What’s new in transglot

NewTwo native game-engine codecs ship. Unity Localization’s String Table CSV — the wide, multi-locale sheet its CSV extension exports — imports and exports directly, selecting the locale column per pull. Unreal Engine’s Portable Object (.po), the format the Localization Dashboard exports, round-trips keyed the way Unreal keys strings: the identity rides msgctxt and the source string rides msgid, so re-import maps to the right key instead of the source text. Godot’s gettext .po still works too. That brings native file formats to 20.

NewA Shopify connector opens the commerce segment: it localizes products, collections and pages both ways over the GraphQL Admin API using your own store’s access token. It reads each resource’s source fields (with their per-field digest) into keys, translates them, and registers reviewed locales back with translationsRegister — pinned to the current digest, into a separate secondary-locale record — so your primary (source) locale is never overwritten. Shopify is a $9/mo add-on.

NewA Markdown/MDX codec ships, so Markdown and MDX documents round-trip as a translation format — headings, paragraphs, lists, tables, and blockquotes are translated while front-matter, fenced code, and MDX/JSX blocks pass through verbatim. Long-form content and docs now localize alongside your app strings, bringing native file formats to 18.

NewFive more first-party CMS connectors join WordPress: Contentful, Intercom (Help Center), Zendesk Guide, Sanity, and Storyblok. Each syncs content both ways using your own provider credentials. WordPress and Contentful are included on every paid plan; Intercom and Zendesk Guide are $9/mo add-ons, and Sanity and Storyblok are $7/mo add-ons.NewA Figma plugin maps text layers to Figma Variables and modes over the deploy API, so design strings round-trip with transglot. It ships in the repository; it is not yet listed on the Figma Community.NewA GitHub App for continuous localization — extract strings on push, translate, and open a pull request with the results. It is built in-app and self-hosted: it stays inert until an operator registers the App and sets its keys, and shows that state until they do.ImprovedA Markdown/MDX codec remains on the roadmap.

NewSAML single sign-on now completes a real IdP login — dedicated login, assertion-consumer (ACS), and metadata endpoints on the onelogin/php-saml toolkit, with just-in-time user provisioning and an enforce-SSO option for verified domains. Available on Enterprise, alongside the existing SCIM 2.0 provisioning and append-only audit log.NewA public Trust Center — subprocessor list, DPA, and an explicit account of how the exact-match (global, anonymous) and semantic (org-scoped) translation memories handle your data.NewAn OSS program: a granted free tier for qualifying open-source projects, a public application form, and a “localized with transglot” README shield you can embed in any repo.ImprovedThe Trust Center states our compliance posture honestly: SOC 2 Type II and ISO are on the roadmap, not yet certified.

NewA first-party WordPress connector — sync translatable content out of WordPress into a project and finished translations back in, with no manual export. Available on every paid plan.ImprovedWordPress is the first CMS connector; other CMS and commerce connectors, and a Markdown/MDX codec, remain on the roadmap.

NewPublish reviewed translations to a public, edge-cacheable OTA/CDN endpoint — immutable versioned URLs plus a “latest” URL that revalidates cheaply with ETag and 304 — so you can fix a production string without a redeploy or an app-store release.New@transglot/runtime, a zero-dependency runtime i18n SDK that reads strings over the CDN, with framework adapters @transglot/vue, @transglot/react-i18n, and @transglot/svelte.NewA per-project freshness/publish policy and scoped, read-only CDN keys (taicdn_), rate-limited per key and IP. Runtime delivery is available on every paid plan.

NewA per-project AI Profile — capture your brand voice, tone, and worked examples once, and every AI translation for that project is grounded in them, on top of the glossary and inferred style guide. It folds into the same cached prompt, so it adds no latency to ordinary runs. Team plans and up.NewSource-drift self-heal — when a source string changes, only the affected translations are re-QA’d and, per your project policy, re-healed, with a single coalesced source_drift.repaired webhook and notification. Human-edited and reviewed rows are still never overwritten.

NewConfidence-gated auto-approve — the translation agent self-reports a per-string confidence at translate time, and only structurally-clean, high-confidence rows auto-approve; everything uncertain routes to a human. The deterministic validator and QA pre-gate always win, so a confident-but-broken row still goes to review.NewA per-language MQM scorecard, exportable as CSV — score trend, top finding categories, and the share of rows still needing review — a procurement- and renewal-ready quality artifact, built from indexed reads at no extra AI spend.NewA golden-set quality-eval / regression harness (eval:run) that re-translates a committed corpus with the current prompt and scores it on the deterministic tier plus an optional LLM-as-judge tier, so prompt and model changes ship without quality roulette.

NewSemantic translation memory — pgvector embeddings (text-embedding-3-small) retrieve the nearest human-reviewed translation as a $0 suggestion and as few-shot grounding for the AI. Unlike the global exact-match pool, the semantic path is strictly scoped to your own organization — the hard tenant boundary — and reads only reviewed rows.NewFuzzy matches are never auto-written: they surface as suggestions in the editor rail and ground the first pass, but a human always decides. Team plans and up, with a per-project toggle.

NewAssign specific locales or keys to an outside translator, track each assignment from open through submitted, and gate approval — built on the existing per-language grants, review requests, and notifications.NewA two-person integrity split: the reviewer who approves an assignment can never be its assignee. External linguists see exactly their assigned scope and nothing else.

NewXLIFF 2.0 import/export — the modern OASIS unit/segment model, alongside the existing XLIFF 1.2 support, for current CAT/TMS tools.NewJava .properties import/export — key=value resource bundles with Unicode escapes, for Java, Spring, and Kotlin projects.New.NET .resx import/export — the C#/.NET XML resource format, for desktop, web, and MAUI apps. That brings the translation file formats to 17 in all.

NewCSV import/export (RFC 4180) — a key/value/description table for game string tables and spreadsheet-driven teams, with plurals carried as key[one] / key[other] rows and a formula-injection guard on export.NewXLSX import/export — the same columns as an Excel or Google Sheets workbook, so you can hand a sheet straight to a translator and read it back in. That takes the translation file formats to 14.

NewXLIFF 1.2 import/export — the OASIS CAT/TMS interchange standard, bilingual (<source> + <target>), XXE-hardened, with inline-markup units skipped loudly instead of flattened.NewApple `.xcstrings` (String Catalog) import/export, including plural variations, for Xcode 15+ projects.Newgettext `.po` import/export with msgctxt, plural forms mapped through each file’s `Plural-Forms` expression, and `msgfmt --check`-clean output.NewRails-style YAML import/export — locale-rooted nested keys, anchors accepted (with expansion budgets), hyphenated `pt-BR:` root keys.ImprovedBilingual pulls now report keys whose source-language text is missing in an `X-Missing-Source-Keys` header instead of silently substituting target text into `<source>`/`msgid_plural`.

NewInfer a draft style guide per locale — formality, tone, brand voice, regional notes, and do/don’t examples — from your team’s human-reviewed translations. Requires at least 5 reviewed rows for that locale.NewStyle guides are never auto-applied: review the proposal, edit it, and save it explicitly before it affects anything.ImprovedA saved style guide folds into the same cached translation prompt used for every batch, so it doesn’t add latency or cost to ordinary translation runs.ImprovedStyle inference and Translation QA now share the same Team-plan quota fail-fast check, so a project that’s out of AI quota gets a clear error instead of a partial run.

NewPer-project auto-heal modes for QA-flagged machine translations: Off, Suggest (default — proposes a fixed value for a human to accept, writes nothing), and Apply (re-translates the row, validates it, and writes a reversible version).NewApplied fixes are versioned and revertible, and are authored distinctly (`qa:autoheal:<model>`) so it’s always clear a fix was machine-applied.FixedHuman-edited and reviewed rows are now provably never overwritten by auto-heal — enforced by a compare-and-swap guard under a row lock, not just an application-level check.

NewTranslation QA scores your target strings 0–100 against accuracy, fluency, terminology, style, placeholder, plural, and length categories, with per-finding severity and suggested fixes where available.NewA free deterministic pre-gate catches placeholder mismatches, incomplete plural categories, and length overruns before any AI call runs — structurally broken rows are flagged at zero AI cost.NewRun QA over missing, stale, selected, or all rows from the web UI or the REST API (`qa/run`), fanned out in ~30-row chunks.ImprovedRe-running QA on an unchanged project does zero AI work — a row is skipped whenever its source hasn’t moved since it was last checked.

NewGlossaries with per-term case-sensitivity, do-not-translate, forbidden markers, part of speech, notes, and per-locale target renderings, plus a pending/active/rejected review lifecycle.NewBoth a per-project default glossary and an org-shared glossary you can link across many projects — on a naming collision, the project-level term always wins.NewDeterministic glossary enforcement (off, warn, or enforce) runs on every AI translation via a purpose-built Aho–Corasick matcher — a forbidden term in the output always fails the check; a missing mandated term is a warning or a hard failure depending on your setting.NewCSV and TBX (ISO 30042) import/export for moving glossaries in and out of transglot — these are glossary-interchange formats, distinct from the supported translation file formats.NewAI-assisted term mining proposes candidate glossary terms from your source strings as "pending" — a human always approves or rejects before a term goes live.

NewThe `transglot` CLI (npm, zero runtime deps) — `init`, `pull`, `push`, and `status`. Push source files, wait on the AI batch, and pull translations back from any terminal or CI job. Run it with `npx transglot` or `npm i -g transglot`, authed with a `tgl_…` project token.NewA first-party GitHub Action (`.github/actions/transglot`) — pushes source files, optionally waits for the AI batch, pulls translations back, and commits them to a branch or uploads them as an artifact. Inputs: `token`, `url`, `wait`.New`@transglot/react` — a React adapter for the in-context editor: `<TransglotProvider>`, `useIncontextT`, and a Shadow-DOM `<EditDialog>` over the core `@transglot/incontext` SDK.New`@transglot/incontext` — the zero-dependency runtime SDK behind the in-context overlay (watermarking, ALT+click, and the scoped-session API client), published for direct use.

NewFour billing plans — Free, Indie, Team, and Business — with per-plan word limits, seats, and project caps.NewPromo codes for discounted or extended plans at checkout.ImprovedVirtualized editor grid now reconciles plural-form saves row-by-row instead of refetching the whole page, so multi-locale plural edits land instantly.ImprovedOptimistic cell saves (Cmd/Ctrl+Enter) now retry with the server-reconciled value on conflict instead of silently dropping the edit.

NewOrganization roles — owner, admin, developer, translator — with per-project policies gating who can edit, translate, or manage settings.NewTeam invitations by email, scoped to an organization and a starting role.NewProject access tokens (`tgl_…`, sha256-stored, reveal-once) scoped to push, pull, or manage abilities, with optional expiry — for the REST API and webhook automation.ImprovedAPI requests are now rate-limited per token to keep batch jobs from starving interactive traffic.

NewPer-project webhooks for `batch.completed` and `batch.failed`, signed with HMAC-SHA256 so your endpoint can verify authenticity.NewA delivery log (last 50 attempts) showing status, response code, and timing for every webhook fired.NewManual redelivery — replay a webhook with its original, frozen payload straight from the delivery log.FixedBatch-failed webhooks now fire for partial-chunk failures, not only whole-batch failures.

NewQuality Mode (Team plan and above) — forces Claude Sonnet 5 with a corrective retry pass, with an optional per-locale whitelist.NewTranslation memory savings tracking — see words and quota reused instead of re-translated, per project.ImprovedTranslation memory matching now keys on source hash, locale, and model, so a memory hit only reuses translations from an equivalent context.ImprovedAuto-backfill runs automatically when a new language is added to a project, instead of requiring a manual translate-missing pass.

New`/api/v1` REST API — pull (export), push (import, with merge or sync-removals), translate (trigger an AI batch), and batches (status and retry).NewBearer authentication via project access tokens, so CI jobs and scripts can pull and push translations without a browser session.ImprovedExport responses now include an `X-Skipped-Keys` header when a target format can’t round-trip plural forms, so automation can detect lossy exports.

NewFlutter `.arb` and Laravel PHP / Laravel JSON added to the import/export formats, alongside JSON (flat and nested), Android XML, iOS `.strings`, and iOS `.stringsdict` — 8 formats total.ImprovedFormat handling is now capability-aware: simple key-value formats skip plural forms cleanly instead of erroring on export.FixediOS `.stringsdict` export no longer drops the `few` category for locales that define it.

NewAI translation powered by Claude (Haiku 4.5 by default), run as a batch pipeline of roughly 30 keys per chunk with structured JSON output that preserves placeholders and tokens.NewTranslate-missing, translate-selected, and retry-failed actions, plus auto-translate whenever a source string is edited.NewTranslation memory — exact-match reuse by source hash and locale, applied before an AI call so repeated strings cost zero quota.ImprovedAI usage is now metered per organization against the plan’s monthly word quota.

NewRoving-tabindex keyboard navigation across the translation grid — move, edit, and save without leaving the keyboard.NewOptimistic cell saves (Cmd/Ctrl+Enter) with per-row edit locks, so concurrent editors don’t clobber each other’s work.ImprovedFuzzy search (Postgres `pg_trgm`) is now typo-tolerant across keys and source strings.FixedFiltering by "untranslated" no longer loses your scroll position when a save comes back.

NewThe translation grid is now virtualized, keeping scrolling smooth on projects with tens of thousands of keys.NewKeyset-cursor pagination in 100-row windows, replacing offset pagination for large tables.NewPlural forms editing per CLDR category, with RTL layout support for right-to-left locales.

NewProjects, languages, and a translation editor grid for managing multilingual strings.NewOrganizations with member management, ready for roles and seats.NewJSON import and export to get existing i18next, React, and Vue projects into transglot quickly.