Puppeteer

This community server replaces the archived @modelcontextprotocol/server-puppeteer. The fork by Meraj Mehrabi is inspired by the reference implementation but fixes the CDP-attach flow and tracks a recent Chromium release. Use it when a page will not render without JavaScript. What it does: - Takes full-page or element-level screenshots as PNG or JPEG - Evaluates arbitrary JavaScript in the page context and returns the serialised result - Clicks, types, and form-fills by CSS or XPath selector - Navigates with wait-for-selector and wait-for-load-state primitives so flaky sites settle first - Attaches to an existing Chrome instance via --remote-debugging-port (useful for sessions behind SSO) - Optionally intercepts requests so an agent can fill auth headers without shipping a credential into the page Example: Launch with "npx -y puppeteer-mcp-server". Give an agent the task "open the status dashboard at status.example.com, wait for the incidents list, and screenshot the banner" — the agent navigates, waits for the React tree to hydrate, and returns a base64 image plus the extracted title text. Setup: No required env vars. Optional PUPPETEER_EXECUTABLE_PATH for a custom Chromium, PUPPETEER_REMOTE_DEBUGGING_PORT to attach to an already-running browser. Because Chromium is bundled, the first run downloads ~150 MB. Transport: stdio. A Chrome process is spawned per session and torn down on disconnect. Notes: Heavier than the Fetch entry — full Chromium rather than an HTTP client. If your site is static HTML with minor JS use Fetch first and reach for Puppeteer only when the rendered DOM diverges from the initial HTML. Attach mode is particularly valuable for agents that need to operate behind SSO: launch Chrome once, log in manually, then let the agent reuse the authenticated browser context.