bench get-app https://github.com/alyf-de/frappe-translator

[![Listed on Frappe Gems](https://frappegems.com/api/method/frappe_gems.seo.badge?app=alyf-de%2Ffrappe-translator)](https://frappegems.com/gems/apps/alyf-de/frappe-translator)

frappe-translator

AI-powered translation CLI for Frappe apps using the Claude CLI.

Massively parallelizes translation by assembling rich context (source code snippets, term glossaries, comments) and translating into all target languages simultaneously.

With Claude Code Max, this can translate around 500 strings into 30 languages per 4-hour token limit. However, the 4-hour token limit is already exhausted after 1 hour. So we're effectively doing 4 translations per second.

Requirements

• Python 3.10+
• Claude CLI with an active subscription (Pro/Max)
• A Frappe bench directory with apps containing locale/main.pot files

Installation

As a user

uv tool install https://github.com/alyf-de/frappe-translator.git

As a contributor

git clone https://github.com/alyf-de/frappe-translator
cd frappe-translator
uv sync
uv tool install -e .

Usage

Translate missing strings

frappe-translator translate /path/to/bench --mode fill-missing

Translate specific apps and languages

frappe-translator translate /path/to/bench \
  --app frappe --app erpnext \
  --language de --language fr \
  --model sonnet

Skip glossary extraction (single-pass, faster)

frappe-translator translate /path/to/bench --skip-glossary

Check translation coverage

frappe-translator status /path/to/bench

Dry run (see what would be translated)

frappe-translator translate /path/to/bench --dry-run

Clear progress (start fresh)

frappe-translator clear-progress /path/to/bench

How it works

Two-pass pipeline

1. Pass 1 (term extraction): Batches all translatable strings and sends them to Claude to extract key domain-specific terms (e.g., "Invoice", "Purchase Order", "DocType"). Existing translations for these terms are looked up across all PO files to build a glossary. This pass is optional (--skip-glossary) and cached across runs.

2. Pass 2 (translation): Strings are batched (default 50 per call) and sent to Claude with full context:

   • Source code snippets (up to 3 per string, from diverse files)
   • Term glossary with existing translations
   • Extractor comments from the POT file
   • Per-language style instructions (formality, address form)

   The LLM translates each string into all target languages at once. Failed batches are retried individually.

Key design decisions

• All languages at once: Context assembly is 80% of the work. Once rich context is assembled, translating into multiple languages simultaneously is efficient and produces more consistent results.
• Dependency ordering: Apps are processed in order (frappe, erpnext, hrms, ...) so earlier apps' translations inform downstream glossaries.
• Incremental writes: Translations are written to PO files every 50 entries. Re-runs skip already-translated entries.
• Per-language resume: If translation fails for one language (e.g., placeholder mismatch), only that language is retried on the next run.

Run modes

Mode	Description
fill-missing	Only translate entries with empty msgstr in any target locale (default)
review-existing	Re-translate entries that already have translations
full-correct	Re-translate all entries regardless of existing translations

Configuration

Create a frappe-translator.toml file:

[general]
concurrency = 5
batch_size = 50
timeout = 120
model = "sonnet"

[style.default]
formality = "formal"
notes = "Use formal address. Preserve technical terms commonly used in ERP software."

[style.de]
formality = "formal"
address = "Sie"
notes = "Use 'Sie' form. German compound nouns preferred over English loanwords."

[style.fr]
formality = "formal"
address = "vous"

[terminology.de]
"Workspace" = "Arbeitsbereich"

[apps.order]
priority = ["frappe", "erpnext", "hrms"]

Pass it with --config frappe-translator.toml.

Development

uv sync
uv run ruff check src/ tests/
uv run ty check src/
uv run pytest

License

MIT