What's new.

2026-10-12: Formats, connectors and agent safety#

A 0.10.0 update that widens what Localization OS can read and connect to, and tightens how AI agents drive it.

• 87 file formats under 151 extensions, up from 11: OpenDocument, PowerPoint and Visio, InCopy, FrameMaker, QuarkXPress Tags, DITA and DocBook, MadCap Flare, HTML Help, EPUB, email, wiki markup, AsciiDoc, reStructuredText and LaTeX, software resources from Android and Apple to Qt, Flutter and Fluent, broadcast captions, and bilingual files from Trados, memoQ, Phrase, Wordfast and WorldServer, including Trados Studio packages. See File formats.
• 28 connector kinds, including GitHub, GitLab and Bitbucket, Microsoft 365 files, Google Docs, ServiceNow, Sitecore, Sanity, Plytix and Drupal, with rebuilt help-center connectors. Seven are live-verified and the rest are beta; see the Connector index.
• Byte-identical exports: an untranslated file comes back exactly as uploaded, and translated files keep their line endings, byte-order mark and, where the text fits, their encoding. Legacy files are read in the source language's code page. See Encodings.
• Review holds for ambiguous dates, changed line breaks and words joined at a formatting change; named placeholders may now move; missing plural forms are filled with the plural translation; abbreviation-aware sentence splitting for new documents. See QA checks.
• Office documents: a per-project Translate comments option, right-to-left layout for spreadsheets, InDesign and SVG, and an Excel export that writes into your own workbook.
• Terminology and memory: MultiTerm XML termbase import, Trados .sdltm memory import, spreadsheet imports, and resource formats in key-based projects.
• MCP: tool annotations and structured output, external content marked as data, background jobs and idempotency keys, tools filtered by the key's scopes, spend confirmations bound to a plan, local file reads limited to allowed folders, and opt-in OAuth sign-in for remote MCP clients. See MCP for agents.
• Connector scope: the Jira query and the Confluence space and query are now admin-only once a credential is stored.

2026-10-08: Localization OS 0.10.0#

A release that changes how the product looks, how machine translation speaks, and how much the install can take at once. The full notes are on What's new in 0.10.0.

• The Console look in light and dark, with Light, Dark or Match my computer stored with your account. Built to WCAG 2.2 Level AA in both themes.
• MT formality per target language, as a customer default, on a project, or on one document, with language-specific registers such as Sie or du.
• Repeat propagation: approving a segment fills its open repeats in the same document, checked, badged, and undoable.
• Locked, editable prompts on a tabbed workflow page, with Apply and Restore default.
• Review holds and a TM changed badge in the editor, and global search across projects, documents, memory, glossaries, workflows and connections.
• Chunked TMX and glossary imports, faster audit writes, and per-engine parallel jobs.
• Security and privacy: audit key rotation, refusal quarantine, fingerprint-only agent transcripts, an exact erasure scope, two-step sign-in, expiring API keys, and dependency and native-library scanning before every merge.

2026-09-22: Quality judges and auto-fix#

A workflow's QE gate can now be scored by a typed quality judge, and failing segments can repair themselves before they reach a person.

• Quality judges in Setup AI: two new engine kinds in their own Quality judges group. Jev (TypeSafe, hosted) takes your TypeSafe API key, talks only to TypeSafe's fixed address, and is priced and counted against the spend cap. Jev-compatible local server connects any server that speaks the Jev System One protocol, such as SemIf or Kev, on the same machine or over Tailscale, with an optional key, at no cost. Each has a Test button, and a judge appears only in a step's QE judge picker. An install that already has a TYPESAFE_API_KEY can turn it into an engine with one click. See Quality judges.
• Auto-fix failures: an opt-in setting on a Translate step whose QE judge is a Jev judge. An LLM post-editor rewrites a failing segment from the judge's scores and the judge re-scores it, for 1 to 5 rounds, with an optional cross-check. A failing segment the judge is unsure about is never auto-fixed, segments that still fail go to review with every round's score and text, and a fully local loop costs nothing. See Workflows.
• Inspect a judge: a page per judge engine showing the questions it is asked, its decision rule, a Try it form that scores one segment and shows the request, the raw answer, the verdict and the exact post-editor prompt, and a link to the LLM Ledger filtered to that engine.
• Command line, API and MCP: engine inspect, engine jev-try and engine import-jev-env, the auto-fix flags on workflow add-step and edit-step, lqa jev-refine --judge, and matching API routes and MCP actions.

2026-08-31: Localization OS publicly launched#

Localization OS is now publicly available. Highlights of the product at launch:

• Multi-agent LLM translation workflows: configurable pipelines of translate, quality-check, human review, TM-match, and external vendor steps, with conditional branching and bounded correction loops.
• Translation memory: every approved translation is remembered and reused, including a TM-match step that fills and locks exact matches so the rest of the pipeline skips them.
• Glossaries: a terminology authority of required and do-not-translate terms that is enforced in the output, not just suggested.
• LLM-based LQA: quality scoring and rulebook checks that feed a workflow's branching and gates.
• A post-editing CAT editor: inline review and correction for the slice of content that needs a human, with edits written back to translation memory.
• Memory: a per-customer notebook of terminology decisions and style rules, some written by hand and some learned from corrections.
• Key-based projects: continuous localization for software strings, with a change feed built for CI.
• Connectors: push-model integrations with systems including Airtable, Jira, Strapi, Contentful, Shopify, HubSpot, Akeneo, and WordPress.
• The Evaluations Lab: side-by-side comparison of translation setups, with statistics that refuse to call a winner when the evidence is thin.
• Image localization (beta): detect text in an image, translate it through the ordinary pipeline, and re-render it in another language.
• Vendors: first-class external translation and review vendors, with rate cards, an optional spend-approval gate, and a reject-and-rework loop.
• Reports and content routing: cost, token, volume, and quality reporting over existing runs, and rules that place inbound content automatically.
• Web and CLI parity: every feature ships on both the web UI and the CLI, with the API and MCP surfaces mirroring it where applicable, for automation and CI.
• Self-hosted by design: Docker or plain Python, SQLite by default with optional PostgreSQL, and enterprise identity, audit, and encryption controls.