Claude Translator

Fair question, and it is checkable rather than something you have to take our word for. doctranslator.com/fr is a large Astro site whose French pages were built by this pipeline. Compare it against the English original and the markup is byte-identical:

Nothing structural moved, so there is nothing for the browser to lay out differently. Check it yourself — it takes about ten seconds:

curl -s https://doctranslator.com/   | grep -o 'class="[^"]*"' > en.txt
curl -s https://doctranslator.com/fr | grep -o 'class="[^"]*"' > fr.txt
diff en.txt fr.txt && echo "markup identical — only the text changed"

And the scores

PageSpeed Insights performance scores, measured 25 August 2026. Accessibility, best practices and SEO were 100 on every run of both pages, in both strategies.

These are medians, deliberately. Lighthouse scores move: five consecutive runs of the same English mobile page returned 98, 98, 88, 98, 98, with largest-contentful paint swinging between 1.8s and 3.2s. That is network and CDN variance rather than page quality. If you measure once and get 91, that is why — run it a few times and compare the two languages against each other rather than against a number on a marketing page.

And the language

Markup identity is measurable. Whether the Spanish is any good is a different question, and until 2.0 the pipeline had no answer to it — the tests checked request shapes and scaffolding, and nothing scored a translation.

node scripts/i18n/tqa.mjs --lang es --dry      # sample size and cost, no API call
node scripts/i18n/tqa.mjs --lang es,fr,de     # scorecard + per-locale JSON
node scripts/i18n/tqa.mjs --lang es --repeat  # judge the same sample twice

It scores a stratified sample on the MQM error typology at its conventional weights — minor 1, major 5, critical 10 — and weights the sample by how often each string appears, so the header everyone reads counts for more than a one-off footnote. Three things keep the number from being decorative:

The judge is a different model from the translator

Models score their own output generously. If no second provider key is configured it uses the same one and says so, in the run and in the report.

The sample is seeded, and the variance is reported

A seed reproduces a score exactly; --repeat scores the same sample twice and prints the gap. A quality number nobody can re-derive, quoted without its noise, is a marketing number.

A unit the judge cannot assess is excluded, not counted as clean

An early version printed 100.00 out of 100 from a sample where every single unit had failed to parse. It now refuses to report a score at all in that case.

Read it as a comparison — between locales, between models, before and after a prompt change — not as a grade. It is one model’s assessment of another’s work rather than a human review, and the report says so on its face.

Writing 1,234.56 to a German reader is one of the most visible marks of a machine translation. Models are unreliable at separator conventions, so the model is told to leave numbers alone and Intl reformats them afterwards, deterministically.

Deliberately left alone: version numbers, times, IP addresses, ISO dates, phone numbers and any ungrouped number — during development a greedy pattern turned 192.168.1.1 into 1.921.681,1, and that case is now a test. One result that looks like a bug and is not: Spanish does not group four-digit numbers, so 1,234.50 is correctly 1234,50 in es and 1.234,50 in de.

You need

• Node.js 20 or newer
• A site that builds to static HTML — it reads your build output, not your source
• An API key for whichever model you choose — or none at all, if you run one locally

As a Claude Code skill

Clone it into your skills directory, then ask Claude to localize the site. It follows the bundled instructions, including the failure modes that cost us real money to find — and the rules for when this is the wrong tool entirely.

git clone https://github.com/ConveyThis/claude-translator.git \
  ~/.claude/skills/claude-translator

In any project

One command. init copies the pipeline into scripts/i18n/, writes a config, declares parse5 and adds the derived i18n/ paths to your .gitignore. It never overwrites anything without --force, and it prints every file it touched.

cd your-project
npx claude-translator init
npm install                       # parse5, the only dependency

The scripts land in your repository rather than staying in node_modules on purpose — they resolve paths from the project they sit in, they are short enough to read, and the licence exists so you can change them.

git clone https://github.com/ConveyThis/claude-translator.git
cp -r claude-translator/scripts  your-project/scripts/i18n
cp claude-translator/i18n.config.example.json your-project/i18n.config.json
cd your-project && npm install --save-dev parse5

Configure

The smallest config that runs. buildDir is where your generator writes HTML; baseUrl is your canonical origin, no trailing slash.

{
  "buildDir": "dist",
  "baseUrl": "https://example.com",
  "provider": "anthropic",
  "glossary": "glossary.json",
  "locales": [
    { "hreflang": "es", "pathCode": "es", "nativeLabel": "Español" },
    { "hreflang": "de", "pathCode": "de", "nativeLabel": "Deutsch" }
  ]
}

echo "ANTHROPIC_API_KEY=your-key-here" >> .env   # and gitignore it

Run it

Then deploy the build directory exactly as you deploy it today — the localized pages are ordinary files sitting next to your existing ones. Full options, every failure mode, and the reasoning behind each design decision are in the README.

Six real limits. None of them is a crippled feature — they are the shape of the approach, and the scripts tell you when you hit one instead of failing quietly.

Client-side hydration

Islands and framework payloads re-render in the browser, over the substituted HTML. The extractor counts the affected pages so you find out in seconds, not after a full run.

Documents

Linked PDFs, DOCX and XLSX stay in the source language — the pipeline only ever touches HTML.

Churn

The memory is keyed by source hash, so it can tell you what share of your site changed since the last run. High churn means paying to re-translate, repeatedly.

Editing a translation

Find the hash in the memory file, edit the string, rebuild. There is no editor, no reviewer and no workflow.

Volume

You pay your own model provider directly, at their rate, with your own key — or nothing at all, if you run the model yourself.

Modified network use

AGPL-3.0 §13 obliges you to publish modifications if you run a changed copy as a service for other people. Running it unmodified, or modified internally, is unrestricted.

The first four are the ones people actually hit, and they are all the same shape: something has to happen after the build, which a build-time tool cannot do. That is what the managed product is for, and the free plan is 5,000 words with no card, which is enough to find out whether it helps. You do not need it to use anything on this page.