M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
P

Playwright MCP

Official
Give AI agents full browser control via structured accessibility snapshots
How to install ↓ Compare GitHub source repository ↗
Category
Browser #25 of 32
Stars
★ 97.5k Very popular
Transport
stdio (local process)
Runtime
Node.js
Credentials
No credential needed
License
Apache-2.0
Last commit
43FMRS · D

Playwright MCP is Microsoft's official AI-agent browser control server, driving interaction through accessibility snapshots rather than screenshots, and connecting to mainstream MCP clients over stdio via npx; the README does not enumerate specific tool names, so the tool list is left empty.

Strongest · Maintenance 17/20 Weakest · Reliability 3/20

Reliability
3/20
Security and permissions
8/20
Maintenance
17/20
Documentation
9/20
Setup experience
6/20
Why each score
Reliability 3/20
Critical mismatch: the object claims the server lives at the repository root, but the README explicitly states the actual MCP server implementation (tools, accessibility snapshots, element-ref interactions) lives in a separate repository, github.com/microsoft/playwright-mcp, distributed via `npx @playwright/mcp@latest`. The microsoft/playwright repo itself is the Playwright testing/automation library; no manifest is cached, and there is no evidence in the provided material of MCP init logic, a tool list, or handshake code at the declared path. Declared tool behavior cannot be verified against real code. Per the static calibration rule, unverifiable claims are deducted, not assumed best-case, so this scores very low.
Security and permissions 8/20
Because no actual MCP server code exists at the declared root path, permission scope, credential handling, and dangerous-operation confirmations cannot be assessed here. No red-line evidence (malware, credential exfiltration, unconfirmed destructive defaults) was found, so the 0-4 red-line band does not apply. However, the underlying Playwright library does carry automation capabilities (network interception, filesystem writes, arbitrary page navigation) that an MCP layer built on it would inherit; without server-specific documentation of scoping or confirmation flows, only a middling score is warranted.
Maintenance 17/20
Repository-level evidence is strong: Apache-2.0 license, 94,512 stars, 148 open issues, not archived, maintained by Microsoft, and a widely used, actively developed project. Browser version badges (Chromium/Firefox/WebKit) in the README are kept current, indicating an active release cadence. This is solid, verifiable evidence of sustained maintenance and clear ownership.
Documentation 9/20
The README is well organized, covering multiple usage paths (Test, CLI, MCP, Library, VS Code extension) with code examples, giving generally solid documentation quality. However, for the MCP capability specifically, only a minimal install JSON snippet and one interaction example are given; tool parameters, limits, troubleshooting, and cost details are deferred entirely to the external playwright-mcp repository, leaving a clear documentation gap for the object actually being reviewed.
Setup experience 6/20
The MCP install command in the README (`npx @playwright/mcp@latest`) points to a different npm package and different GitHub repository than the object under review (microsoft/playwright repository root). No executable MCP server entry point or config matching the declared path was found in this repository to substantiate the setup steps. Per the static calibration rule, an install path that reads simply in prose cannot be scored highly without verifiable execution evidence at the declared location.

Static review · not runListed 2026-08-26

Read the FMRS scoring method →

Fit and risk

What it can accessUses the networkControls a browser

Best for

  • Users who want an LLM to drive open-ended browser tasks
  • Developers already using MCP-capable clients like Claude Desktop, VS Code, Cursor, or Windsurf
  • Cross-browser (Chromium/Firefox/WebKit) automation without hand-writing test scripts

Not for

  • Teams needing a full end-to-end test runner with parallel execution and trace viewer — the README points to Playwright Test for that instead
  • Coding-agent setups that prioritize token efficiency and want to avoid loading large tool schemas and accessibility trees — the README recommends the lighter-weight Playwright CLI for that case
  • Environments requiring strict isolation from real browser sessions, cookies, or local files, since the agent gets genuine browser control

Required permissions

  • Local Node.js/npx runtime to launch the server process
  • Ability to download and run Chromium/Firefox/WebKit browser binaries
  • Outbound network access to load whatever pages the agent is directed to visit

Risks and side effects

  • The server grants agents real control of a browser, so a manipulated prompt could cause navigation to or submission of data on unintended sites
  • If a saved storage state (login session) is loaded, the agent may gain access to authenticated account data
  • The README does not describe any authentication or access-control mechanism built into the MCP server itself — safety depends on the host client and how target URLs are constrained

Setup

Before you start

Runtime：Node.js

Add an MCP server entry to your client (VS Code, Cursor, Claude Desktop, Windsurf, or Claude Code) with command npx and args ["@playwright/mcp@latest"]. VS Code users can alternatively use the one-click install badge provided in the repo. Claude Code users can run claude mcp add playwright npx @playwright/mcp@latest. On first run, npx fetches the latest package and starts a local process that the client talks to over stdio.

claude_desktop_config.json 
{"mcpServers":{"playwright":{"command":"npx","args":["@playwright/mcp@latest"]}}}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add playwright -- npx @playwright/mcp@latest

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After adding the documented config to your MCP client (VS Code, Claude Desktop, etc.), Playwright's browser-control tools should appear in the client's tool list; then ask it to navigate to https://demo.playwright.dev/todomvc — a returned accessibility snapshot confirms the connection works.

Troubleshooting

  1. Verify Node.js is installed and `npx @playwright/mcp@latest` runs successfully on its own
  2. Check that the MCP client config's command/args match the documented example exactly
  3. Confirm the client supports the MCP stdio transport
  4. If the browser fails to launch, check whether Playwright's browser install step is needed and whether the target site is network-reachable

Things to try

Once connected, you can ask your AI assistant things like:

  • Navigate to https://demo.playwright.dev/todomvc and add a few todo items.
  • Open https://playwright.dev/ and take a screenshot of the page.
  • Check all completed todos on https://demo.playwright.dev/todomvc.
  • Fill in the email and password fields of the sample form and submit it.

Use cases

Letting a coding agent or chat assistant navigate, click, and fill forms on real web pages
Driving browser automation from an LLM without relying on screenshots or vision models
Having an AI agent verify multi-step web flows such as form submission or checkout

Supported clients

Claude Desktop
Claude Code
VS Code
Cursor
Windsurf

Listed from the project's documentation, not tested by this site.

Overview

Playwright MCP is a Model Context Protocol server from the Playwright project (Microsoft), launched locally via npx, that gives AI agents full control of Chromium, Firefox, and WebKit. Agents perceive pages through structured accessibility snapshots rather than screenshots or vision models, and use element refs (e.g. e5, e10) to click, type, and interact deterministically. The README states tool coverage spans navigation, form filling, screenshots, network mocking, and storage management. The server itself is maintained in the separate microsoft/playwright-mcp repository and is presented as an official companion project within microsoft/playwright's README and Resources section.

Similar servers

Compare with other MCP servers

Source revision d9f2fd3e2232 Data synced 2026-10-12 Read the FMRS scoring method