Microsoft's browser-automation MCP server, from the microsoft/playwright-mcp repository. Install commands, configuration options and the tools it exposes, read from the project README on 2 September 2026.
Interactive examples · No account is connected on this page
Updated
Playwright MCP is a Model Context Protocol server from Microsoft that gives an MCP client browser automation through Playwright. It exposes tools such as browser_navigate, browser_click and browser_snapshot, and it reads pages as structured accessibility snapshots rather than screenshots, so no vision model is needed. The README lists Node.js 18 or newer as the requirement, and most clients add it with one npx command.
Pick an example to see which Playwright MCP tools a task would call and what the result looks like. Examples are illustrative; nothing on this page connects to Playwright.
Playwright MCP needs Node.js 18 or newer and nothing else. Claude Code and VS Code add it with a one-line command, and every other client takes the standard JSON block that runs npx @playwright/mcp@latest. Command-line flags choose the browser, headless mode and device emulation. Start it with --port when several clients must share one browser.
The server ships as the npm package @playwright/mcp, and the README names a single hard requirement: Node.js 18 or newer. Everything else comes from the MCP client you already use. The README names VS Code, Cursor, Windsurf, Claude Desktop, Goose, Grok and Junie as supported hosts, and says any other MCP client works the same way, because the wire format is the protocol rather than an integration.
node --versionTwo clients publish a one-line install command in the README. Claude Code registers the server by name and command; VS Code takes the same object as JSON on the command line.
claude mcp add playwright npx @playwright/mcp@latest
code --add-mcp '{"name":"playwright","command":"npx","args":["@playwright/mcp@latest"]}'Every other client takes the standard configuration block. Windsurf and Claude Desktop are documented as using exactly this, with no client-specific fields.
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": [
"@playwright/mcp@latest"
]
}
}
}Command-line options go in the args array after the package name. --browser accepts chrome, firefox, webkit and msedge. --headless runs without a window; the README notes the server is headed by default. --device emulates a device such as an iPhone 15, and --config points at a configuration file.
npx @playwright/mcp@latest --browser msedge --headless --device "iPhone 15"The README covers a case that trips people up: running a headed browser on a machine without a display, or from worker processes. The fix is to start the server where a DISPLAY exists and pass --port to enable the HTTP transport, then point clients at the URL instead of a command.
npx @playwright/mcp@latest --port 8931
{
"mcpServers": {
"playwright": {
"url": "http://localhost:8931/mcp"
}
}
}Official documentation: Read the README on GitHub.
The README documents three ways to run Playwright MCP. npx over stdio is the default, started and stopped by the client. HTTP mode with --port runs a long-lived server that clients reach by URL, which fixes headed browsers on machines without a display. Docker runs the mcr.microsoft.com/playwright/mcp image for containerised setups.
| What differs | npx (stdio) | HTTP (--port) | Docker |
|---|---|---|---|
| Command | npx @playwright/mcp@latest | npx @playwright/mcp@latest --port 8931 | docker run -i --rm --init --pull=always mcr.microsoft.com/playwright/mcp |
| Client config key | command + args | url: http://localhost:8931/mcp | command + args |
| Process lifetime | Started and stopped by the MCP client | Long-lived; clients connect to the URL | Long-lived when started with -d |
| Case the README gives for it | The default setup shown for every client | Headed browser without a display, or worker processes | Containerised runs; the sample passes --headless --browser chromium --no-sandbox |
| Image or package | @playwright/mcp on npm | @playwright/mcp on npm | mcr.microsoft.com/playwright/mcp |
Commands transcribed from the microsoft/playwright-mcp README, read 2026-09-02. Pin @playwright/mcp to a version instead of @latest if you need reproducible runs.
Playwright MCP gives a model a browser through structured accessibility snapshots instead of screenshots, so no vision model is needed. Its tools cover navigation and clicking, typing and form filling, drag, drop and file uploads. Extra capability groups for vision, PDF and devtools are opt-in through --caps, and profile flags control whether sessions persist.
The README describes the server as one that lets language models interact with web pages through structured accessibility snapshots, bypassing the need for screenshots or visually-tuned models. That is the design choice that separates it from screenshot-driven browser agents: the model receives a labelled element tree, so it can name a target instead of guessing at pixel coordinates.
browser_navigate, browser_click and browser_hover cover moving between pages and pointing at elements found in the snapshot.
browser_type, browser_fill_form, browser_select_option and browser_press_key handle text entry, whole-form fills, dropdowns and raw key input.
browser_drag, browser_drop and browser_file_upload cover the interactions that plain clicking cannot express.
The --caps flag turns on additional capability groups the README names as vision, pdf and devtools. They are off unless you ask for them.
--isolated keeps the browser profile in memory instead of on disk, and --storage-state points at a storage state file so an isolated session can start already signed in.
Playwright MCP is Microsoft’s Model Context Protocol server for browser automation. It gives an MCP client a real browser through Playwright and describes pages as structured accessibility snapshots, so the model names elements instead of guessing at pixels. It needs Node.js 18 or newer and is usually added with one npx command.
Playwright MCP is one server among many in the MCP server directory, and it covers a narrow job: giving a model a real browser. Servers such as GitHub MCP hand the model an API, while this one hands it a rendered page, which is the only option when the site has no API. Clients differ in how they register it, which is why the Cursor MCP and Claude MCP setups are written up separately.
The protocol underneath is public. The current Model Context Protocol specification revision is 2026-07-28; it defines a JSON-RPC 2.0 message format with stateless, self-contained requests, three participants (hosts, clients and servers), and three server-side features — resources, prompts and tools. Playwright MCP is a tools server: everything it offers arrives as a callable tool rather than as a resource.
If you are choosing between a browser-driving server and a purpose-built extractor, the trade is throughput against generality. A browser renders anything but costs a page load per step, which is the ground covered on web scraping. For agents that need to plan across several of these servers, see AI agent builder.
GitHub MCP ServerHow to install GitHub MCP Server: the remote endpoint at api.githubcopilot.com, the local Docker image, PAT and OAuth auth, and the toolsets to turn on.
Cursor MCPHow to add MCP servers to Cursor: project and global mcp.json, stdio, SSE and Streamable HTTP transports, static OAuth, config interpolation and allowlists.
Claude MCPHow Claude MCP works on three surfaces: claude mcp add in Claude Code, .mcpb extensions in Claude Desktop, and remote custom connectors on claude.ai and Cowork.n8n MCP ServerHow n8n does MCP: the instance-level MCP server at /mcp-server/http, the MCP Server Trigger and MCP Client Tool nodes, and the community n8n-mcp package.
Reddit MCP ServersReddit publishes no MCP server of its own. Three community Reddit MCP servers compared on install command, credentials, tools and whether they can post.Playwright MCP needs Node.js 18 or newer.
Use your client’s CLI (claude mcp add playwright npx @playwright/mcp@latest) or the standard JSON config.
Ask the agent to open a page and return a snapshot before letting it type or click.
Navigate, snapshot and read the elements — the only option when a site offers no API for the data you need.
Type, fill forms, select options and click through a flow in an isolated profile, stopping before irreversible steps.
Start an isolated browser with --storage-state so the session begins signed in without writing a profile to disk.
Run the server over HTTP with --port when several clients or agents need the same browser.
These answers cover what people ask before relying on Playwright MCP: whether Playwright needs a separate install, the exact Claude Code command, how to run headless on a server, how to keep a login between sessions, whether it is an official Microsoft project, and how it differs from the Playwright test runner.
The README lists Node.js 18 or newer as the requirement and the npx command fetches @playwright/mcp itself, so no separate install step is documented. The --browser flag selects the engine from chrome, firefox, webkit and msedge. The README's tool listing has a heading for browser installation, but that section is empty, so if your client reports a missing browser binary, install the Playwright browsers on that machine.
Run claude mcp add playwright npx @playwright/mcp@latest. That is the exact command the README publishes for Claude Code. Claude Desktop is different: the README points at the general MCP install guide and says to use the standard JSON configuration block shown above. More client-specific notes are collected on the Claude MCP page.
Yes. Pass --headless; the README notes the server runs headed by default, so headless is opt-in. For containers, the README's own long-lived Docker example runs /app/cli.js --headless --browser chromium --no-sandbox --port 8931 --host 0.0.0.0 from the mcr.microsoft.com/playwright/mcp image, which is the configuration to copy if you are deploying it as a service rather than launching it per session.
Two flags govern this. --isolated keeps the browser profile in memory, so nothing persists after the run. --storage-state takes a path to a storage state file, which is how you start an isolated session already authenticated. Without --isolated, the browser profile persists on disk in the normal Playwright way.
It is published in the microsoft/playwright-mcp GitHub repository, and the container image is served from Microsoft's own registry at mcr.microsoft.com/playwright/mcp. This page is a third-party write-up of that project; check the README before you rely on any flag here, since command-line options change between releases.
The test runner is driven by test files you write and assert against. Playwright MCP wraps the same browser automation library in a Model Context Protocol server so a model decides the next action at runtime, from an accessibility snapshot of the current page. They share the browser layer and nothing above it.
A browser server gets an agent onto a page. Getting the result into a spreadsheet, a CRM or a ticket is a separate problem, and that is what the automation side of this site covers: which step triggers the browser run, where its output lands, and what the workflow should do when a page layout changes.