Build it into your systems
Direct integration, delta syncs, in-place review. Every language version stays current without anyone coordinating it.
OnePlatform · For engineering-led teams
Everyone has the same models. The difference is what you can ground them in.
Your company has already decided how it says things in every language it sells in. Those decisions sit in delivered files, in people's heads and in email threads — nowhere a system can read them. This is where they become structured, and what your own AI tools query when they need them.
TextUnited · MCP call
terminology.lookup({ term: "checkout flow", locale: "de-DE" })
approved"Bezahlvorgang"
rejected"Kassenablauf", "Checkout-Prozess"
approved_byProduct · DE · Sep 2025
Automatically saved in the systemOwned by you
Where those decisions are now
Your company settles how it speaks a dozen times a week. None of it leaves anything a system can read.
| A translator settles a term | Which of three correct German words your company uses — and the two it does not. |
| The team in Lisbon writes the page | In Portuguese, from scratch. Their wording becomes the version, and it was never a translation of anything. |
| Legal rejects a phrasing | In one market, while approving it in another. The reason exists; the record of it does not. |
| A reviewer fixes what the model produced | Rewrites the two sentences that were wrong, ships the file, and moves on to the next one. |
Those are four of them. Add the style guide marketing wrote, the terminology technical documentation keeps separately, and the naming the product team settled two releases ago. Each one is a decision made by somebody qualified — and each ends up in a delivered file, a Google Doc, a mail thread, or one person's spreadsheet.
Not lost, exactly. Just nowhere a machine can reach.
Where they live instead
The answers exist. Getting one means finding the person who has it.
Language data used to mean one thing: does this target match this source. That was enough while a translator was the one asking, because everything the match left out — whether it was still current, which market it came from, whether anyone had approved it — they filled in themselves.
Now the thing asking is a model, and it needs the rest of the content to tell it what your company is allowed to say. Which was never a matching problem.
Somebody in your company can answer each of these. Usually one person, from memory, if they happen to be online.
Which term does this business unit use for that component, and which two did it reject?
Which phrasing did legal clear for Germany and refuse for Austria?
What changed when the spec was revised, and who signed it off?
Is this still current, or did we replace it two releases ago?
Put those answers where a system can reach them and they stop depending on who is available.
None of that is a property of translation. A page written in Lisbon and never translated into anything carries exactly the same decisions — a term someone chose, a claim legal approved, a phrasing the market accepted. It belongs in the same record.
What you can do with it
Four ways to work against the record.
Give your assistant something to stand on
Your own AI queries approved terminology through the MCP server while it drafts, and can propose a new one. What becomes binding stays a human decision with a name against it.
Automate the flow
Build the pipeline visually, in the automation tool your team already runs.
Check content against what you have decided
Measure a document against whatever you decided should hold: the original it came from, or the regulation it has to satisfy. A rule is a source too.
Getting your existing content in is engineering, not a setting. Segmenting and labelling an archive that was never translated, thresholds, connectors — we build those with you. We are API-first, so none of it is locked to us either.
And a way in for everyone else: the same browser interface, included for every user on the plan, so the departments who never asked for an API are not waiting on your team to translate a deck.
What it takes to run
Unlimited users, isolated tenant, your own models.
Unlimited users — freelancers, in-country reviewers and whole departments cost the same as a handful of each. A person kept outside the system makes decisions that go nowhere the next project will look, which is the real cost of a seat licence.
Isolated per tenant — sealed to your account, never used to train anything outside it, never another company's advantage.
Business units on one record — separate administrators, cost centres and billing, over shared language data. Where a unit genuinely needs its own terms, that is a domain rather than a second termbase.
Your own approved models — connect the AI your company has already cleared, rather than the one we happen to prefer.
SSO — because you were going to ask on the second call.
Before the technical call
Questions
We write a lot of content natively, not as translations. Does any of this apply?
Terminology and style are not properties of a translation — they are properties of how your company speaks, and they apply to Portuguese written in Lisbon exactly as they do to Portuguese translated from English. Getting natively written content structured into the record is project work rather than an upload.
How do our agents actually use this?
Through the MCP server, which exposes terminology as tools an assistant can call, read and write. Your tool does the generating; we hold what it stands on.
Why not just put our documents in a vector database?
Similarity retrieval answers what looks related. The question here is what you are allowed to say, and no distance metric knows that a term was rejected for that market or superseded two releases ago. In practice the two sit together: the index finds candidates, the record says which are authoritative.
What happens when we change models?
Nothing in the record moves. It sits outside the model and applies to whatever is generating. A prompt-based glossary is a property of the model you happen to be using this quarter.
Our translation runs through an agency. Does that have to change?
Not to begin with. Most engagements start by bringing that stream in, because it is the one with a budget and an owner — but the record is the point, not who performs the translation.
How does the data stay ours?
Isolated to your tenant, never used to train anything outside it, and exportable in the standard interchange formats — not a proprietary dump, but files a competing system can open.
Bring your stack, and your files.
The useful first conversation is not a demo of our interface. It is your CMS, your repository or your assistant, a look at where the record would sit in it — and what you already have in eight languages that could go in on day one.