microsoft/playwright-mcp ? reverse-engineered prompt

I want to build a server that can give a large language model the ability to control a web browser. It should use Playwright for the actual browser automation.

The main thing is that it should understand the structure of a webpage—its buttons, links, and input fields—without needing to look at screenshots or use a vision model. This should make it fast and lightweight. The goal is for an LLM or an agent to be able to navigate websites, fill out forms, and click on things based on structured data from the page's accessibility tree.

Please build this as a Node.js server using TypeScript. It should be designed so that different AI clients and agent development tools can connect to it and use its browser automation capabilities for tasks like exploratory testing or autonomous workflows.