Internationalizer

AI-native internationalization pipeline for software projects. Translate, validate, and manage i18n files using LLMs.

العربية · বাংলা · Čeština · Dansk · Deutsch · Ελληνικά · Español · Suomi · Français · עברית · हिन्दी · Indonesia · Italiano · 日本語 · 한국어 · Bahasa Melayu
Nederlands · ਪੰਜਾਬੀ · Polski · Português · Română · Русский · Svenska · తెలుగు · ไทย · Türkçe · Українська · Tiếng Việt · 粵語 · 简体中文 · 繁體中文

Why Internationalizer?

Most i18n tools are either runtime libraries (i18next, react-intl) or key-management SaaS platforms (Crowdin, Lokalise). None of them solve the actual translation problem well:

Manual translation doesn't scale past a few languages
Machine translation APIs (Google Translate, DeepL) ignore your terminology, tone, and UI conventions
Generic LLM translation works better, but without glossaries and style guides, you get inconsistent results

Internationalizer is different. It's a CLI pipeline that combines LLM translation with:

Per-language glossaries — enforce consistent terminology across your app
Per-language style guides — control tone, formality, pluralization, and typography
Translation memory — skip unchanged strings, save money on API calls
Deterministic validation — catch missing or extra keys, protected-structure drift, glossary issues, and plural or ICU errors before they ship
Explicit approval — keep provider or adopted provenance separate from human review
Fluent and pseudolocales — preserve translator context and exercise accented or bidirectional layouts without an API call

Installation

Install from npm:

npm install -g internationalizer

Or run without a global install:

npx internationalizer --help

The npm package installs the matching prebuilt binary from npm via platform-specific optional dependencies.

Install with Go:

go install github.com/Tom-R-Main/Internationalizer/cmd/internationalizer@latest

Or build from source:

git clone https://github.com/Tom-R-Main/Internationalizer.git
cd Internationalizer
go build -o internationalizer ./cmd/internationalizer

npm Packages

Git tags and npm package versions must match, for example v0.1.0 and 0.1.0
The root internationalizer package depends on platform packages such as internationalizer-darwin-arm64
Supported npm targets: macOS arm64/x64, Linux arm64/x64, Windows x64
CI publishing requires a GitHub secret named NPM_TOKEN

Quick Start

Create a config file in your project root:

# .internationalizer.yml
source_locale: en
target_locales: [fr, de, es, ja]
bundles:
  - id: app
    source: locales/en.json
    target: locales/{locale}.json
    format: json

llm:
  provider: gemini
  model: gemini-3.8-flash
  api_key_env: GOOGLE_AI_STUDIO_API_KEY

Set your API key:

export GOOGLE_AI_STUDIO_API_KEY=your-ai-studio-key

Preview what will be translated:

internationalizer translate --dry-run

Run the translation:

internationalizer translate

Validate and approve the exact generated artifacts:

internationalizer validate
internationalizer review list --status needs_review
internationalizer review approve --locale fr --all
internationalizer validate --require-approved

Commands

`translate`

Find missing or stale keys and translate them via an LLM.

internationalizer translate                    # translate all locales
internationalizer translate -l fr              # translate French only
internationalizer translate --dry-run          # preview without API calls
internationalizer translate --adopt-existing   # baseline existing translations without API calls
internationalizer translate --refresh-policy   # refresh prompt/style/model-stale entries
internationalizer translate --batch-size 20    # smaller batches
internationalizer translate --concurrency 2    # fewer parallel calls

Translation state independently reports missing, source-stale, policy-stale, current, and manually edited conditions, so a manual edit cannot conceal a source or policy change. Policy-stale values are reported but only retranslated with --refresh-policy. Manually edited values are never overwritten automatically. Use --adopt-existing when introducing the manifest to existing translations or when explicitly accepting a manual edit as the new provenance baseline. Adoption does not imply human approval; use review approve for that separate decision.

`pseudo`

Generate deterministic test locales without a provider or translation-memory lookup. Accented output defaults to en-XA; bidirectional output defaults to ar-XB. ICU and Fluent runtime syntax, code, links, and markup are preserved.

internationalizer pseudo                         # accented en-XA
internationalizer pseudo --strategy bidi         # bidi ar-XB
internationalizer pseudo --dry-run                # show planned artifacts only
internationalizer pseudo --locale qps-ploc        # choose another valid locale tag

The generator refreshes only artifacts it previously recorded as pseudo output. Use --force to replace any other existing target deliberately.

`review`

Inspect and approve the exact target content currently bound to its source and translation policy. Generated, cached, and adopted values all begin in needs_review; approval is invalidated by later source, policy, or target changes. Pseudolocales are tracked separately as test artifacts.

internationalizer review list --status needs_review
internationalizer review approve --locale fr --bundle app --key common.save
internationalizer review approve --locale fr --all

`validate`

Check all locale files against their source bundles. Default validation checks structural coverage (the percentage of required target keys present), reports extra keys as warnings, and fails for missing keys, interpolation mismatches, or invalid ICU MessageFormat structure.

internationalizer validate                     # human-readable output
internationalizer validate --json              # machine-readable JSON
internationalizer validate -q                  # exit code only
internationalizer validate --strict             # enforce translation quality rules
internationalizer validate --require-state      # require current manifest provenance
internationalizer validate --require-approved   # also require explicit approval

--strict also reports translated coverage. A linguistic value identical to its source is untranslated unless the glossary explicitly contains an exact same-source, same-target entry for the complete value; ignore_case is honored, but a glossary term embedded in a longer value is not an exemption. Strict mode fails on extra keys, source-identical values, changed interpolation/HTML/code/ Markdown-link structure, glossary violations, and configured plural forms.

--require-state verifies each target against .internationalizer.lock. It fails when a key is untracked, or when its recorded source, translation policy, or target hash is stale. --require-approved implies --require-state and also fails if the exact current artifact has not been approved. Both can be combined with --strict.

Human and JSON reports use stable finding codes:

Code	Meaning
`missing_key` / `extra_key`	Source and target key sets differ
`blank_translation`	A non-empty source has an empty strict-mode target
`source_identical`	A strict-mode linguistic value remains untranslated
`protected_structure_mismatch`	Interpolation, HTML, code, or link structure changed
`glossary_violation`	No approved target term or variant was found
`plural_form_missing`	A configured locale plural form is absent
`icu_message_syntax`	A source or target ICU message is malformed
`icu_argument_mismatch`	ICU argument names, types, or formatter styles differ
`icu_selector_mismatch`	Selectors differ or a plural category is invalid for the target locale
`untracked`	No manifest record exists for the target
`source_stale`	Source content changed after the recorded translation
`policy_stale`	The generated prompt or model settings changed
`target_modified`	Target content differs from the manifest record
`needs_review`	Current provenance exists, but the exact target is not approved

`detect`

Inspect existing configuration and discover catalogs, including nested apps.

internationalizer detect
internationalizer detect --json
internationalizer config check --json

Discovery reports runtime evidence, uncovered catalogs, and unresolved choices. i18next dependencies suggest message_syntax: i18next; ICU integration evidence requires a runtime decision. JSON is a storage format, not a message grammar.

`config plan` and `config apply`

Create a proposal, review its diff, then explicitly apply the saved plan:

internationalizer config plan --json \
  --add-bundle web=exf-app/web/src/i18n/locales/en.json \
  --syntax web=i18next --syntax default=i18next \
  --confirm-source tmp/english-keys.json --out config-plan.json
internationalizer config apply --plan config-plan.json --no-input --json
internationalizer translate --dry-run --json

This example assumes an existing source_path: tmp/english-keys.json marketing config. Select paths and syntax for your own runtime; discovery does not decide which artifacts ship. Plan/apply preserves existing provider settings, locale overrides, glossary paths, and bundle IDs. It rejects stale plans and recognizes an already-applied configuration. --no-input disables prompts; it does not authorize additional actions.

See the onboarding and JSON contract for initial setup, filters, error codes, and retry behavior. internationalizer commands --json describes installed workflow entry points and their effects.

JSON compatibility: validate --json now emits a schema_version: 1 envelope. Consumers of its previous array output must read data.reports instead.

`glossary`

Manage per-language glossary terms that are enforced during translation.

internationalizer glossary list --locale fr
internationalizer glossary add --locale fr --source "Dashboard" --target "Tableau de bord"
internationalizer glossary remove --locale fr --source "Dashboard"

`tm`

Manage translation memory (JSONL cache of previously translated strings).

internationalizer tm stats                     # show record counts
internationalizer tm export                    # dump as JSON
internationalizer tm clear --force             # delete all records

Configuration Reference

# .internationalizer.yml

# Source language (default: en)
source_locale: en

# Languages to translate into (required)
target_locales: [fr, de, es, ja, yue, zh-CN, zh-TW, ar]

# One or more source-to-target mappings (required).
# {locale} is replaced with each configured target locale.
message_syntax: auto # default; overridden per bundle below
bundles:
  - id: app
    source: locales/en.json
    target: locales/{locale}.json
    format: json
    message_syntax: i18next
  - id: docs
    source: README.md
    target: docs/i18n/{locale}.md
    format: markdown
    message_syntax: plain
  - id: browser
    source: browser/locales/en-US/browser.ftl
    target: browser/locales/{locale}/browser.ftl
    format: fluent

# Backward compatibility: source_path still maps targets to sibling files
# such as locales/fr.json. Prefer bundles for new projects.
# source_path: locales/en.json

# LLM provider settings
llm:
  # Provider: "anthropic", "openai", "gemini", or "openrouter" (default: gemini)
  provider: gemini

  # Model name defaults by provider:
  #   anthropic:  claude-opus-5
  #   openai:     gpt-5.6-luna (reasoning effort defaults to max)
  #   gemini:     gemini-3.8-flash
  #   openrouter: deepseek/deepseek-v4-pro-0813
  model: gemini-3.8-flash

  # Environment variable containing the API key
  api_key_env: GOOGLE_AI_STUDIO_API_KEY

  # Base URL for OpenAI-compatible endpoints (optional)
  # base_url: https://api.openai.com

  # OpenAI GPT-5-series Responses API reasoning effort
  # (default: max for the OpenAI provider)
  reasoning_effort: max

  # Optional LLM settings for individual target locales. An override using the
  # global provider inherits unspecified global settings. A different provider
  # uses that provider's defaults for unspecified settings.
  locale_overrides:
    yue:
      provider: openrouter
      model: deepseek/deepseek-v4-flash-0731
      api_key_env: OPENROUTER_API_KEY
    zh-CN:
      provider: openrouter
      model: deepseek/deepseek-v4-flash-0731
      api_key_env: OPENROUTER_API_KEY
    zh-TW:
      provider: openrouter
      model: deepseek/deepseek-v4-flash-0731
      api_key_env: OPENROUTER_API_KEY

# Keys per LLM call (default: 40)
batch_size: 40

# Parallel LLM calls (default: 4)
concurrency: 4

# Directory containing per-locale style guide Markdown files (default: style-guides)
style_guides_dir: style-guides

# Directory containing per-locale glossary JSON files (default: glossary)
glossary_dir: glossary

# Path to translation memory file (default: .internationalizer/tm.jsonl)
tm_path: .internationalizer/tm.jsonl

# Versioned source, policy, target, and provenance state
# (default: .internationalizer.lock; commit this file)
manifest_path: .internationalizer.lock

# Optional translation and strict-validation rules
validation:
  plural_style: i18next-v4 # generate and validate target-locale plural forms

Locale identifiers must be well-formed BCP 47 tags such as fr, pt-BR, or sr-Latn-RS. Canonical-equivalent target locales are rejected as duplicates, and locale-specific provider overrides match canonical-equivalent spelling. In the example above, locales without an override—including Japanese—inherit the global Gemini configuration.

File format and message syntax are separate settings. message_syntax accepts auto (the default), i18next, icu, or plain, globally and per bundle:

i18next protects {{name}}, nested paths such as {{user.name}}, escaping modifiers such as {{- name}}, formatting modifiers, and repeated placeholders. It enables i18next v4 locale-specific plural keys. Other braces are literal; {.sift,.claude,.codex,.agents} is not parsed as ICU. Custom interpolation delimiters and nesting expressions are not part of this profile.
icu always parses the message as ICU, including malformed input that cannot be recognized by automatic detection. Parsing errors never fall back to text.
plain treats braces as text and does not impose an interpolation grammar.
auto infers each message's grammar from its source. Select an explicit mode for catalogs mixing prose and code. Fluent resources require auto and use their own grammar; Markdown documents cannot select icu.

Validation, provider and TM output checks, adoption, pseudolocalization, and review approval share the selected syntax. HTML <code> contents and Markdown code spans are preserved exactly. Explicit syntax profiles enforce protected content even without --strict; --strict additionally checks translation quality and glossary rules. Source errors are reported once per bundle with source_path, blocked_by_source, and blocked_locales in JSON reports.

The syntax setting is part of the policy hash. This prompt-contract update makes previously recorded policies stale, including those using auto. Use translate --dry-run to inspect them, translate --refresh-policy to regenerate, or translate --adopt-existing to validate and record existing values under the new policy. Adoption leaves entries needing explicit review. Dry-run reports “Would translate” with planned keys and blocked job counts.

ICU MessageFormat values are parsed structurally. Simple arguments, select, plural, selectordinal, number, date, and time are supported, including nested messages, plural offsets, exact-number selectors, and #. Validation checks syntax, argument types and formatter styles, plural offsets, select branch identity, and target-locale CLDR plural categories. Provider output that breaks these invariants is rejected before a locale file or translation-memory record is written.

Fluent (.ftl) resources are handled as semantic source documents rather than flattened maps. Message values, terms, and attributes become independent units; comments are passed to the provider as developer context and included in source provenance. Serialization preserves resource comments and ordering. Validation protects variables, references, functions, selector defaults and branches, and data-l10n-name markup slots while allowing target-locale selector variants and natural reordering of named rich-text elements.

With i18next-v4, recognized source plural families are expanded during translation to the target locale's CLDR categories. A target-only category uses the source family's _other value as its translation template. Strict validation requires those target categories; source-only categories are optional for target locales that do not use them.

Style Guides

Style guides are Markdown files that get injected into the LLM translation prompt. They control tone, formality, typography, and other language-specific conventions.

style-guides/
  _conventions.md    # shared rules for all languages
  fr.md              # French-specific rules
  ja.md              # Japanese-specific rules
  ar.md              # Arabic-specific rules

Shared conventions (`_conventions.md`)

Define rules that apply to all languages: interpolation syntax, HTML preservation, string type conventions (buttons vs. labels vs. errors), etc.

Per-language guides (`{locale}.md`)

Define language-specific rules: formality register (tu vs. vous), punctuation (guillemets, inverted question marks), plural forms, date/number formatting, and a terminology glossary.

Style guides are durable policy inputs, not generated output. Internationalizer reads them but never rewrites them. Their content is hashed separately from the glossary and prompt contract, so an application code change does not make a translation stale. Editing a guide intentionally marks that locale for policy review; changing internal prompt wording does not, unless the prompt contract version also changes.

See examples/react-app/style-guides/ for a working example.

Glossary Format

Glossary files are JSON arrays stored in {glossary_dir}/{locale}.json:

[
  {
    "source": "Dashboard",
    "target": "Tableau de bord",
    "variants": ["Panneau de contrôle"],
    "enforcement": "error",
    "ignore_case": false,
    "whole_word": true
  }
]

variants lists other approved target forms. enforcement may be error, warning, or omitted for the default error behavior. Terms are injected into the LLM prompt as a terminology table, ensuring consistent translation across your application. An exact entry such as {"source":"API","target":"API"} also exempts that complete source-identical value from strict untranslated-value findings; it does not exempt a longer value merely containing API.

Translation Memory

Translation memory is stored as a JSONL file (one JSON record per line). Each record contains:

The bundle, key, source value, translated value, and canonical target locale
Source, style-guide, glossary, prompt-contract, and combined policy hashes
The provider and model that produced the translation
A timestamp

On subsequent runs, strings with the same source and policy hashes are served from the cache without calling the LLM. The default path is under the ignored .internationalizer/ directory, so it remains a local cache. Set tm_path to a tracked location if your project intentionally shares translation memory. The reviewable .internationalizer.lock manifest is versioned separately. Manifest schema v2 records provenance origin and review status independently, so “generated successfully” and “approved by a person” cannot be conflated.

Supported Formats

Format	Extensions	Mode
JSON	`.json`	Key-value (nested, dot-notation flattened)
YAML	`.yml`, `.yaml`	Key-value (preserves comments and ordering)
Markdown	`.md`, `.mdx`	Preamble and H2-level sections
Fluent	`.ftl`	Semantic messages, terms, attributes, comments, and selectors

Markdown targets contain invisible internationalizer:unit comments before H2 sections. These stable markers let Internationalizer add, move, or edit one source section without retranslating unrelated sections. Existing unmarked documents receive markers on their next successful update.

Project Type Detection

internationalizer detect identifies your i18n setup by checking:

Existing configured bundles and source-locale filenames/directories
Nested package.json dependencies and localization-module ICU references
Runtime evidence separately from file format, with uncertainty made explicit

Scanning is bounded and excludes dependency, build, hidden, and data directories. Dynamic plugin registration and unconventional catalog paths need explicit configuration; static detection cannot prove that ICU integration is absent.

Architecture

cmd/internationalizer/     CLI entry point and command definitions
internal/
  config/                  YAML config loading with defaults
  detect/                  Project type auto-detection
  fluentpattern/           Fluent pattern validation and safe text transforms
  formats/                 Format adapters (JSON, YAML, Markdown, Fluent)
  glossary/                Per-locale glossary management
  llm/                     LLM provider interface + implementations
    anthropic.go           Anthropic Claude backend
    openai.go              OpenAI / compatible backend
    gemini.go              Google Gemini via AI Studio backend
                           OpenRouter uses openai.go with custom base_url
  locale/                  BCP 47 identity and CLDR plural categories
  message/                 ICU MessageFormat parser and structural comparison
  policy/                  Stable translation-policy hashing
  pseudo/                  Provider-free accented and bidi test locales
  review/                  Explicit artifact approval workflow
  state/                   Versioned translation manifest
  styleguide/              Style guide loader
  tm/                      JSONL translation memory
  translate/               Translation orchestrator
  validate/                Locale validation and diffing

Comparison to Alternatives

Feature	Internationalizer	i18next	Crowdin	Generic LLM
LLM-powered translation	Yes	No	Partial	Yes
Per-language style guides	Yes	No	No	No
Glossary enforcement	Yes	No	Yes	No
Translation memory	Yes	No	Yes	No
CLI / local execution	Yes	N/A	No	Manual
Git-friendly files	Yes	Yes	Partial	Manual
No SaaS dependency	Yes	Yes	No	Varies
Open source (AGPL-3.0)	Yes	Yes	No	Varies

License

AGPL-3.0

See THIRD_PARTY_NOTICES.md for dependency notices.

Contributing

See CONTRIBUTING.md for development setup and guidelines. All contributions require DCO sign-off.

Navigation Menu

Repository files navigation

Internationalizer

Why Internationalizer?

Installation

npm Packages

Quick Start

Commands

`translate`

`pseudo`

`review`

`validate`

`detect`

`config plan` and `config apply`

`glossary`

`tm`

Configuration Reference

Style Guides

Shared conventions (`_conventions.md`)

Per-language guides (`{locale}.md`)

Glossary Format

Translation Memory

Supported Formats

Project Type Detection

Architecture

Comparison to Alternatives

License

Contributing

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
bin		bin
cmd/internationalizer		cmd/internationalizer
docs		docs
examples		examples
internal		internal
packages		packages
scripts		scripts
style-guides		style-guides
test		test
.gitignore		.gitignore
.golangci-lint-version		.golangci-lint-version
.golangci.yml		.golangci.yml
.internationalizer.example.yml		.internationalizer.example.yml
.internationalizer.yml		.internationalizer.yml
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
THIRD_PARTY_NOTICES.md		THIRD_PARTY_NOTICES.md
go.mod		go.mod
go.sum		go.sum
package.json		package.json

Latest commit

History

Folders and files

Repository files navigation

Internationalizer

Why Internationalizer?

Installation

npm Packages

Quick Start

Commands

translate

pseudo

review

validate

detect

config plan and config apply

glossary

tm

Configuration Reference

Style Guides

Shared conventions (_conventions.md)

Per-language guides ({locale}.md)

Glossary Format

Translation Memory

Supported Formats

Project Type Detection

Architecture

Comparison to Alternatives

License

Contributing

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

`translate`

`pseudo`

`review`

`validate`

`detect`

`config plan` and `config apply`

`glossary`

`tm`

Shared conventions (`_conventions.md`)

Per-language guides (`{locale}.md`)