VMGM Software
Start here

Tools

Browser tools for AI agents: which one to use.

There are a dozen ways to let an agent use a browser. Most of the choice comes down to one question: should it use your browser, with your logins, or a clean one of its own?

In short

  • Browser tools for AI agents split by whose browser they drive: your own signed-in Chrome, or a clean browser of their own.
  • Claude in Chrome works in the tab you have open, with your logins, and asks before risky steps; Chrome Bridge does the same from Claude Code in fewer turns.
  • Playwright MCP, agent-browser and Chrome DevTools MCP start a fresh browser, which suits testing and debugging but not sites you are signed into.
  • Browser Use, Stagehand and Skyvern are for automations you run again, and cloud browsers such as Browserbase and Steel run many at once.
  • If a site has an API or a connector, an agent should use that instead of clicking through a browser.

Which one should you use?

  • The site has an API or a connectorUse that. Clicking through a browser is the fallback, not the default.
  • Do something on a site you are signed into, in the tab in front of youClaude in Chrome
  • The same, from Claude Code, in fewer turns and for less moneyChrome Bridge
  • Hand off one whole web task in a single callJev Ultrafast
  • Several tasks with your logins, without taking over your browserego lite
  • A coding agent checking the site it just builtagent-browser or Playwright MCP
  • Why a page is broken or slow: console, network, performanceChrome DevTools MCP
  • How the site behaves in SafariSafari MCP
  • An automation you will run againStagehand, or Skyvern for form-heavy portals
  • Hundreds of runs, or no Chrome on the machineA cloud browser with Browser Use or Stagehand

Tools that use your own browser

Your logins come along, so the agent can work in Gmail, a CRM or a dashboard you are already signed into.

  • Claude in Chrome

    by Anthropic
    Chrome extensionYour signed-in Chrome

    Doing a task on the page in front of you, on a site you are already signed into.

    It works in the tab you already have open, with your logins, and asks before risky steps. You can step in mid-run to solve a 2FA prompt or make a choice.

    Watch out: Each step is a screenshot or page read, then a click, so long tasks take many turns and a lot of tokens.

    Part of a paid Claude plan

  • Chrome Bridge

    by Francesco Sorrentino
    MCP server and Chrome extensionYour signed-in Chrome

    The same logged-in Chrome from Claude Code, in fewer turns and for less money.

    Pages come back as short element names (n1, n2) the agent can click straight away, and one call fills a whole form. In its author’s benchmark against Claude in Chrome, on the same model, a form took 3 turns instead of 14, and it answered 15 of 15 tasks correctly against 12.

    Watch out: A small open-source project from one developer. Read the code before you hand it your Chrome, and do not retry a form submit blindly: it submits twice.

    Free, MIT licence

  • Jev Ultrafast

    by Browser Use
    Python agent, with community MCP serversYour Chrome, or a headless one

    Handing off one whole web task, such as “find a flight”, in a single call.

    A small, fast model (TypeSafe’s Jev) picks each click, so your main model does not spend a turn on every step. Browser Use’s demo runs a Google Flights search in 7.1 seconds.

    Watch out: When a pop-up, CAPTCHA or login blocks it, it stops, and you finish by hand or with a step-by-step tool.

    Free, MIT licence; you pay for the model calls

    Also: jev-ultrafast-mcp, a community MCP server

  • ego lite

    by CitroLabs
    A browser for MacIts own browser, with your logins copied in

    Several agent tasks at once, with your logins, while you keep using the browser yourself.

    Each agent gets its own Space, so it never takes over your tab, and several can run side by side. Its makers say hard tasks finish up to 3.45 times faster than with agent-browser.

    Watch out: The Spaces share one browser profile, so every agent has every login you put in it. Only sign in to accounts you would hand to a new hire.

    Free, MIT licence, no account

    Also: ego lite on GitHub

Tools for testing and debugging a site

A clean browser with no logins, for a coding agent checking the site it just built.

  • agent-browser

    by Vercel Labs
    Command-line toolIts own browser

    A coding agent opening, clicking through and checking a site.

    It returns the page as a short accessibility tree with references, which costs far fewer tokens than screenshots.

    Watch out: It starts without your logins unless you connect it to your own Chrome.

    Free, Apache 2.0 licence

  • Playwright MCP

    by Microsoft
    MCP serverIts own browser

    Tests, CI and anything that needs a fresh browser every run.

    The best-known browser MCP, built on Playwright, which drives Chrome, Firefox and WebKit.

    Watch out: A poor first choice for Gmail, Notion or an internal dashboard: it is not signed in unless you save a login state for it first.

    Free, Apache 2.0 licence

  • Chrome DevTools MCP

    by Google’s Chrome team
    MCP serverIts own Chrome

    Finding out why a page is broken or slow.

    It gives the agent what Chrome’s DevTools show you: the console, network requests and performance traces.

    Watch out: Built for debugging, not for doing tasks on sites you are signed into.

    Free, Apache 2.0 licence

  • Safari MCP

    by Apple’s WebKit team
    Built into SafariSafari on your Mac

    Checking a site in Safari, where Chrome-only testing misses bugs.

    It is part of Safari, so there is nothing to install. The agent can open pages, take screenshots and read the console and network, in real Safari.

    Watch out: It needs Safari 27 or Safari Technology Preview, and it is Safari on the Mac, not on an iPhone.

    Free, part of Safari

Tools for automations you keep

For a web task you will run again and again, written as code.

  • Browser Use

    by Browser Use
    Python libraryIts own browser, or theirs in the cloud

    An agent that takes a task and returns a result, on sites it has never seen.

    The most widely used open-source browser agent. You give it a task in plain English and it works out the clicks.

    Watch out: It runs its own browser, so it is not a way to use the tab you have open.

    Free, MIT licence; you pay for the model, and for their cloud browsers if you use them

  • Stagehand

    by Browserbase
    TypeScript libraryIts own browser, or Browserbase’s

    Automations you maintain as code, with AI only at the steps that break.

    You write normal code and call act, extract or observe where a page is unpredictable. It caches actions, so repeat runs are faster.

    Watch out: You are writing and maintaining code, not just asking.

    Free, MIT licence

  • Skyvern

    by Skyvern
    Open-source app, or their cloudIts own browser

    Form-heavy portals: insurance, government and supplier sites, with logins and 2FA.

    It reads the page by sight, so it copes with sites whose code changes or makes no sense.

    Watch out: The AGPL licence has rules for hosting it yourself, so check them before you build on it.

    Free under AGPL 3.0, or their paid cloud

Browsers in the cloud

Browsers on someone else’s servers, for running many at once. Not a way to use the tab you have open.

  • Cloud browsers

    by Browserbase, Steel, Kernel, Lexmount and Browser Use
    Hosted serviceBrowsers on their servers

    Many browsers at once, long sessions, or no Chrome on your machine.

    You rent browsers by the session and point Browser Use, Stagehand or your own code at them. Most offer live view, recordings and a way for a person to take over.

    Watch out: Paid by usage, and your logins do not come along.

    Paid, usually by the browser-hour

    Also: Steel, Kernel, Lexmount, Browser Use Cloud

A prompt that makes your agent pick

Paste this into your agent’s instructions, such as CLAUDE.md, and cut the tools you have not installed.

Browser tool rules
Use these rules to pick a browser tool for each web task, and tell me which one you picked and why, in one sentence, before you start.

1. If the site has an API or a connector, use it. Clicking through a browser is the fallback.
2. Needs my logged-in session, on the tab in front of me: Claude in Chrome.
3. Same logged-in Chrome, but many steps: Chrome Bridge, for element refs and one-call form fill.
4. One self-contained goal where every click does not need you: Jev Ultrafast.
5. Several tasks with my logins while I keep using the browser: ego lite.
6. Testing or checking a site: agent-browser or Playwright MCP. Debugging console, network or performance: Chrome DevTools MCP.
7. An automation we will run again: Stagehand, or Skyvern for form-heavy portals.
8. Many runs at once, or no local Chrome: a cloud browser.

While you work:
- Prefer accessibility snapshots and element refs to screenshots. Do not paste whole pages into the conversation.
- Batch actions: fill a form in one call where the tool allows it.
- Stop and hand it to me at a login, 2FA, CAPTCHA, payment or anything that cannot be undone.
- Do not take over my foreground tab when another tool can work in its own window or Space.
- Afterwards, tell me which tool you used, what worked, and whether to use a different one next time.

Frequently asked questions

What is the best browser MCP?

It depends on whose browser the agent should use. For your own signed-in Chrome, Claude in Chrome, or Chrome Bridge for fewer turns. For testing a site in a clean browser, Playwright MCP or agent-browser. For debugging, Chrome DevTools MCP.

Why not just use Playwright MCP for everything?

It starts a fresh browser with no logins. That is right for tests and wrong for “update this in our CRM”, where you would have to set up and keep a saved login for it.

Is it safe to let an agent use my logged-in browser?

Only with care. A page can contain text written to trick the agent. Keep a tool that asks before risky steps for your real accounts, stop it at logins, payments and anything that cannot be undone, and do not let it run on its own on pages you do not trust.

Why are screenshots slower than element refs?

A screenshot costs many tokens and the agent has to find each button in it, often one click per turn. A short list of elements with names like n1 and n2 lets it act straight away, and several actions can go in one call.

Which of these work with Cursor and Codex?

Every MCP server here works with any MCP client, including Claude Code, Cursor and Codex. Claude in Chrome is Anthropic’s, and ego lite connects through a skill that supports Claude Code, Codex and Cursor.

None of these makers pay us to list them. Checked against each maker’s own pages on ; speed figures are the makers’ own.

We set up browser tools for engineering teams, with the rules that make an agent use them safely.

Talk to us about your team