LocalizationAI

A Godot 4.7 editor plugin that translates .csv and .po localization files with an LLM — either a local Ollama server (free, offline) or OpenRouter (paid, hosted). You drop your file into a node graph, pick the languages, hit Run, and watch the strings fill in.

What it does

You build a small graph — Source → Translate → Export — and the plugin handles the rest. CSV translation tables and gettext .po catalogs both work; existing translations stay put, only empty cells get filled. Pick 75+ target languages in one pass. Add a Prompt node to inject custom instructions (tone, terminology, formality) globally or per language. Point a Directory Source at a folder to batch a whole project's worth of files through the same translate node.

Speed comes from two knobs that stack: parallel chains (different files at once, top toolbar) and parallel requests (concurrent strings within one file, on the translate node). OpenRouter handles 8+ in flight comfortably; small local models usually want 1–2.

Long runs are safe to interrupt. Pause lets in-flight requests finish, then waits. Stop flushes a *_progress.csv / *_progress.po partial file you can feed back through the graph to resume exactly where you left off — already-filled cells are skipped. The same partial gets written if Godot crashes, the OS reboots, or memory runs out. Speaking of which: a built-in memory guard aborts cleanly when free RAM drops below 800 MB so a too-big local model can't swap your desktop into a freeze.

Translations are tuned for game text. The system prompt preserves %s, {name}, BBCode tags, line breaks, and capitalization style. A post-processing pass strips the usual LLM noise (Translation: prefixes, echoed source, surrounding quotes the source didn't have). Save the whole graph — nodes, connections, models, languages, parallel counts — to a JSON workflow under addons/localization_ai/workflows/ and load it next session.

No Python packages to install. The translator is stdlib-only; you just need python3 on PATH. Ollama models can be listed / pulled / deleted from the editor's Models tab. OpenRouter keys are kept in a local keyring under Godot's user:// folder (%APPDATA%\Godot\app_userdata\<project>\ on Windows, ~/Library/Application Support/Godot/app_userdata/<project>/ on macOS, ~/.local/share/godot/app_userdata/<project>/ on Linux) so workflow JSONs stay safe to share.

Install

1. Copy addons/localization_ai/ into your project's addons/ folder.
2. Project → Project Settings → Plugins → enable LocalizationAI.
3. Make sure python3 is on your PATH.
4. For local mode: install Ollama and ollama pull llama3.1:8b (or whatever fits your RAM). For hosted: get a key at openrouter.ai.
5. Open the LocalizationAI main-screen tab.

Quick start

1. Right-click the canvas → Add File Source → pick a .csv or .po.
2. Add Translate → choose provider, model, source + target languages.
3. Add Export → pick an output folder.
4. Connect File Source → Translate → Export (left-to-right green ports).
5. ▶ Run.

A starter graph lives at addons/localization_ai/workflows/basic_workflow.json. The 📖 Wiki has a Getting Started walkthrough and a full Node Reference covering every setting.

Standalone CLI

The translator works without Godot:

python3 addons/localization_ai/scripts/translate.py \
    --input strings.csv --output strings_translated.csv \
    --stopped-output strings_progress.csv \
    --provider local --model llama3.1:8b \
    --target-lang bg,da,tr --source-lang en \
    --api-url http://localhost:11434 \
    --workers 4

Model management:

python3 addons/localization_ai/scripts/manage_models.py \
    --action list --api-url http://localhost:11434

Architecture (one paragraph)

The editor side is @tool GDScript driving a GraphEdit. Real work happens in addons/localization_ai/scripts/translate.py, launched as a background process. The two sides talk through three temp files per run (progress / control / prompts), so pause / stop / live progress work without ever blocking the editor. CLAUDE.md has the deep dive.

License

MIT — see LICENSE.

Use at your own risk

Architecture and design are mine; AI did a lot of the typing. There's no test suite or CI, so back up your .csv / .po files before pointing this at anything important and spot-check the output — LLMs occasionally drop placeholders or translate things they shouldn't. OpenRouter / Ollama costs and rate limits are yours; API keys live in a plaintext keyring under Godot's user://.