Localization OS
Localization OS is a self-hosted, MT-first translation management system. You get the whole repo: host it, modify it, run it on your own servers.
Your AI, your hardware. No per-word fees.
Schedule a technical walkthrough → or email brent@patoislabs.com
A real run, uncut: TM match, translation, LQA, post-edit.
The problem
Somebody asked it in a meeting: "Why are we paying this much for a cloud TMS? Couldn't we build our own?" And the conversation died in that meeting.
Localization OS is the missing starting point: a working, self-hosted TMS you could run internally tomorrow, and shape into whatever your team needs.
The model
You get the entire codebase: host it, connect your models, change anything. Self-hosting means no per-word fees. Your data never leaves your building, because there's no "our side" for it to go to. Commercial terms are set in each customer's license agreement; talk to us about yours.
And if Patois Labs disappears tomorrow, nothing turns off. The software and the source are already running on your servers.
The product
LLMs translate everything; humans handle the exceptions. Whole documents in one pass, not segment by segment. Automated LQA speaks MQM, and its findings feed the post-edit pass. Describe a workflow in a sentence; the assistant builds it.
Runs entirely on a laptop. Point it at a frontier API, your approved models, or the GPU under your desk. Zero code changes. TM, glossaries, style guides, and an LLM-native memory layer feed every prompt. Fully offline on local models, or through the APIs your company has approved: that traffic follows your configuration, nobody else's.
A modular monolith with named ports: one process by default, your infrastructure by choice. Disable any feature and it's gone. Everything is an API first, with an MCP server on top, so your agents can drive the whole system. Deploys in minutes: Docker or plain Python, SQLite by default, no Kubernetes required.
Your data
Most TMS APIs give your ML team a keyhole. This one records every LLM call and every human correction, analysis-ready for your warehouse. It doesn't just do the work. It produces the dataset your next model trains on. Your content, your corrections, your dataset, under your existing vendor agreements.
The LLM Ledger: every request, every payload, and the export that follows.
Ready for your IT review
SAML SSO · SCIM provisioning · Role-based access control · Full audit trail · Encrypted at rest · Built to WCAG 2.2 AA · Web UI and CLI in feature parity
SSO and provisioning are interop-tested against a real Okta org; every action lands in a tamper-evident audit chain; a model refusal is quarantined for review, never stored as a translation.
What's inside
Exact and fuzzy leverage, applied where it counts: inside the prompt. Standard TMX imports on day one, in chunks, so a big file never stalls everyone else.
Enforced in the translation, not displayed beside it. Standard TBX imports too.
Sie or du, vous or tu, polite or honorific: set it per customer, project, or document, for each target language.
RAG across everything you feed it: TM, terminology, style guides, project notes.
Screenshots ride along with strings, so the model sees the UI it is translating.
One-off document jobs or continuous, key-based localization. Both first-class.
Full context in one pass. Segment mode when you want it.
MQM findings from an LLM reviewer, fed straight into the post-edit pass.
Describe the pipeline you want; the assistant assembles it; you hit run.
A real CAT editor for the strings that need a human. Approve a sentence once and its repeats follow.
A monthly spend cap, department budgets, and a kill switch. Models on your own network cost nothing.
Every feature is an API first, with an 84-tool MCP server on top. Web and CLI in parity.
Office and OpenDocument, InDesign and InCopy, bilingual files from Trados, memoQ and Phrase, subtitles, help systems, and the resource files your developers ship. 150 file extensions in all.
A file with nothing to translate returns byte for byte. Translated files keep their line endings and encoding wherever the text allows, and spreadsheets, InDesign and SVG files get right-to-left layouts for Arabic, Hebrew and Persian.
An agent's paid call needs a plan and a confirm bound to that plan. Text from outside systems arrives marked as data, never as instructions, and connector results tell the agent to ask you before it publishes anything.
Every supported format made the full round trip, thousands of times over: upload, translation, export, re-parse. Into 29 languages, including Arabic, Hebrew, Japanese and Thai, on local models.
Connections
28 connectors, from help centers and headless CMSs to code repositories and shared drives. Content comes in, runs through the same pipeline as everything else, and goes back as each system's own language version. The source is never overwritten.
Beta Built and tested against the system's published API; not yet run against a live account.
Everything else connects through the full API, a signed webhook intake, or the MCP server. Every connector ships as source, like everything else: adapt it to your setup, or switch off the ones you don't use.
Don't see your system? Request it and I'll reach out.
How connectors work →
How it works
Content comes in, machine translation does the bulk, quality and post-editing run automatically, and only the exception slice reaches a human. Then it ships, with the whole run recorded.
Every request and response is captured whole. Your ML team finally gets all of it.
Your hardware. Your models. Your data.
Let's talk. A 30 minute technical walkthrough: see it run, ask anything, decide if it fits your environment.
Built by a localization veteran of Intento, Lilt, LinkedIn, and Cricut.
Schedule a technical walkthrough →
or email brent@patoislabs.com