One install, four passes, no runtime dependency on us.
The script finds your strings. Your rulebook constrains the AI. Two automatic scans check the result. And in Build mode everything ships as files inside your own site — so if BeMyWords vanished tomorrow, your pages would still be translated.
It reads the site you already have.
In Embed mode the script reads your sitemap and learns new text as real visitors browse — including copy inside components and CMS fields. In Build and Sync modes, you push keys from your repo instead.
- Works on Tilda and hosted sites, Astro, Rails, React, Vue, iOS, Android, Flutter.
- Ignores anything you mark do-not-translate.
Your rulebook runs before the model does.
Glossary, tone, substitutions, product names that never change. The AI receives them as hard constraints, not suggestions — so "Skiwo" stays "Skiwo" in all five languages.
- Per-language prompts and overrides.
- One translation memory across every project.
Two scans, then a number you can act on.
Checkup reviews every language of every key against your rules; the language audit flags impostor strings — English text living in your Norwegian tab. You get one score and a short list of things worth a human eye.
Your build ships the files. Not a runtime service.
In Build mode your build fetches finished translation files and ships them inside your site — no extra network hop for your visitors, no vendor in your critical path. Or serve live from the script, or sync over the API.
- JSON, YAML, .strings, .xml, .arb — whatever your stack expects.
- Export everything, any time, no lock-in clause.
Point it at your own inference endpoint and the price becomes 199 kr.
Any OpenAI-compatible endpoint — hosted or self-hosted, a separate vision endpoint too if you want. Paste the endpoint and key, and switching to the flat plan is one click. It stays flat no matter how many words you push through it.
See how the switch works