Telemetry window: Nov 2025 → Jun 2026. Download window: Jul 2025 → Jun 2026. Read the full data POV →
What playwright-mcp does
One MCP server. Three things your agent can suddenly do.
Drive a real browser
Spin up a Chromium browser, navigate, click, type, scroll. The same primitives Playwright ships, exposed as MCP tools your agent can call.
264K browser sessions logged
Hand pages to your agent
Screenshot, interactive snapshot (annotated screenshot + semantic tree), full snapshot, text snapshot. Pick the resolution your agent needs.
643K screenshots vs ~100K DOM snapshots
Run custom JavaScript
The execute-code escape hatch lets the agent run arbitrary Playwright JS — for debugging, off-spec automation, or anything the standard tools don't cover.
401K execute-code calls
Get started in 30 seconds
Three paths to a working setup. Pick the one that matches how you work.
Paste this into Claude Code, Cursor, ChatGPT, or any coding agent. It'll figure out which client you're on and wire playwright-mcp into the right config.
Install and set up playwright-mcp for my coding agent. Package: playwright-mcp on npm Source: https://github.com/qabyai/playwright-mcp Detect which AI coding agent I'm using (Claude Code, Cursor, opencode, or Claude Desktop) and: 1. Locate or create the right MCP config file for that agent. 2. Add a "playwright-mcp" server entry with command "npx" and args ["-y", "playwright-mcp"]. 3. Restart the agent so it picks up the new server. 4. Confirm by asking the agent to list MCP tools — init-browser, get-screenshot, get-interactive-snapshot, get-full-snapshot, get-text-snapshot, and execute-code should appear. Prerequisite: Playwright's Chromium browser. If missing, run: npx playwright install chromium
Prerequisites: Node.js, Playwright with Chromium installed. Full setup walkthroughs and IDE-specific gotchas live at the playwright-mcp docs site.
The MCP tools you'll call
Six tools, ordered by real-world call count from the last 12 months of telemetry.
| Tool | What it does | Calls (12mo) |
|---|---|---|
| init-browser | Launch Chromium and navigate to a URL. | 264,268 |
| get-screenshot | Take a PNG screenshot of the current page. | 643,424 |
| get-interactive-snapshot | Annotated screenshot + semantic tree of buttons, links, and inputs. | 31,486 |
| get-full-snapshot | Complete page structure including all visible content. | 29,273 |
| get-text-snapshot | Text-only extraction — headings, paragraphs, lists. | 39,595 |
| execute-code | Run custom Playwright JavaScript against the page — debugging or off-spec automation. | 401,482 |
How real teams use it
Three findings from 1.42M tool calls across 5,904 domains that surprised us.
Localhost is the #1 site tested
127.0.0.1 beat every public domain — 272 users, 18,762 events. Agents test the change the developer made an hour ago, not staging.
Agents drive by sight, not DOM
643K screenshots vs ~100K DOM snapshots — a 6:1 ratio. The accessibility-tree paradigm is losing to vision. Vision tokens got cheap; agents prefer the camera.
SSO + GitHub dominate public domains
~158 users automate auth walls (Google, Microsoft, Atlassian). 75 users automate GitHub itself. Killer near-term capability: getting past the login.
Full POV — including the 41% activation cliff and the 187 MCP clients breakdown — what 230K downloads taught us about AI agents in CI/CD →
playwright-mcp vs Microsoft's @playwright/mcp
Different packages. Same protocol. Different shape. Both are valid choices — here's the honest delta.
| playwright-mcp (this one) | @playwright/mcp (Microsoft) | |
|---|---|---|
| Maintainer | QAby.AI team (independent) | Microsoft, alongside Playwright |
| 12-mo downloads | 230,105 | 60.4M (~267x larger) |
| Shape | Opinionated tools for AI-assisted test generation: interactive snapshots, annotated screenshots, focused execute-code escape hatch. | Broader Playwright API surface exposed as MCP tools. |
| Best for | Testing and end-to-end automation — inner-loop verification, test generation, and dev-loop debugging. | Agents that want raw access to the full Playwright API. |
The point of citing Microsoft's package isn't to claim parity — it's to confirm the category is real. Agentic browser automation grew ~40x year over year. Both packages benefit.
Frequently asked questions
What is Playwright MCP and what is it used for?
Playwright MCP is an open-source Model Context Protocol server that gives AI coding agents (Claude Code, opencode, Cursor, Codex, Cline, and 180+ other clients) a way to drive a real browser as part of their tool loop. In the wild, agents use it mostly for inner-loop verification ("did my change work?") and for getting past login walls so they can test the thing behind authentication.
How do I install playwright-mcp?
Install globally with `npm install -g playwright-mcp` or run on demand with `npx -y playwright-mcp`. Then add the server to your MCP-compatible client (Claude Code, Cursor, Claude Desktop, opencode, etc.) — config snippets are in the Get Started section above.
Which MCP clients does playwright-mcp work with?
Any MCP-compliant client. Telemetry from 1.42M tool calls shows 187 distinct clients on the wire, including Claude Code, opencode, Cursor, Visual Studio Code, codex-mcp-client, Gemini CLI, Cline, Windsurf, Trae, Kiro, GitHub Copilot, Antigravity, LM Studio, mcporter, and the qwen-cli-mcp-client.
Do agents prefer screenshots or DOM snapshots for browser testing?
Screenshots, decisively. In 1.42M tool calls we logged, agents fired get-screenshot 643K times (about 2.4 per browser session). All three DOM-snapshot calls combined totaled ~100K, less than a sixth of screenshot volume. Multimodal vision tokens got cheap enough that agents prefer to look at the page rather than parse the DOM tree.
What websites do developers test most with Playwright MCP?
Localhost (127.0.0.1) is the #1 target (272 users, 18,762 events). After that: search engines, auth and SSO walls (Google, Microsoft, Atlassian logins), GitHub, and a long tail of dev tools, social platforms, and Chinese-language platforms. The dev loop, not staging, is the modal use case.
What's the difference between playwright-mcp and Microsoft's @playwright/mcp?
Different packages, different vendors, same protocol. Microsoft's @playwright/mcp shipped 60.4M npm downloads in the last 12 months riding inside official Playwright distribution. Our playwright-mcp shipped 230,105 in the same window. The bigger story is the category itself: agentic browser automation grew ~40x year over year, validating the workflow for both packages.
Is playwright-mcp free? Can I use it commercially?
playwright-mcp is open source and free to install via npm. See the LICENSE in the GitHub repo for the exact terms (no derivative works clause). Commercial use of the package as distributed is permitted; if you want a managed, hosted version with regression coverage on every merge, see QAby.AI.
Two ways to run playwright-mcp
Use the open-source package on your machine, or let the QAby.AI team run it managed in CI with a regression layer on top.
Self-host
One npx away. Run playwright-mcp locally, point your coding agent at it, ship.
- →Install with
npx -y playwright-mcp - →Full source on GitHub — fork it, customize it
- →All 6 MCP tools, your machine, your browser
- →You handle install, upgrades, browser deps, CI wiring
QAby.AI
playwright-mcp running managed in CI on every merge, with AI agents owning the discover → build → run → heal loop.
- →Agents discover your flows and build the tests
- →Regression runs on every merge — managed CI, parallel browsers
- →Tests heal when your UI changes — no selector babysitting
- →Owned by your team, not your QA vendor