Recommended Free Tools
Playwright MCP connects an MCP-compatible AI client to a Playwright browser. You install the local server with npx @playwright/mcp@latest, register it in your client, then ask the assistant to open pages, inspect their accessibility structure, and interact with links, fields, buttons, and other controls. The server is an integration layer, not a separate browser product. This guide explains the setup, first tasks, configuration choices, security boundaries, troubleshooting, and the managed remote alternative.
What Playwright MCP server tools do
The Microsoft Playwright MCP project describes itself as an MCP server that provides browser automation through Playwright. An MCP client (such as an AI desktop app, coding assistant, or another compatible host) starts the server and exposes its browser capabilities to the model.
Instead of requiring a vision model to interpret a screenshot, the server represents a page with a structured accessibility snapshot. The assistant can use that representation to identify headings, links, text boxes, buttons, checkboxes, and other accessible elements, then request navigation or interaction. This often gives the model a more explicit target than raw pixels, although success still depends on the website, authentication flow, JavaScript behavior, permissions, and the package version you run.
The normal interaction loop
- The MCP client launches the Playwright MCP server.
- You ask the assistant to navigate to a URL or perform a browser task.
- The server returns a structured page representation, commonly an accessibility snapshot.
- The assistant chooses an element from that structure and requests an action such as clicking, typing, selecting, or navigating.
- The server reports the resulting page state, allowing the assistant to continue.
Documented examples include opening a demo shopping page and adding items, filling a form, taking a screenshot, running Playwright code, and mocking an API. These are task examples rather than guarantees that every site or control will behave identically.
#1 Best Overall
Prerequisites and version-sensitive requirements
- Node.js: The surfaced Playwright documentation is inconsistent: one getting-started page states Node.js 20 or newer, while the repository README states 18 or newer. Check the current package requirements before installation; use Node 20 or newer when you can to avoid the older-runtime edge case.
- An MCP client: You need a client that can register and start an MCP server. The exact menu labels, configuration file location, and JSON or TOML format vary by client and can change.
- Permission to automate: Use accounts, sites, and data you are authorized to access. Browser automation can submit forms, change records, or expose private content.
Install and connect the local server
1. Verify Node.js
Run node --version. If your client or the current package documents a newer minimum, upgrade Node.js before continuing. Keeping npm current also reduces package-resolution problems.
2. Register the server in your MCP client
Create a server entry named playwright using the command below:
npx @playwright/mcp@latestMany clients represent that entry with a JSON object similar to:
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest"]
}
}
}Treat this as a pattern, not a universal file format. Follow your client’s current MCP setup page for the exact path, key names, and restart procedure. Some clients provide a command-line registration command; others use JSON, TOML, or a graphical “add server” screen.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitches3. Restart and approve the connection
Restart or reload the client so it starts the server. If the client asks whether to trust or enable the server, review the command and working environment first. A server with browser access can reach any site your profile can reach.
4. Run a harmless first request
Ask the assistant to open a public demo page and describe the available controls from its accessibility snapshot. Then ask it to perform one reversible action, such as entering text into a test form without submitting it. This confirms that the server starts, the browser launches, and the client can pass tool results back to the model.
Browser behavior you should choose deliberately
Browser engine
The project documents browser selection, including Chromium-family choices and Firefox/WebKit-related options. Accepted value names are version-sensitive, so inspect the current release documentation before adding a browser flag. Choose the engine that matches the site you are testing; do not assume a Chromium-only result represents Firefox or WebKit behavior.
Headed versus headless
The setup material documents headed mode as the default. Use --headless when you do not need a visible window, such as in CI or a server process. Headed mode is useful while diagnosing selectors, login prompts, permission dialogs, and unexpected navigation because you can watch the browser.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Persistent and isolated profiles
A persistent profile keeps cookies and login state between sessions. An isolated context starts fresh and loses in-memory state when it closes unless you explicitly provide storage state. Persistent profiles are convenient for development but contain sensitive data; never point automation at a personal profile that holds unrelated accounts.
Connecting to an existing browser
The repository documents connecting through Chrome DevTools Protocol (CDP) or an extension. Extension mode can reuse the existing browser profile and its logged-in session. Treat that session as sensitive: an AI agent operating through it may be able to read or modify everything the profile can access. Prefer a dedicated profile with least-privilege accounts.
Timeouts, output, and capability settings
Configuration support includes timeouts, browser capabilities, output controls, and configuration files. Exact option names and syntax can change with package releases. Start with defaults, then add one setting at a time while consulting the release documentation. Keep navigation and action timeouts long enough for the application’s real load time, but finite so a stalled page does not hold a worker forever.
How to use Playwright MCP in practice
Navigate and inspect
Use a precise request such as: “Open https://example.test, return the page title and list the controls in the main content.” The assistant should first navigate, then inspect the returned accessibility structure. If the page is still loading, ask it to wait and inspect again rather than guessing selectors.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Interact by meaning, not coordinates
Ask for actions tied to accessible names and roles: “Fill the Email textbox with …,” “Click the Submit button,” or “Select the United States option.” This is more robust than telling the model to click a screen coordinate. If a control has no useful accessible name, improve the application’s labels or ask the assistant to inspect the surrounding structure before acting.
Handle multi-step workflows
- State the goal and the boundary, for example, “Prepare the form but do not submit.”
- Have the assistant inspect the current page before each consequential action.
- After navigation, wait for a meaningful condition such as a visible heading or a known form field.
- Confirm the resulting state before proceeding to the next step.
- Require an explicit approval before purchases, account changes, messages, or deletions.
Use screenshots and code execution carefully
The documented quick reference includes screenshot requests and running Playwright code. A screenshot is useful for visual verification, but the accessibility snapshot remains the primary structured representation described by the project. Code execution can be powerful for unusual interactions; restrict it to trusted tasks and review generated code before allowing it to run against production accounts.
Local MCP or managed remote MCP?
| Decision | Local Playwright MCP | Microsoft Playwright Workspaces remote MCP |
|---|---|---|
| Execution | Your machine or controlled runtime | Microsoft-managed cloud browser over streamable HTTP |
| Setup | Node.js, package launch, and client registration | Azure subscription, enabled Playwright workspace, endpoint, and authentication |
| State model | Local profiles, cookies, isolation, or supplied storage state | Workspace-managed sessions and access controls |
| Availability | Package availability depends on your environment | Microsoft’s quickstart labels the remote feature preview |
| Cost note | The cited setup material does not state a local package charge | Sessions consume workspace capacity and might incur charges |
The remote service is optional; it is not required for the local server. Microsoft’s remote quickstart recommends Microsoft Entra ID for authentication. Treat access tokens like passwords: do not commit them to source control or place them in prompts, agent instructions, or logs. During evaluation, require approval for tool calls so you can observe what the agent is doing.
Reliability, security, and operational practices
- Use test data first: Validate flows on a staging site or disposable account before allowing production actions.
- Separate profiles: Keep a dedicated persistent profile for automation; do not share personal cookies with an agent.
- Limit permissions: Give the account only the roles needed for the task.
- Set bounded waits: Use finite navigation and action timeouts and retry only idempotent operations.
- Make approval explicit: Pause before irreversible submissions, purchases, invitations, or deletion.
- Log safely: Redact passwords, tokens, personal data, and page contents that should not leave the controlled environment.
- Pin deliberately:
@latestis convenient for trying the current release, but production deployments should test updates before adopting them.
Troubleshooting common failures
The client cannot start the server
Check that Node.js is installed and on the client’s PATH. Run npx @playwright/mcp@latest in a terminal to expose download or permission errors, then verify the client’s command and argument syntax. Restart the client after changing its configuration.
The browser opens, but the assistant sees little or nothing
The page may still be loading, may require a login, or may expose weak accessibility labels. Ask for a fresh snapshot after waiting for a specific heading or control. Confirm that the account is signed in and that the site is not presenting a consent, bot-check, or permission dialog.
A click or fill action targets the wrong element
Ask the assistant to list matching roles and accessible names, then make the request more specific. Improve duplicate labels in the application where possible. Avoid coordinate-based instructions and verify the page state after every action.
Login state disappears
You are probably using an isolated context or a new profile. Use a dedicated persistent profile or provide supported storage state, and check that the profile directory is writable. Never copy a personal browser profile into an automation environment.
CDP or extension attachment fails
Confirm that the browser was started with the expected debugging endpoint or that the extension is installed and enabled. Check port access and version compatibility. If attachment is unreliable, start a clean browser context instead of reusing an existing session.
Free tools Windows power users keep installed
One-click scans. No signup required.
Remote sessions are unavailable or unexpectedly costly
For Playwright Workspaces remote MCP, verify the workspace is enabled, the endpoint and token are valid, and that capacity is available. The preview documentation warns that sessions may consume capacity and incur charges; review workspace limits before running unattended jobs.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Or skip the browser setup
If your goal is a reliable website image or PDF rather than interactive browser control, ScreenshotNeo provides a single-call screenshot API and an MCP server. It accepts cookie and consent banners before capture and removes more than 60 known consent platforms, newsletter popups, and chat widgets; each step can be disabled. Bot checks, blank pages, failed loads, timeouts, and cache hits are not billed, and response headers identify the page verdict and billing result. Its MCP tools—take_screenshot, get_page_info, and capture_pdf—work with Claude, Cursor, and other MCP clients.
cURL
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpPython
import requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)Node.js
const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);See the ScreenshotNeo documentation for parameters. It supports full-page captures with lazy images, CSS-selector element shots, dark mode, 12 device presets and custom viewports, retina scale, PDF paper settings and page ranges, HTML/CSS rendering, custom JavaScript and CSS, pre-capture clicks, hidden selectors, selector or network-idle waits, request blocking, headers, cookies, user agents, authorization, timezone, geolocation, transparent backgrounds, resizing, configurable-TTL caching, signed image links, asynchronous webhooks, bulk capture of up to 100 URLs per call, usage data, OpenAPI, and familiar parameter names for easier migration. Every plan includes every feature: 1,000 shots monthly free with no card; paid plans start at $5 for 3,000 shots. Create a free ScreenshotNeo account to try it.
FAQ
Is Playwright MCP itself a browser?
No. It is an MCP server that controls a Playwright-backed browser supplied by your environment.
Do I need Microsoft Azure to use it?
No. Azure is associated with Microsoft’s separate managed Playwright Workspaces remote MCP option. The local package runs from your environment.
Can I rely on a complete fixed list of tool names?
No. Tool availability and schemas can vary by package version and enabled capabilities. Check the running client’s exposed tools and current documentation.
Frequently Asked Questions
Can Playwright MCP bypass a CAPTCHA or bot challenge?
The documented material does not promise that. Treat bot checks and CAPTCHA pages as site-specific failures requiring an authorized human or an approved test path.
Should I use headed mode in CI?
Usually no; use the documented --headless option for unattended runs, and switch to headed mode while diagnosing a failure.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →How should I protect a persistent profile?
Use a dedicated automation profile, least-privilege accounts, restricted filesystem access, and never include its cookies or tokens in logs or prompts.
The Bottom Line
Use Playwright MCP when an AI client needs interactive, Playwright-powered browser control. Start locally with npx @playwright/mcp@latest, choose profile and browser settings deliberately, and require approval for consequential actions. For clean, non-interactive screenshots or PDFs, ScreenshotNeo avoids browser setup and bills only successful clean captures.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
