MCP servers for browser automation in 2026: Playwright, Chrome DevTools, Browserbase and pilotfish
By the pilotfish team · Last updated
Four kinds of browser MCP server are worth knowing in 2026. Playwright MCP and Chrome DevTools MCP run Chrome on your machine and suit testing and debugging. Browserbase runs browsers in its cloud for your agent to drive. pilotfish has two: an open server that lets your agent drive a local browser with fewer tokens, and a managed one that takes the whole task, signed in and past bot checks. Disclosure: we build pilotfish; every competitor detail below comes from its own README, linked at the end.
Key takeaways
- Testing and debugging your own app: Playwright MCP or Chrome DevTools MCP. Both run Chrome on your machine and start with
npx. - Cloud browsers you drive yourself: Browserbase MCP, hosted at
https://mcp.browserbase.com/mcp. - Your agent drives a local browser, with fewer tokens: the open pilotfish MCP,
npx -y @gaiventures/pilotfish-mcp. Free, MIT, no account. - Hand over the whole task, signed in, past bot checks: pilotfish managed, at
https://pilotfish-be.gai.ventures/mcp. Pay as you go from $5. - A site that offers its own tools to agents: that is WebMCP, which works inside the page. pilotfish uses those tools when a page has them.
What is a browser MCP server?
An MCP server gives an AI app (Claude, Cursor, VS Code, Codex CLI and others) a set of tools over the Model Context Protocol. A browser MCP server's tools are a browser's: open a page, read it, click, type, scroll, take a screenshot. The AI app's model decides each step, and the server carries it out and says what the page shows now.
They differ in three ways that matter more than the tool list: where the browser runs (your machine or someone's cloud), whose model picks the steps, and what happens when a site asks for a sign-in, a one-time code or a CAPTCHA.
Where does the browser run, and how do you connect?
| Server | Publisher | Browser runs | Connect with | Account or keys |
|---|---|---|---|---|
| Playwright MCP | Microsoft | Your machine | npx @playwright/mcp@latest (stdio) | None for the server |
| Chrome DevTools MCP | Chrome DevTools team | Your machine, in Chrome | npx -y chrome-devtools-mcp@latest (stdio) | None for the server |
| Browserbase MCP | Browserbase | Browserbase's cloud | Hosted https://mcp.browserbase.com/mcp, or self-hosted npx @browserbasehq/mcp | Browserbase account; self-hosted also needs BROWSERBASE_API_KEY, BROWSERBASE_PROJECT_ID and a model key |
| pilotfish (open) | gAI Ventures | Your machine, in Chrome, Edge or Chromium | npx -y @gaiventures/pilotfish-mcp (stdio), or its own HTTP server | None: no account, no API key |
| pilotfish managed | gAI Ventures | pilotfish's cloud | https://pilotfish-be.gai.ventures/mcp (Streamable HTTP), or the npx -y @gaiventures/pilotfish-mcp-managed relay | Sign in with OAuth, or an API key; credits from $5 |
A local server is faster to try and sees what you see: your own browser profile, your own sign-ins. A cloud server keeps the browser off your machine, keeps running when your laptop sleeps, and can run many at once.
Playwright MCP
Microsoft's server reads pages through Playwright's accessibility tree rather than screenshots, which keeps it fast and model-agnostic. It opens a visible browser by default; --headless hides it and --isolated keeps the profile in memory instead of on disk. --allowed-origins and --blocked-origins limit which sites the browser requests, but its README is explicit that they are not a security boundary and do not affect redirects. It is the natural pick for testing your own web app.
{
"mcpServers": {
"playwright": { "command": "npx", "args": ["@playwright/mcp@latest", "--headless", "--isolated"] }
}
}
Chrome DevTools MCP
The Chrome DevTools team's server is built for debugging: console messages with source-mapped stack traces, network requests, and performance traces with insights. It takes --headless, --isolated and --slim (fewer tools). Its README asks you not to share sensitive information with it, notes that performance tools may send trace URLs to Google's CrUX API (--no-performance-crux turns that off), and that Google collects usage statistics.
Browserbase MCP
Browserbase runs the browsers in its cloud, and its server uses Stagehand, Browserbase's own agent library, behind six tools. The hosted server at https://mcp.browserbase.com/mcp is the setup Browserbase recommends, and its README says the hosted version covers the model's cost (Gemini). Self-hosted over stdio, you supply your Browserbase API key and project id, and a model key. You pay Browserbase for browser time; your agent still decides what to do on each page.
pilotfish open MCP
Our open server drives Chrome, Edge or Chromium on your machine with your agent's own model, like Playwright MCP, and is built to spend fewer of that model's tokens. After the first look at a page, the agent gets only the lines that changed, and browser_query pulls out just the values it needs. Paying, placing an order, sending and deleting wait for your yes. When a page needs a person (a sign-in, a one-time code, a CAPTCHA), the result says NEEDS A PERSON, you deal with it in the browser window, and the agent calls browser_resume. PILOTFISH_ALLOWED_HOSTS limits which hosts it opens. When a page offers its own tools through WebMCP, the agent can call them with browser_page_tool. It is free, MIT licensed and needs no account.
{
"mcpServers": {
"pilotfish": { "command": "npx", "args": ["-y", "@gaiventures/pilotfish-mcp"] }
}
}
To share one browser between several clients, start it as an HTTP server with npx -y @gaiventures/pilotfish-mcp --http --port 8931 and point them at http://127.0.0.1:8931/mcp.
pilotfish managed MCP
The managed server takes a different job: your agent hands over the whole web task in words, and pilotfish does it in its own cloud browser with its own models, then returns the answer and any files. Your agent makes a call or two per task instead of one per click. It signs in with your saved logins (kept encrypted, never shown to a model), reads one-time codes from a connected inbox, runs stealth and residential IPs, solves common CAPTCHAs, and hands you the live browser when a step really needs you. Payments, orders and messages wait for your approval unless the task allows them up front.
{
"mcpServers": {
"pilotfish": { "url": "https://pilotfish-be.gai.ventures/mcp" }
}
}
A client that supports MCP sign-in connects with that alone; others add an Authorization: Bearer header with an API key. Clients that only start local servers use npx -y @gaiventures/pilotfish-mcp-managed, which forwards to the same URL. It is pay as you go from $5 of credits, at the rates on the pricing page.
Which browser MCP server should you use?
- You are testing or debugging a site you build: Playwright MCP for flows, Chrome DevTools MCP for console, network and performance.
- You want cloud browsers and will script them: Browserbase.
- Your agent should do the steps itself, on your machine, without a big token bill: the open pilotfish MCP.
- You want the task done, including sign-ins, codes and bot checks, with an answer you can check: pilotfish managed. Our benchmarks show what that costs per finished task.
Browser MCP servers and WebMCP
A browser MCP server works from outside the page, the way a person would. WebMCP works from inside it: the website registers its own tools for agents, so an agent can call "search flights" instead of finding the right boxes to fill. They are complements, not rivals, and we wrote up how they fit together.
Before you connect one
- A browser server can read anything its browser shows, including pages you are signed in to. Use a separate profile, or a cloud session, for anything you would not hand to an assistant.
- Limit where it can go where the server lets you, and remember that an allow-list of origins is not always a security boundary.
- Prefer a server that asks before irreversible steps (paying, sending, deleting) over one that just clicks.
Frequently asked questions
What is the best MCP server for browser automation?
It depends on the job. Playwright MCP and Chrome DevTools MCP are the usual picks for testing and debugging your own site on your machine. Browserbase gives you cloud browsers to drive. The open pilotfish MCP lets your agent drive a local browser with fewer tokens, and pilotfish managed takes the whole task in its cloud browser, signed in and past bot checks.
Which browser MCP servers run in the cloud?
Browserbase's hosted server (https://mcp.browserbase.com/mcp) and pilotfish managed (https://pilotfish-be.gai.ventures/mcp). Playwright MCP, Chrome DevTools MCP and the open pilotfish MCP run the browser on your own machine.
Do I need an API key for a browser MCP server?
Not for Playwright MCP, Chrome DevTools MCP or the open pilotfish MCP. Browserbase needs a Browserbase account, plus API key, project id and a model key when you host it yourself. pilotfish managed signs you in with OAuth or takes an API key.
Can a browser MCP server sign in to websites and handle 2FA?
A local server uses the browser you sign in to yourself. The open pilotfish MCP stops with NEEDS A PERSON at a sign-in, code or CAPTCHA and carries on after you finish it. pilotfish managed signs in with your saved logins, reads emailed codes from a connected inbox and solves common CAPTCHAs, and hands you the live browser when it cannot.
Is a browser MCP server the same as WebMCP?
No. A browser MCP server drives the browser from outside the page. WebMCP lets a website publish its own tools inside the page for agents to call. pilotfish supports both: it drives any site, and calls a page's WebMCP tools when the page offers them.
Is the pilotfish MCP free?
The open MCP is free and MIT licensed, with no account. The managed MCP is pay as you go from $5 of credits: browser time, residential traffic and model tokens, each at a published rate.
Sources
- Playwright MCP README (@playwright/mcp 0.0.85, checked 17 Nov 2026)
- Chrome DevTools MCP README (chrome-devtools-mcp 1.13.0, checked 17 Nov 2026)
- Browserbase MCP README (@browserbasehq/mcp 3.0.0, checked 17 Nov 2026)
- pilotfish MCP repository (@gaiventures/pilotfish-mcp 0.1.58 and @gaiventures/pilotfish-mcp-managed 0.1.45)
- pilotfish open MCP docs, managed MCP docs and pricing