Open source · MIT · Bring your own LLM key
Your coding agent moves hardcoded strings to keys and translates them. glot checks every LLM output with deterministic code and refuses anything broken or too long, with the exact reason, until the retry is valid.
Detects your i18n stack, extracts strings, translates, checks. Run it again after every UI change.
Benchmark, October 18, 2026
320 ICU messages per model and condition, 8 locales. Same prompts, same models.
| Model | Raw prompt | With glot |
|---|---|---|
| Claude Haiku 5.5 | 14.1% | 0% |
| Claude Sonnet 5.5 | 8.8% | 0% |
How it works
Or all three at once, idempotent: npx glot-i18n localize.
Finds your i18n library and catalogs, writes glot.toml and the rules your coding agent reads.
$ npx glot-i18n init
next-intl · messages/{locale}.json
wrote glot.toml, CLAUDE.md rules
Moves hardcoded UI strings to keys and rewrites the components for next-intl or react-i18next.
$ npx glot-i18n i18nize
42 strings → keys in 9 components
app/checkout/page.tsx +12 −12
Only the delta. Each output is checked; a refusal goes back to the model with the reason. Still invalid after retries: nothing is written.
$ npx glot-i18n translate
✗ refused pl inbox.unread: missing few, many
✓ fixed pl attempt 2 · check passed
The guarantees
Every card shows a real refusal, word for word as glot prints it.
Same set as the source, none invented, none translated. ICU, i18next and MessageFormat 2.
fr "Bon retour, {nom} !"
Refused by glot: Placeholder `{nom}` does not exist in the source; remove it or fix its name (the source only has {name}).
Every category the target locale needs: four for Polish, six for Arabic.
pl "{count, plural, one {…} other {…}}"
Refused by glot: `{count, plural, …}` is missing plural categories required by pl: few, many.
glossary.toml in your repo: required and forbidden translations, do-not-translate terms.
fr "Ouvrez votre espace"
Refused by glot: Glossary: `Workspace` must be translated as `Espace de travail` in fr.
Measured with your app's real font and box. Too wide is refused, never truncated.
de "Zahlungserinnerung an den Kunden senden"
Refused by glot: Too wide for its box in the app: one line of 120px at Geist 14px/400; it renders 276px wide.
glot.lock hashes source and target. Only the delta is translated; hand edits are never overwritten.
$ glot translate --locales fr
Refused by glot: 1 edited by hand
glot check runs every rule offline on every catalog and exits 1 with PR annotations.
$ glot check --format github
Refused by glot: ::error file=messages/de.json,line=12,title=glot tag_mismatch::[de] home.cta: Markup tags differ from the source: missing `<b>`
Pixel budget
German runs a third longer than English. glot fit renders your running app in each locale, measures every text against its box, asks for a shorter variant that keeps the meaning and placeholders, and renders again.
"Zahlungserinnerung senden"
"Mahnung" · claude-haiku-5-5 · $0.0002
Real run on the demo app, German locale.
Try it, nothing to install
Edit the translation. The same validator that guards glot translate, compiled to WebAssembly, runs in your browser on every keystroke.
Start from a broken translation
Loading the validator…
No refusals yet. Try one of the challenges.
Pixel width and glossary need your app and your repo: try them with the CLI.
Agents
glot mcp gives your agent your keys, where they are used, and the glossary. Every string it adds goes through the same validator.
In-app editing
The dev overlay finds the key behind the text you clicked, shows it in every locale with live validator issues, and saves to your catalog files.
Where glot differs from hosted translation platforms and other developer-first tools.
| Feature | glot | Lokalise | Crowdin | Lingo.dev | Tolgee |
|---|---|---|---|---|---|
| Price | Free (MIT) + your LLM usage | From $149/mo, 14-day trial | Paid plans after a 14-day trial, free for open source | Free CLI with your own key, hosted engine $0 base + usage | Free tier and free self-host, cloud from €58/mo |
| Open source | |||||
| Bring your own LLM key | |||||
| Deterministic validation of every string | |||||
| Git as source of truth (lockfile) | |||||
| MCP server for coding agents | |||||
| In-context editing in your app | |||||
| Rendered pixel-budget fitting |
Checked against each vendor's public website and docs on October 18, 2026. Partial means limited, plan-dependent or through an integration. No means we found no such feature in public docs. Plans change: every cell is linked to its source.
glot is MIT and free. You pay your LLM provider with your own key: 850 strings cost $0.035 with the default model in our run. glot translate --dry-run estimates first, and a budget caps it.
No. Your app keeps next-intl, i18next or FormatJS with plain JSON catalogs. glot works on the files, in your repo and in CI.
glot sends the exact reason back and retries twice, then tries the fallback model. If it still fails, nothing is written: the key is reported as rejected and stays untranslated.
The strings to translate and their context go to your LLM provider. Not your source files. Nothing goes to glot's authors; there is no telemetry.
No. The CLI and the MCP server do the whole job locally. The app adds a shared dashboard, review and GitHub sync for teams.
One command in your repo. Nothing broken reaches your catalogs.