MCP serverPartly tested (needs an account)
Control a browser
Playwright MCP
Microsoft's MCP server that lets an agent open web pages, click, type and read the page. Good for testing your own site and for web tasks.
- Works with
- Claude Code, Codex CLI, Gemini CLI
- Good for
- Web developers, Build an app, Daily work and research
What it does
What it is
MCP is a standard way to connect tools to an AI agent. This MCP server gives the agent a browser. The agent can open a page, read what is on it, click, type and take notes.
It uses the page structure as text, so it does not need to look at screenshots.
Who it helps
- Web developers who want the agent to test their own pages.
- Anyone who wants the agent to collect information from public pages.
How to start
Run the install command for your agent, then ask for one small browser task on your own site.
Be careful
A browser can do a lot. Do not let it work on pages where a click can buy something or send a message. Keep it away from your signed-in browser profile.
How to install
claude mcp add playwright npx @playwright/mcp@latestcodex mcp add playwright npx "@playwright/mcp@latest"How to use it
After install, ask the agent to do something in a browser, for example "open my site, click the sign-up button and tell me if the form works". It reads the page as text, not as a picture, so it is fast.
What to watch out for
- It controls a real browser and can open any web address, fill forms and click buttons. Do not point it at pages where a click can cost money or send messages.
- If you let it use a browser profile that is signed in, it can act as you. Use a clean profile.
- A web page can contain hidden text that tries to steer the agent.
- The @latest tag means you get a new version each time. Pin a version if you want a stable setup.
Licence and author
- Licence
- Apache-2.0
- Author
- Microsoft
- Source code
- github.com/microsoft/playwright-mcp
How we tested it
We installed it on an empty test computer on 12 Oct 2026. Partly tested: installed in a clean test system, but its main features need an account or key we did not use.
No web page was opened, so the browser download and real page actions were not tested. Tools such as file upload and script evaluation increase the risk; read the tool list.
Technical details of the test
- Date
- 12 Oct 2026
- Method
- In a clean container with no account we started npx -y @playwright/mcp@latest, sent the MCP initialize and tools/list requests, and it answered with 25 tools. We also registered it with claude mcp add, and claude mcp list reported it as connected. We did not open a browser page.
Checked again on 12 Oct 2026
Sources
- github.com/microsoft/playwright-mcp