16.3k stars · AGPL-3.0 · v1.9.6 (2026-05-26) · Track this in Scout
Self-hosted machine-translation API — removes the per-character meter from running a bilingual site.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A self-hosted machine-translation API built on the Argos Translate models. It runs entirely offline on CPU, exposes a plain HTTP endpoint, and ships a small web interface for one-off translations.
What it is good for. The moment a project commits to two languages the per-character bill starts, and a bilingual site plus bilingual docs plus bilingual video descriptions plus bilingual UI strings is a very large number of characters. DeepL and Google both meter that; this removes the meter. Point Tolgee's machine-translation provider at a local instance and the first draft of every Russian string costs nothing — a native speaker then edits the ones that matter, which is the part no machine does well. The same endpoint drafts the Russian video descriptions, the Russian llms.txt and the Russian blog posts. It is honestly not as good as DeepL. But for a first pass that a native speaker is going to correct, free and unlimited beats excellent and metered.
- No API key, no per-character cost, and no text leaving the machine it runs on
- Plugs into Tolgee (01) as a translation provider, so the two behave as one system
- CPU-only, and English↔Russian is one of the better-supported model pairs
- Quality is clearly below DeepL. Fine as a draft you edit; not fine as text you publish unread
- AGPL-3.0 — safe to run as a service you use, but read it before exposing a modified copy to users
- Language models are large downloads and the container wants a couple of gigabytes while translating
docker run -d --name libretranslate -p 5000:5000 \
