Translating a website is the same work as feedback: someone goes through the site, clicks a specific text, proposes the wording and the team approves it. fydbek adds an AI pre-translation that knows the industry, tone and terminology of the whole site, and a translation memory, so the same text is always translated the same way. Nothing is written to your website, fydbek stores the translations. You download approved translations as message files, or your AI agent writes them into the code over MCP.
Turning it on
An organization admin turns translations on in the project settings, Translations tab: the source language of the site (for example en) and the target languages (cs, de, fr). A project can have up to 5 target languages. Then open the Website translations page and crawl the website.
How fydbek finds texts
- Crawling reads the sitemap (
/sitemap.xmlor the URL from the settings). Without a sitemap it follows links from the home page, up to 200 pages. It only reads the source language version and skips paths like/de/.... It runs again once a week. - The widget in translation mode also finds texts that appear only after a click (menus, tabs, dialogs). Open the site with
?fydbek&translate=de, or click Translate in the widget bar. - A text is identified by its content, not by its position on the page. The same text on several pages (menu, footer) is translated once. Links and emphasis inside a text are tags
<0>...</0>that must stay in the translation. - Besides texts, image descriptions (
alt), titles, field hints and the page title and description are translated too.
What is not translated
Elements with translate="no", the class notranslate, data-fydbek-ignore or data-fydbek-mask, scripts, code in <pre> and form fields. Texts without letters (numbers, prices), bare URLs and e-mails.
Keys and context in code (optional)
You do not have to touch the site. If you want, give an element data-fydbek-key="home.hero.title" (the key in your message file) and data-fydbek-context="Header button" (a note for translators and AI). Keys can also be assigned without changing the site: import the source message file (for example en.json) in the Export and import tab.
Brief, glossary and AI pre-translation
AI proposes a brief from the site content (industry, audience, tone, rules for numbers and units) and a glossary of terms that are always translated the same way or not translated at all. The team edits and confirms both. A person starts the pre-translation, never automatically, and sees a price estimate first. AI translates page by page with the page context, nearby texts and the element role (button, heading, menu) and keeps short elements short. The result is always a draft that a person approves.
AI runs on Claude through Google Vertex AI in the EU. Only website texts, the brief and the glossary go to the model, no comments or personal data. The model treats website content as data to translate, not as instructions.
Statuses and approval
| Status | Meaning |
|---|---|
| Missing | The text has no translation yet. |
| AI draft | AI pre-translation or a match from an import, nobody has seen it. |
| Edited | Edited by a person, waiting for team approval. |
| Client review | Sent to the client, who approves, edits or comments. |
| Approved | Goes to the export and to the translation memory. |
The team approves. A client with the role Approves translations approves texts the team sent for review. A guest with the role Translator edits but does not approve. Set the role on a share link or on a guest. An AI agent over MCP only proposes.
Export
- next-intl: nested JSON by key, links and emphasis as
<link>and<b>fort.rich. Curly braces from the translation are escaped, placeholders of the source stay. - i18next: JSON by key with numbered tags for the
Transcomponent. - Original and translation: flat JSON for texts in a CMS (for example Makeswift) that are copied by hand.
Texts without a key are under unmapped, keyed by the original. An agent in the website repository can find them in the code and write the key back to fydbek (MCP tool set_string_key, prompt apply_translations).
MCP and API
The MCP server has the tools get_translation_status, list_strings, get_string, get_glossary, export_translations, set_string_key and update_translation. An API token needs translations:read, and translations:write for proposals. See MCP for AI agents.
Price
Manual translations and memory matches are free. AI pre-translation is billed by characters of the source text and language, once a month by invoice. Usage is in the organization settings, Translations tab.
Support
Write to hello@fydbek.com.