Use Playwright integration to browse the web, take screenshots, and interact with web pages in a headless Chromium browser. Use whenever the user asks to screenshot a page, navigate to a URL, check what is on a website, fill a form, click through a flow, or extract content from a rendered page. Use even when you think you know the page content, websites change constantly, only the live browser reflects current rendering, JS-injected content, and dynamic state. Do not use for fetching plain HTML or JSON that does not need JavaScript rendering (use WebFetch), scraping that violates the site's terms of service, anything requiring credentials the user has not provided, or local file access (the Playwright container has no /workspace mount).
Use Playwright integration to browse the web, take screenshots, and interact with web pages in a headless Chromium browser. Use whenever the user asks to screenshot a page, navigate to a URL, check what is on a website, fill a form, click through a flow, or extract content from a rendered page. Use even when you think you know the page content, websites change constantly, only the live browser reflects current rendering, JS-injected content, and dynamic state. Do not use for fetching plain HTML or JSON that does not need JavaScript rendering (use WebFetch), scraping that violates the site's terms of service, anything requiring credentials the user has not provided, or local file access (the Playwright container has no /workspace mount).
Access via MCP Hub: search_tools to discover methods, execute_code with the injected playwright global. Runs as headless Chromium in a hardened container: no filesystem sharing with /workspace.
Workflow
search_tools { query: "browser", detail_level: "full_schema" } to find the right method.
execute_code using exact parameter names from the schema.
Pitfalls
Screenshots return inline as base64: multi-content [{type:"text",...}, {type:"image", data:"<base64>", mimeType:"image/jpeg"}]. The filename parameter writes inside the Playwright container only. Do NOT access /tmp/playwright-mcp-output/ from /workspace, it does not exist there.
No file bridge: to get a PDF or screenshot into /workspace, use the base64 return value and write it from the Claude container.
Navigate, wait, act: after browserNavigate, wait for the page to settle (wait_for_selector or page idle) before clicking or reading. Do NOT click immediately after navigate.
Page state persists within one execute_code call but resets on the next call: plan multi-step flows inside a single execute_code block.
Prefer viewport screenshots: fullPage: true can crash Chromium on ad-heavy or iframe-heavy pages. Use it only for simple pages when the user explicitly requests full-page.
After a crash ("Target crashed" / "Target page, context or browser has been closed"): call browserClose({}) first, then retry with browserNavigate.
Public URLs only: no auth-walled sites unless you script the login flow explicitly.
Confirmation rule
Most Playwright use is read-only (screenshots, scraping). Confirm with the user before interacting with auth-walled or paid services, or performing any write/destructive action on a live site.
