@modelcontextprotocol/ext-apps consumers ship ~140K of unused zod v4 locale files in browser bundles

[les2]

I discovered while developing an mcp-app using the sdk (client side only) that the bundle contained all of zod's locales because of a start import that prevented vite's / rolldown's tree-shaking from working.

I was able to work around this by developing (with the help of AI) a little vite plugin to strip it:

// Zod v4 re-exports every locale via `export * as locales` from its classic
// entrypoint, which defeats tree-shaking when the MCP SDK imports `* as z from
// 'zod/v4'`. Replace the locales index with an `en`-only stub so the namespace
// still exists but the other ~50 locale files never enter the graph.
const zodLocalesEnOnly = {
  name: "zod-locales-en-only",
  enforce: "pre",
  load(id) {
    if (/[\\/]zod[\\/](?:v4[\\/])?locales[\\/]index\.js$/.test(id)) {
      return 'export { default as en } from "./en.js";';
    }
    return null;
  },
};

See https://github.com/les2/mcp-ext-apps-zod-locale-bloat for a minimal demonstration of the issue.

Below is an AI issue description:

@modelcontextprotocol/ext-apps consumers ship ~140K of unused zod v4 locale files in browser bundles

Summary

Any browser app that depends on @modelcontextprotocol/ext-apps (and therefore on @modelcontextprotocol/sdk) bundles all ~50 zod v4 locale files — Spanish, French, German, Japanese, Chinese, Korean, Russian, etc. — because of a tree-shaking dead-end in the transitive zod chain. On our widget (Ember + Vite + Rollup) this was ~140K of dead weight, ~20% of the final bundle.

The end result is that every browser bundle currently ships translated error strings like 数値過大：期望, Trop court, demasiado corto, zu kurz — none of which are ever displayed, since the SDK only uses zod's English error messages at runtime.

A minimal reproduction is available at https://github.com/les2/mcp-ext-apps-zod-locale-bloat — npm install && npm run compare builds twice (without and with the workaround) and prints the size delta.

[hydrosquall]

I noticed this might duplicate #649