How to stop Claude from scanning your entire codebase every session

If you use Claude or Cursor on a real project, you have probably hit the point where it starts reading every file just to understand what the codebase looks like. It is not doing anything wrong. It just has no other way to get context. So it scans, and that scan costs tokens every single session.

I wrote about WebDNA a while back. The short version is that it is a Next.js plugin that generates a JSON file at build time describing your routes, components, API endpoints, and brand tokens. AI agents read that file first instead of crawling your source.

Since the first post I have added a two-tier manifest system so you can exclude private routes and sensitive components from what the AI sees, a fallback API route at /api/webdna for agents that cannot access .well-known directories, and a linter that catches missing descriptions and undeclared auth scopes before they become a problem.

If you want to try it on your project, the full setup guide is at webdna.omraj.guru. It takes about five minutes and works with any Next.js project using App Router or Pages Router.