Can a prompt replace Weglot?

Build a translating reverse proxy for a single website you control.

Requests arrive on a locale-prefixed path — /fr/about — and the proxy fetches the source page, translates the visible text, caches the result and serves it. Translation goes through a machine translation API with a key, or a self-hosted LibreTranslate instance for a site where cost matters more than fluency.

Extraction is the part that decides whether this works. Parse HTML properly and walk text nodes, never regex the markup. Skip script, style, code and pre blocks entirely. Preserve inline markup inside a translated sentence by replacing each inline element with a numbered placeholder before translation and restoring it after — and write the test for that first, because a link that migrates to the wrong word is the failure mode people notice. Translate the title, meta description and image alt text too.

Cache per URL and per locale, keyed by a hash of the extracted source strings, so a page that has not changed is never retranslated. Store the cache in SQLite with the source hash, and add a management page listing every string with its translation and an override field. Overrides win over the machine output permanently and are what makes the system usable after week one — product names, tone, the phrase your industry uses.

Emit hreflang alternates on every page, including a self-reference and x-default, and add each locale to the sitemap. Getting this wrong is how a translated site ends up competing with itself in search.

Do not build a JavaScript widget that translates in the browser. Client-side translation is invisible to crawlers, which is most of the reason to translate a site at all.