AI machine translation for developers

Translation that respects your placeholders

Every plan translates with Claude Haiku 4.5 by default. Quality Mode, available on Team and Business, forces Claude Sonnet 5 with a corrective retry pass, and an optional per-locale whitelist scopes that extra care to the strings that matter most.

Translations are generated in batches of roughly 30 keys as structured JSON, so variables and tokens like {count} or %{name} come back intact instead of being rewritten as free text.

Edit a source string and transglot re-translates it into every target locale automatically — there is no separate sync step. Add a new language and every existing key is backfilled with a first-pass translation, so you start from a full grid rather than a blank one.

Exact matches — keyed by source hash, locale, and model — are reused from translation memory at no quota cost, with savings tracking so you can see the words and cost it saved. Translation memory is a shared, anonymous pool across the platform rather than a private per-organization store, and it is a per-project toggle if you would rather not use it. Everything else is metered per organization against your plan’s monthly word limit, with overage billed at $0.003 per word on paid plans and a hard cap on Free.

What AI translation gives you

• Claude Haiku by default, Sonnet in Quality Mode — every plan translates with Claude Haiku 4.5; Quality Mode (Team and Business) forces Claude Sonnet 5 with a corrective retry pass and an optional per-locale whitelist.
• Placeholder-preserving structured output — translations are generated in batches of roughly 30 keys as structured JSON, so variables and tokens survive the round-trip.
• Auto-translate on source edit — edit a source string and transglot re-translates it into every target locale, with no separate sync step.
• Auto-backfill on new language — add a language to a project and every existing key is automatically backfilled with a first-pass translation.
• Translation memory — exact matches (by source hash, locale, and model) are reused instantly at no quota cost, with savings tracking. It is a shared, anonymous pool rather than a private per-organization one: entries carry no project or organization, so a match can come from anywhere on the platform. Turn it off for a project with a single toggle.
• Semantic (fuzzy) translation memory — on Team plans and up, pgvector embeddings retrieve the nearest human-reviewed translation from your own organization and surface it as a $0 suggestion and as few-shot grounding for the AI. Unlike the exact-match pool, this fuzzy path is scoped to your organization — the hard tenant boundary — reads only reviewed rows, and is never auto-written.
• Per-project AI Profile — capture your brand voice, tone, and a few worked examples once (Team and up) and every AI translation for that project is grounded in them, on top of your glossary terms and the inferred style guide.
• Quota-metered, with retry — AI usage is metered per organization against your plan’s monthly word limit, with a retry-failed action for any batch that didn’t complete.

From source string to shipped translation

• Chunk and structure the batch — keys are grouped into batches of roughly 30 and sent to Claude as structured JSON, so the model returns machine-readable output instead of free text that can drift or drop tokens.
• Translate with placeholders inventoried — every placeholder and token in the source ({count}, %{name}, positional args) is inventoried and expected back in the target.
• Validate, and retry once if needed — each chunk is validated for placeholder integrity and, in Quality Mode, a corrective retry re-runs a failing row on Claude Sonnet 5 with the reason attached.
• Reuse and meter — exact matches are served from translation memory at zero quota cost; everything else is metered per organization against your plan’s monthly word limit.

Compared with generic machine translation

• Placeholders — free-text MT rewrites {count} or %{name} into words and breaks interpolation at runtime; structured JSON inventories every placeholder and expects it back.
• Plurals — ICU plural syntax gets flattened or corrupted elsewhere; here each CLDR plural category is translated in place and validated for completeness.
• Keeping in sync — someone re-exports, re-uploads, and re-runs MT every time a source string changes; here editing a source auto-translates it and adding a language auto-backfills every key.
• Repeat strings — you pay to translate the same string again in every project and export; here exact matches are reused from translation memory at zero quota cost, with savings tracked.
• Quality control — one generic model and no second pass means you find the bad strings in production; Quality Mode forces Claude Sonnet 5 with a corrective retry on the strings that matter.

The details that keep translations trustworthy

Beyond the first pass, transglot is built so machine output stays safe to ship: metered, reusable, and always reviewable in the same grid.

• Quality Mode, scoped to what matters — Team and Business plans force Claude Sonnet 5 with a single corrective retry, and you can whitelist the specific locales where the extra care is worth it.
• Translation memory, honestly described — exact-match reuse is keyed by source hash, locale, and model, and tracks the words and cost it saved. The pool is global and anonymous, not scoped to your organization, so if that is not the trade you want, switch it off per project with use_translation_memory.
• Metered per organization — AI usage counts against your plan’s monthly word limit, with overage billed at $0.003 per word on paid plans and a hard cap on Free.
• Retry without re-doing everything — if a batch doesn’t complete, a retry-failed action re-runs only what’s left, so you never re-translate rows that already succeeded.
• Reviewed in the same editor — every AI draft lands in the keyboard-first grid, so a human can confirm or adjust it without switching tools or exporting a file.