Marta
Madrid
The first is texting someone, which you were doing anyway. The other four happen on their own, on the material you just produced.
You use Glotta instead of the system keyboard and write normally. As you finish the sentence, the translation appears above the keys. One tap swaps it in and you send that.
The keyboard also reads the previous lines of the conversation, so pronouns, gender and register are the right ones — which an isolated sentence can never guarantee. You have not opened an app and you have not studied anything.
Ci vediamo dopo il lavoro?
Every sentence you write lands in a list, deduplicated: write the same thing three times and you do not get three rows, you get one row with the counter at three.
Repetition is the signal. What you repeat is what you actually need, and no editor could have guessed it for you.
The same sentence three times: one row, counter at three.
A spaced-repetition algorithm schedules the reviews, but on a deck no editor chose: your sentences, ordered by how many times you have needed them.
The algorithm is SM-2, the same principle as Anki. The difference is not the mechanics: it is that the deck wrote itself while you were living your day.
SM-2, like Anki. The difference is the deck: your sentences, ordered by how many times you have needed them.
Tap a sentence and it comes apart: the verb with its conjugations, the words worth keeping, the reason for each. And you can ask it questions in a chat.
The lexical analysis runs on the device with NLTagger; the dictionary is Apple's own system one, so a definition is never a network request.
¿Te que nos por la noche?
You feel like
Impersonal: the subject is the thing you fancy, not the person. After “que” it takes the subjunctive.
Anche
A simulated conversation in a messaging style, built from the sentences you wrote today or yesterday. You answer in the language you are learning.
It tells you whether that was right, nearly, or not — with the correction and the reason. It is the one moment that asks for active recall, and it is built on material you have already used.
Natural and correct. It is exactly how a native speaker would answer.
It replaces the system keyboard in every app. It translates finished sentences, corrects within one language, swaps text in with a tap, and reads the context of the conversation.
All three surfaces read the same store on the phone. What they do not yet do is follow you to a second device — there is no sync, and it is the most serious limitation Glotta has today.
Above the keys, in «Highlight»
Ti va divernosdomani sera?
The sentence stays yours; over the words worth keeping, the equivalent appears. You learn in context without losing the thread.
The free tier has the first one. The other three come with Premium, because they are the ones that genuinely ask the engine to choose which words to move, not just to translate.
On the typing path. Cancellable, cached, and if it cannot keep up it steps aside instead of blocking the keyboard.
Keystroke
the host app knows nothing
Sentence end
SentenceDetector
Cache
shared LRU · the instant path
Translation
on-device · cloud if you choose it
The bar
above the keys, in time for the next sentence
Off the typing path, where it can take the time it needs. This is where a sentence becomes something to learn from.
Sentence
stored with its translation
Dedupe
one row, a counter going up
Words
NLTagger, both languages
SM-2
when to bring it back
Review
tomorrow's deck
60–70 MB
A keyboard extension lives in very little memory. So Glotta only translates complete sentences, checks the cache first, works in a cancellable task, and if it really cannot manage, steps aside rather than taking the keyboard down with it.
0 bytes
In secure fields iOS hands the keyboard no context at all: there, Glotta reads nothing, translates nothing and learns nothing. The setting that would allow it exists, and it is off by default.
3 levels
Writing in the language you are learning is not a translation: it is a review. A cascade of on-device engines tries strongest-first and stops at the first one that can genuinely answer.
1 rule
A rewrite that drops a number, eats an emoji, or answers with an explanation is thrown out before it reaches you. The way a generative model fails is by giving a confident, false answer.
None of this is visible while you type, and that is the point: the interesting part of Glotta's engineering is the part that saves you the wait.
It exists, and it is written down. The manifesto says what Glotta is betting on, what could disprove it, and what is missing today.