Agentic localization with human review
Teams let their own coding agent run the translation. When a linguist approves a translation, it is sealed, and anyone can check that no word has changed since.
Your team's coding agent sends the jobs. Context goes with every string: the glossary, screenshots, translation memory. The strings you pick go to a linguist, and anyone can verify the approval.
Which model translates best, by language and content type?
Share of strings in which professional linguists found no errors. Higher is better.
The best model changes from row to row, so you can set one for each language and content type. How routing works
AI in a typical TMS vs Localize.io
| What | Localize.io | A typical TMS with AI |
|---|---|---|
| Human review | Localize.ioA linguist you name reviews each string and seals it; anyone can check the seal, and a changed word no longer verifies. See it in the tour: Review by a person | A typical TMSA "reviewed" status set in the tool, taken on trust. |
| What it costs | Localize.ioEvery model call is recorded with its cost. See it in the tour: The record | A typical TMSModel spend is often hidden, or shown only as a total. |
| Long files | Localize.ioSplit into batches sized to each model's output limit. See it in the tour: Send work | A typical TMSTypically sent in one request: a long answer gets cut off, and is still paid for. |
| When a check fails | Localize.ioEvery attempt is logged, at most three; a string that still fails comes back with its best attempt and a question for you, and the rest of the job is delivered. See it in the tour: Checks | A typical TMSThe model is often re-asked silently, and billed again. |
| What the model did | Localize.ioThe prompt, the reply, the tokens and the cost are visible on every call. See it in the tour: The record | A typical TMSPrompts, reasoning and tokens typically stay inside the platform. |
| Which model | Localize.ioYour configuration picks the model, per language, content type, project and pass, and every job shows which setting decided it. See it in the tour: Routing | A typical TMSThe platform picks the model, and how it decides isn't shown. |
| Translation memory | Localize.ioEvery string gets its nearest past translations by meaning, with a similarity score, even when no fuzzy match exists. Approved wording is reused; outdated memory isn't. See it in the tour: Example 3: translation memory | A typical TMSA fuzzy match of 60–70% or more, if there is one. Shorter or reworded strings often get nothing. |
Seal
Change one character, and the seal breaks.
A linguist approved “Analytics” in Japanese as 分析. Anyone can check that seal, with POST /v1/signatures/verify or without us.
"valid": true"valid": falseModels
The right model for every language, and every content type.
Interface strings, help articles and marketing copy are different work. Give each the model that handles it best. When you have a glossary, a second model checks that its terms are kept.
Set it once. Change it any time: the next job uses the new setting.
A route is set by
ui and Spanish marketing apart.Change a route, and the next job follows it:es-ES marketingMistral Large 4GPT-6.1 Sol
How it works
Send strings. Get them back translated.
Your agent sends segments through the API, each string with its id and its context, and gets the same segments back, translated. A file in your own format stays on your side: your agent turns it into segments and writes the translations back.
Segments in, segments out
{"id": "sign_in",
"source": "Sign in",
"context": "header button"}{"id": "sign_in",
"target": "Iniciar sesión"}{"id": "get_key",
"source": "Get API key",
"max_length": 20}{"id": "get_key",
"target": "Obtener clave de API"}Context
Each string gets only the context it needs.
A screenshot when the screen explains the string, a glossary term when the string uses it, a past translation when one is close, a rule from your style guide when the tone matters. The same goes for design specs, product docs, reviewer notes and whatever else you keep.
Every asset is split by language, so a job into German gets German terms, memory and rules, and nothing of the other languages.
Approved wording is reused. Outdated or wrong memory isn't.
Step 2: your assetsThe examplesExample 3: translation memory
Three strings, three different contexts
A button or a noun? The screen says: the button under the Pro plan.
The name of the page has to match the name in your product.
A close line from last year's campaign: its tone carries over, “business” doesn't.
Reviewers
A person reviews it, where it matters.
Send a translation, ours or anyone's, to a linguist you name: ours, or your own people. They read each string with your terms, memory, screenshots and style guide, correct what needs it, and seal every string. Then the same checks run on their work before it comes back to you.
Choose by price, by what is already waiting for them, by how fast they work and when they are at work. Changed your mind? Withdraw it, or send a new version to the same person.
- Translation onlyModel translations, checked. No person involved.
- Review onlyTranslations you already have, from any vendor or your own team, read and sealed by a linguist.
- BothTranslate everything, then send the strings that matter, such as legal or marketing copy, to review.
And it makes the next job better: what a linguist notes and fixes again and again becomes suggested changes to your glossary, style guide or source. Accept them, and every next job works from them. How
One string, back from a real review
“アナリティクス is just a transliteration and sounds like a specific service name such as Google Analytics. It should refer to one of cookie categories.”
- Less setupNo converting, sorting or preparing first. Bring what you have.
- New material counts from the next jobAdd a glossary or screenshots for the whole account or one project, whenever you like.
- Every result is checkedBefore it comes back to you, not after someone notices.
- A full recordHow each translation was made: the route, what the model was shown, what it answered, the checks, the cost.
- Fits how you workA REST API for your pipeline, MCP for your agents.
Connect in minutes. One key per integration, over REST or MCP.
Read the docsTranslate your first strings in a few minutes.
White label, on-premises or your own models? Enterprise →