AI Translation Software for Business

diff

The project overview: source keys, translated strings, review findings and coverage for twelve locales.

Every string goes through Glossa Flux, the fast pass, the moment the English changes. The ones your brand lives on go through Glossa Deep, the careful one. Both end at the same four checks, where a row passes, is repaired once, or is held.

Start for free Watch the run

What you get that a translate button cannot give you.

A monthly word allowance instead of an invoice per key, and unchanged strings cost nothing to re-run. Past the allowance a paid plan meters at $0.0008 a word, printed here rather than found on an invoice.

flux · the fast pass

every push

Drafts the whole locale set at once.

One fan out per commit, not a queue. Unchanged rows cost nothing.

$ transglot push

diff 3 keys changed

flux 9 locales, in parallel

gates 27 of 27 pass

pull request opened

deep · the careful pass

Team

For the strings your brand lives on.

Flag a key set and it rides the slower lane, with the brand voice profile in the prompt.

plurals

224

CLDR rule sets, held

Russian takes four forms, Arabic six. The rule comes from CLDR, never from a guess.

in the prompt

What Glossa is told before it writes.

your glossary, locked
the length budget
the screenshot
the key and its neighbours
the brand voice

never

by construction

Overwrites a human edit.

A row a person touched is a row the engine leaves alone on the next run.

translation memory

never metered

The second time you say it, it is free.

An exact match comes out of your own memory instead of being written again, and it is not billed. A file nobody changed costs nothing to re-run, which is what makes adding the eleventh language cost what the first did.

auto approve

your threshold

Let the easy ones through, keep the rest.

Set the confidence a row has to clear to skip review. Everything under it queues for a person, so a reviewer spends their day on the strings that actually needed them.

vision

It can look at the screen.

On the strings that need it, the draft is written against a screenshot of where the string renders rather than against a character count.

and then it stops

one retry, then a refusal

The part nobody else will write down.

A row that fails is repaired once and re-checked. If it fails again it is held, the branch check stays red, and nothing broken reaches a customer. Everyone else flags it and ships it.

placeholders held
plural forms complete
glossary terms enforced
inside the length budget
repaired once
then refused

First-pass translation in seconds, with your placeholders intact.

live: pick a string or type one, choose a lane, run. Flux drafts every push; Deep is slower and reads the brand voice first. Nothing on this plate leaves the page

glossa · engine consoleacme / checkout / checkout.cta

source

Glossa Flux · fast, drafts every push16 / 24 chars

Glossa Fluxplaceholders · plurals · glossary · length

nothing run yet_
pick a source and a lane, then run. Four locales stream here, gate by gate

this run

words drafted0

memory hits0 · 0 tokens spent

human edits untouched0

refused0

time0.0 s

auto-approve at0.90 and above

Flux drafts on every push and lets the gates and the confidence figure decide what a person has to look at. Most rows clear; the rest wait in review.

the engine never overwrites a human edit: a locale someone touched is kept as it is, not redrafted, and the meter counts it

A simulation, in your browser. Nothing is sent anywhere: the drafts, the confidence figures, the memory hit and the timings are a fixture and illustrative. The two lanes, the four gates, the one repair pass and the human-edit rule are the shipped mechanism.

However it arrives, it takes the same road.

git pushlive editorconnector syncagent over MCP

format read, placeholders held

keys sync

every locale, in parallel

flux draft

placeholders, plurals, glossary and length

four gates

auto repair

back the way it came

publish

pull requestCDN bundleheld for review

A row that fails a gate is repaired once and re-gated. If it still fails it is refused and held for review, with the rendered screen attached.

All of it, on one screen.

What the cycle has used, what memory served for nothing, and how far each workspace has got.

The command centre: words reused from memory instead of retranslated, AI words this cycle, the average quality score, held-back locales and the decisions waiting on a person.

reused at $0

one pipeline, every locale

glossa deep, the careful pass

The strings your brand lives on, minded.

Turn quality mode on for a project, or for just the locales that carry money or law. Everything else stays on the fast pass, where it belongs.

You pick the scope: The whole project, or only the locales that carry money or law. Everything else keeps the fast pass.
Reviewed in the real screen: Re-rendered at real size, in place, so tone, fit and glossary are checked where a reader meets them.
Escalates with evidence: A refusal reaches a human with the rendered screen and the failing gate attached. Approve or fix in one step.

How quality mode works

Glossa Deep is included from the Team plan.

The AI settings screen with automatic translation, model selection and translation memory controls.

deep, from the Team plan

qa guardrails

Four gates. Every row, every commit.

Whichever pass drafted it, no row ships without clearing all four. A failure gets one automatic repair, and is refused if that does not fix it.

Placeholders

Every variable survives the trip: never dropped, never reordered into nonsense.

en Hi, {name}!

ja {name}さん、こんにちは

Plurals

All the categories a language needs. Russian takes four, Arabic takes six, and the rule set comes from CLDR rather than from a guess.

ru файл

ru файл / файла / файлов

Glossary

Locked terms and forbidden terms are enforced, not suggested, in every locale.

term "Transglot", do not translate

held in every mapped locale

Length

German is long and buttons are small. Rows are measured against the component budget, not a character count.

de In den Einkaufswagen legen

de In den Warenkorb

the engine, in parts

What the pipeline actually does.

One command, any repository

The CLI reads your format, pushes the diff and streams the run back by locale.

$ npx transglot push

diff: 1 key changed

flux: drafting mapped locales

gates: 4 of 4 pass, 1 repaired

file formats

JSON, YAML, .strings, .stringsdict, ARB, .po, XLIFF 1.2 and 2.0, RESX, properties and more. Seven of them carry plurals.

Memory before generation

Translation memory and locked glossary terms are bound before a draft is written, so a term you settled last year outranks anything the model prefers today.

translation memory
glossary locks
style rules

Context, not just text

Where a screenshot exists, the draft sees the component the string ships in, along with its length budget.

rendered screen
length budget
key history

Every locale ICU knows

No fixed list to run out of.

Your agents speak Transglot

Sixteen MCP tools, the same verbs and the same gates as everything else.

agent which keys fail length in de?

mcp 3 keys, re-drafted

mcp gates 4 of 4 pass

Refusal is a feature

A row that cannot be repaired is held with its failing gate and its rendered screen, and the branch check stays red. The alternative is a broken string that nobody sees until a customer does.

The same five questions, three very different answers.

This is why localization stopped being a project with a start date and became a build step nobody schedules.

The same five questions,
Question	Glossa Fluxthe fast pass, always on	Glossa Deepthe careful pass, quality mode	Agencies and TMSthe old way
Trigger	Every push, automatically, before anyone asks.	Quality mode. You pick the project, or the locales.	Email a spreadsheet, wait for the quote.
Fan out	Every mapped locale in parallel, not a queue.	Deliberate and multi step, on the locales you chose.	One language at a time, per release.
Review	Four gates on every row: placeholders, plurals, glossary and length.	Re-checked in the rendered screen, at real size.	PDF back and forth over email.
On a failure	One repair pass, then refused. It does not ship broken.	Escalated to a reviewer with the screen attached.	Found by your users.
Cost	Included on every plan, Free included.	Included from the Team plan.	Priced per word, every word, every release.

Glossa Flux

the fast pass, always on

Trigger: Every push, automatically, before anyone asks.
Fan out: Every mapped locale in parallel, not a queue.
Review: Four gates on every row: placeholders, plurals, glossary and length.
On a failure: One repair pass, then refused. It does not ship broken.
Cost: Included on every plan, Free included.

Both passes ship with the same four gates. Overage on either meters at $0.0008 a word.

FAQs

The questions teams ask before they let it write.

Does every string really go through the fast pass?

Yes. Glossa Flux is the default path for every key on every plan, Free included. There is no manual send to translation step: the engine runs on the push, the connector sync or the API call that changed the string.

When should a project go on the careful pass?

Turn quality mode on where nuance is money: the marketing site, pricing, onboarding, anything legal. It is a project switch, and you can narrow it to the locales that matter most. Everything else ships perfectly well on the fast pass. Deep is included from the Team plan.

What happens when a row fails a gate?

It gets one automatic repair pass: placeholders reinserted, length re-fitted, glossary terms restored. Then all four gates run again. If it still fails, the row is refused and held for review with the rendering attached, and the branch check stays red.

Which languages can it draft?

Any BCP-47 tag ICU recognises, which is 708 locales in the current ICU table, plus a structural fallback for tags outside it. There is no fixed list to run out of, and plural handling comes from the 224 CLDR plural languages rather than from a guess.

Which formats and frameworks does it read?

Twenty file formats: JSON flat and nested, Android XML, iOS .strings, .stringsdict and .xcstrings, Flutter ARB, Laravel PHP and JSON, XLIFF 1.2 and 2.0, gettext .po, YAML, CSV, XLSX, Java properties, RESX, Markdown, Unity CSV and Unreal PO. Seven of them carry plurals.

Do you train on my content?

No. Your strings are sent to the subprocessors listed in the Trust Center to be translated and nothing else, and nothing you send is used to train a model. The shared exact-match memory pool is anonymised and can be switched off per project on every plan. The Trust Center is the authoritative record of who can process your content and in which region.

works with what you already run