript, screenshots or raw CDP calls like Accessibility.getFullAXTree, and clicks by coordinates, which go through iframes and shadow DOM.

The default setup is a CLI plus a skill. You paste one setup prompt into your agent and it installs everything, no config to edit. If your client only speaks MCP, browser-harness-mcp exposes the same helpers as 23 tools, from browser_click to raw browser_cdp:

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

MIT and free in your own Chrome, no API key.

The catches. The agent sees your whole logged-in Chrome, not a sandbox. On Chrome 144+ you tick remote debugging once in chrome://inspect and click Allow per connection (browser-harness mac-approve clicks it on macOS). One daemon controls one tab, so two agents acting at once can race. There's no documented clean-profile mode for the MCP server and no Firefox or WebKit. In our eval it was slower than Playwright MCP and Chrome DevTools MCP.

Two free servers beat us at specific jobs:

• Playwright MCP for tests. --isolated starts every session clean, --browser runs Firefox and WebKit, and it reads pages as accessibility snapshots with element refs, so it needs no vision model. It was the fastest tool in our eval, with about twice our tokens. If your team already lives in Playwright, use it.
• Chrome DevTools MCP for debugging your own app. It's from the Chrome DevTools team and records performance traces, runs Lighthouse audits, inspects network requests and reads console messages with source-mapped stack traces. Browser Harness can reach that data through raw CDP, but this has purpose-built tools for it. It had the highest raw score in our eval (within noise) and used about three times our tokens.

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

Cloud Browsers: Browser Harness on Our Cloud, or Kernel Over MCP

Our own cloud MCP server runs on cloud browsers, but it takes whole tasks, not clicks (next section). For step-by-step work in the cloud I use the Browser Harness CLI on our cloud browsers, $0.02 per browser-hour. Its MCP server only covers your local Chrome, so if your client only speaks MCP, I'd pick Kernel.

Kernel runs every browser as Chromium in its own unikernel VM. execute_playwright_code runs Playwright code inside that VM, next to the browser, and browser_repl is a persistent Node.js runtime there with raw CDP; the README calls it "unrestricted VM code execution, not a sandbox." You pay per second of active runtime, $0.06 per hour headless and $0.48 headful, and a browser stops billing 5 seconds after the last activity. The gotcha: the default inactivity timeout is 60 seconds, so an agent that pauses longer loses the browser unless it sets timeout_seconds. MIT, hosted with OAuth sign-in:

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

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

Steel reads the page as an accessibility tree with @eN refs and acts by ref. Its standout is steel_session_handoff: a person takes over for a login, CAPTCHA, payment or local file, then hands the browser back. Our server has a live URL you can watch, but no handoff. It's MIT and not on npm, so npx github: builds it on first run.

If you already pay Browserbase, its server is the shortest path, and your agent keeps step control without reading raw pages. Plans, proxy rates, stealth and every tool for all five are in the cloud browser round-up.

Give It a Goal: Browser Use Cloud MCP

The Browser Use Cloud MCP server is hosted, so there's no local server to run. run_session takes a task in plain words and runs a full Browser Use agent on one of our cloud browsers, then returns right away. Your agent polls get_session for status, output, a cost breakdown and a live URL. Each poll returns a status, not a page, which is the whole point: your coding agent reads the task, a few checks and the answer. output_schema gets you JSON back, profile_id starts the task already logged in, and keep_alive plus send_task handle follow-ups in the same browser.

You pay $0.02 per browser-hour, model tokens at provider cost plus 20%, and residential proxy traffic at $5/GB, with the US proxy on by default (pricing). New projects can run 10 at once, more as you spend.

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

It runs API V3, which our docs say is faster and cheaper than V4 but less accurate. It also wasn't in the eval. Every client config is on /mcp.

For free and local, the Browser Use open-source server launches a local Chrome with step tools, plus retry_with_browser_use_agent, which hands the task to a full browser-use agent when direct control fails. That agent runs on your own OpenAI API key. Models, cost and how to watch a run for every goal-taking server are in the browser agent round-up.

What We Measured

We ran four step-by-step tools on 100 of the 200 tasks in BU Bench V2.1, one run per task, with the same agent (Codex with GPT-6 Luna), the same fresh Browser Use cloud browser and the same judge. Browser Harness ran as its CLI and skill, not the MCP server.

Scores: Chrome DevTools MCP 61.0, Browser Harness 60.4, Playwright MCP 60.0, agent-browser 56.3. 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.

Where they differed was tokens. Median input tokens per task: Browser Harness 2.1M, agent-browser 3.3M, Playwright MCP 4.4M, Chrome DevTools MCP 6.7M. So Browser Harness had the lowest median agent cost per task, $0.042 against $0.066 for Playwright MCP and $0.097 for Chrome DevTools MCP. By median time Playwright MCP was fastest at 8.3 minutes, then Chrome DevTools MCP at 8.7, Browser Harness at 9.0 and agent-browser at 11.3.

Limits: one run per task, half the benchmark, one agent and one model, and a fresh cloud browser every time, not your logged-in Chrome. The data file doesn't record whether agent-browser ran as its CLI or its MCP server. Versions and every per-task score are in the data file, and the full table with tool calls and solve rates is in the step-by-step round-up.

All Eleven Servers

The Deep Dives

• The step-by-step round-up covers the five tools of the first kind (Browser Harness, Playwright MCP, Chrome DevTools MCP, agent-browser and Browser MCP): how each works under the hood, install, license, gotchas, where each beats us, and the full eval.
• The cloud browser round-up covers Browser Use Cloud, Browserbase, Kernel, Steel and Hyperbrowser: prices, stealth, concurrency and tools.
• The browser agent round-up covers Browser Use Cloud, the Browser Use open-source server, Hyperbrowser and agent-browser's chat tool: models, cost and how to watch a run.

Cloud prices are per browser-hour as listed by each vendor on October 18, 2026, before plan fees, included hours and proxy traffic.