A Five-Year-Old Intranet App Gets an AI Sidecar via MCP and npm

Many organizations sit on legacy internal tools that are too costly to rewrite but still gate critical data behind browser logins. This pattern lets an AI agent operate those tools through their existing APIs and authentication, turning a dead-end maintenance burden into automatable infrastructure without opening security holes.

An old internal web app with near-zero daily users can be given an AI interface without touching its codebase. The approach wraps the app's existing login and file APIs behind an MCP server distributed as a private npm package. An agent launches the server via npx, opens the user's local Chrome to complete a real login, captures the session token into process memory, and then calls the same list and download endpoints the browser would hit. Credentials never touch disk or configuration files, and the server process restart clears them completely. The technique works across Cursor, Claude Desktop, VS Code, and Windsurf because MCP is client-agnostic. The npm package bundles Playwright to drive the local browser, zod for input validation, and the MCP SDK to expose three tools: login_with_browser, list_files, and download_file. A logout tool optionally calls the backend to invalidate the server-side session before clearing memory. The entire setup avoids intranet tunneling, public exposure, or unauthenticated backdoors, keeping the old system's security model intact while letting an agent do the repetitive clicking.