Translate your Miliastra Wonderland Stage into every language the game speaks.
NARN is a translation workbench built for Miliastra Wonderland Craftspeople. It reads the CSV the game exports — a quirky, non-standard dialect that breaks ordinary spreadsheet importers — ships 12 read-only glossaries (332 terms, 14 languages) covering the game's characters, elements, reactions, stats, nations, ranks and Miliastra's own vocabulary, and writes back a file the game can import.
It connects to AI models (Claude, GPT, Gemini, DeepSeek, GitHub Copilot, and more — Google AI Studio and OpenRouter both have free tiers) and classical machine translation (DeepL), keeps your terminology consistent with glossaries, never re-translates the same line twice with translation memory, and catches mistakes with AI-powered review — all from one workbench.
Nothing about it is Miliastra-only: if you have a CSV of text in any language, NARN will translate it. That is simply the workflow it was built around.
NARN is an independent, fan-made tool. It is not affiliated with, endorsed by, or sponsored by HoYoverse / COGNOSPHERE PTE. LTD. Genshin Impact and Miliastra Wonderland are their trademarks. The bundled glossaries are unofficial fan-made terminology aids — see TRADEMARKS.md.
Hosted, at narn.zercade.dev — sign in and start translating. Nothing to install and nothing to maintain; connect your own AI provider keys, or use ones already set up for you.
Self-hosted, with Docker — run it entirely on your own machine, with everything staying local:
curl -O https://raw.githubusercontent.com/zercade-dev/app-narn/main/docker-compose.yml
docker compose up -dThen open http://127.0.0.1:3001 and create a vault password — your API keys are encrypted and stored on your machine, never sent anywhere but the AI provider you connect. See Self-hosting in detail below for the plain-Docker alternative and building from source.
- Glossaries keep names, terms, and tone consistent everywhere they appear.
- Translation memory reuses your past translations, so identical or near-identical lines don't get re-translated (and re-billed) from scratch.
- AI review flags fluency, tone, and terminology problems and suggests fixes before you ship.
- Collaborate, on the hosted version. AI is optional — share a project with other people and ask them to help translate, scoped to just the languages they're responsible for. They can still use AI if they want to, with their own provider keys, not yours.
- Classical machine translation. DeepL is built in, for fast, non-AI translation whenever you'd rather not use an AI provider at all.
- Bring your own model, self-hosted. Beyond the built-in providers, point NARN at any OpenAI-compatible endpoint — including a local LLM running on your own hardware, for translations that never leave your machine.
Configure your providers — connect any mix of AI providers and classical MT, side by side.
Invite collaborators — on the hosted version, give people write access to specific languages only.
Translate at scale — every language for every string, side by side.
Track runs and cost — every translation and review run is logged with its cost.
Catch mistakes before they ship — AI review flags issues with one-click fixes.
The command above is the fast path. For the plain-Docker alternative, what the volumes and environment variables do, and building from source, see Self-hosting NARN.
NARN is licensed under the Business Source License 1.1 (BUSL-1.1) — see
LICENSE, which is the authoritative text. In summary: production use is
permitted, including running a non-commercial hosted instance, but offering NARN to third
parties as a commercial hosted or managed service requires a separate licence. On the Change
Date stated in that file, the licence converts to the Apache License, Version 2.0.
Third-party code redistributed with NARN is covered by
THIRD-PARTY-NOTICES.md. Trademark and no-affiliation statements
are in TRADEMARKS.md.