Argonaut

Minimalist document translator with a Qt (PyQt5) interface that uses argos-translate-files as an offline translation engine.

Two engines are available from Settings → Engine:

• Argos Translate (default) — light per-pair models, installed and removed from Settings → Manage language packages….
• NLLB-200 — Meta's nllb-200-distilled-600M (int8, CTranslate2), noticeably better quality and direct translation between any pair of 32 languages. The model (~630 MB) is downloaded on first use and stored under ~/.local/share/argonaut/. Note that the NLLB weights are licensed CC-BY-NC 4.0 (non-commercial).

Requirements

Python 3.10 or newer (the pymupdf wheels Argonaut depends on need it).

pip install PyQt5 argos-translate-lt argos-translate-files langdetect psutil pymupdf

You need at least one language package installed. Open Settings → Manage language packages…, pick the pairs you want and press "Install selected"; the dialog lists every package from the Argos index with its version and download size, filters by installation state, and lets you remove or update the installed ones (a package with a newer version in the index is flagged and can be updated in place). They can also be installed from the command line with argospm install translate-en_es.

Usage

Install it (provides the argonaut command):

pip install argonaut-translator

From a checkout, install in editable mode first:

pip install -e .
python3 -m argonaut

Translating documents

1. Choose the source and target languages (⇄ button to swap them). By default the source is "Detect language": each document's language is detected automatically, so you can mix files in different languages in the same batch.
2. Drag documents or whole folders onto the window (dropped folders are walked recursively) or add them with "Add…". Each file shows its size and the list footer shows the batch total.
3. Optionally pick a folder with "Output…" where all translations are saved (× returns to the default behaviour: next to each original). Tick "Skip files already translated" to leave existing outputs untouched, which makes an interrupted batch resumable.
4. Press "Translate". Each translated file is saved with the target language suffix (e.g. report_es.docx).
5. During translation you can press "Cancel": the operation stops (even mid-file) and the files already translated are kept.

Note about PDFs: they are translated paragraph by paragraph while preserving the layout, so a long document can take a while on CPU. The bar shows the real percentage of translated paragraphs and an estimate of the remaining time, and the status line names the phase and the page it is on (reading, translating, generating, saving) so a long book shows progress instead of an apparently idle bar. Rotated text (e.g. vertical watermarks) is kept untranslated, and links from the original document are not preserved in the translated copy.

Supported formats: .txt .docx .odt .odp .pptx .epub .html .srt .pdf

Settings

Everything below lives in the Settings menu and is remembered between sessions.

• Engine — Argos Translate or NLLB-200 (see above).
• CPU threads — capped at the detected core count and defaulting to all of them. The NLLB engine splits this budget into several parallel CTranslate2 workers of about four threads each, which measured faster than one wide worker; budgets below eight threads are left as a single worker, where splitting measured slower.
• Translation quality — Best quality, or Fast, which decodes greedily: roughly 1.7× faster for a small loss in accuracy. The two modes never share cache entries, since their output differs.
• Theme — System (the desktop's own look, the default), Light or Dark.
• Download speed units — network units (16.8 Mbps) or the bytes a download manager shows (2.0 MB/s).
• Cache — see below.
• History — see below.
• Manage language packages… — the Argos package dialog.
• Delete the NLLB-200 model… — frees the ~630 MB the model occupies.

Translation cache

A segment translated once is reused for the rest of the batch and, by default, across sessions: a paragraph repeated in several files reaches the engine only once. The window reports how many segments each file and the batch as a whole reused.

The cache lives in a small sqlite database at ~/.local/share/argonaut/translation_cache.db (or under $XDG_DATA_HOME). Entries are namespaced by engine, model version and language pair, so upgrading a language package stops serving the old model's output. From Settings → Cache you can turn it off, choose how long unused entries are kept (never, 30 days, 90 days — the default — or a year; expired ones are pruned when a translation starts), see the database's size and entry count, and clear it.

Translation history

Every file a batch finishes is recorded — what was translated, into which language, where the result was written, with which engine and how long it took. Settings → History → View history… lists them newest first; double-clicking a row (or the "Open translation" button) opens the translated file, and rows whose output has since been moved or deleted stay listed, greyed out, since the history is a record of what happened rather than a file browser.

Like the cache, it lives in its own sqlite database (~/.local/share/argonaut/history.db) and the same submenu lets you stop recording, choose how long entries are kept (never, 30 days, 90 days — the default — or a year), see its size and entry count, and clear it. Viewing and clearing keep working while recording is off, and only successful translations are recorded: a file that failed is not history, it is an error.

Interface language

The interface starts in the language your desktop is set to, falling back to English when that is one Argonaut does not speak (Qt is asked for the whole ordered preference list, so a second choice can still match, and region and script are ignored — pt-BR gets the Portuguese strings). In the Language menu you can switch to Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Dutch, Polish or Turkish; the change applies instantly and that is what gets saved (QSettings). Until you choose one, the desktop keeps deciding — so changing the system language changes Argonaut's too.

The strings are JSON, one file per language in src/argonaut/locales/<code>.json, read only when that language is actually used. To add a language, copy en.json, translate the values and list the code in LANGUAGES in i18n.py (the menu order and the name shown for it); missing keys fall back to English. The test suite checks that the menu and the shipped files agree, and that every language carries the same keys, the same {placeholders} and no duplicate keyboard accelerators within a menu.

The languages you translate between are named in that same interface language: French reads "Francés" in Spanish and "フランス語" in Japanese, in the source and target combos, in the package dialog and in the progress and error messages — each list ordered alphabetically by the name shown. Those names are the one generated file in that folder, locales/language_names.json, built from CLDR by tools/gen_language_names.py (pip install babel && python tools/gen_language_names.py); Babel is a development tool, not a runtime dependency. A language with no entry keeps the English name the engine reports.

The Help menu opens the user manual (this README, on GitHub) and Report a bug…, which goes straight to the issue tracker. Both open in your browser: translation is offline, but documentation and bug reports are the two things that benefit from being current.

It also includes "About Argonaut…", a dialog with three tabs: About (version, a short description, the supported formats and links to the issue tracker and the releases), Details (a plain-text report with the Python, Qt and dependency versions, the operating system, the active engine, whether the NLLB model is installed and where the cache lives — copyable with one button, ready to paste into a bug report) and Credits and licenses.

Persistent settings

When the window closes, QSettings stores — besides the interface language — the source and target languages, the output folder and the window size/position; everything is restored on the next launch. If a saved language is no longer installed or the folder no longer exists, the default value is used. Everything in the Settings menu is saved as soon as it changes.

Structure

All modules live in the src/argonaut/ package:

• __init__.py — package version (__version__).
• main.py — entry point; silences dependency warnings and launches the window.
• window/ — the main window, split by responsibility: main_window.py (widgets, menus and window state), engine.py (backend, threads, quality, the cache and history menus and the NLLB model), files.py and file_list.py (the file list and its columns), translation_run.py (driving a batch and reporting its progress), theme.py (light/dark/system palettes), about_dialog.py and history_dialog.py.
• worker.py — thread that translates the file list and emits progress signals.
• pdf.py — fixed PDF translator (paragraphs, progress, cancellation).
• typesetting.py — lays the translation back into the page: line breaking, fitting each paragraph to its box, and a font per script.
• translation.py — language detection, supported formats and the progress/cache wrapper.
• history.py — persistent record of the files each batch translated.
• nllb.py — optional NLLB-200 backend (CTranslate2 + SentencePiece) exposing the same duck-typed API as argostranslate.
• download.py — shared streaming download with progress and cancellation.
• packages.py — Argos package index, download, installation and removal.
• package_dialog.py — dialog to browse, install and remove packages.
• i18n.py — interface language: which one is in force, how it is chosen and saved, and lazy access to the data files below.
• locales/<code>.json — the interface strings (English by default, Spanish, French, German, Italian, Portuguese, Russian, Chinese, Japanese, Dutch, Polish and Turkish).
• locales/language_names.json — the translation languages' names in each of those interface languages (generated; see tools/gen_language_names.py).

Packaging lives at the top level: pyproject.toml (PyPI), io.github.nibblex.Argonaut.yml (Flatpak manifest) and data/ (desktop entry, AppStream metainfo and icon for Flathub).

Tests

pip install -e .[test]
pytest

The suite runs Qt headless (QT_QPA_PLATFORM=offscreen) and needs no language models: translation engines are faked. Coverage is printed at the end of the run; CI publishes the badge on every push to main.

Releasing

To publish a new version, bump __version__ in src/argonaut/__init__.py and add a <release> entry in data/*.metainfo.xml.

PyPI

Publishing is automated: commit the bump, tag it and create a GitHub release for that tag.

git tag -a v1.5.0 -m "Argonaut 1.5.0"
git push origin main && git push origin v1.5.0
gh release create v1.5.0 --title "Argonaut 1.5.0" --notes-file notes.md

Publishing the release triggers .github/workflows/publish.yml, which builds the wheel and the sdist and uploads them to PyPI via trusted publishing. Pushing the tag alone does not publish anything, so the CI run on main (the test suite on Python 3.10 through 3.14) can be used as a gate: a version is permanent on PyPI once uploaded.

Flathub

The manifest is io.github.nibblex.Argonaut.yml. Flathub builds have no network access, so the Python dependencies must be pinned first with flatpak-pip-generator:

python3 flatpak-pip-generator --requirements-file=requirements.txt \
    --output python3-requirements

Test locally, then lint:

flatpak-builder --user --install --force-clean build-dir \
    io.github.nibblex.Argonaut.yml
flatpak run --command=flatpak-builder-lint org.flatpak.Builder \
    manifest io.github.nibblex.Argonaut.yml

First submission: take a screenshot for the metainfo (see the TODO in data/*.metainfo.xml), then open a PR against flathub/flathub (branch new-pr) adding the manifest, per the submission guide.

Author and license

© 2026 Sergio Rodríguez.

This project is distributed under the GNU GPL v3 license, in line with the license of PyQt5, which the interface depends on.