Skip to main content

Translation commands

Translation commands work on whatever is selected, in whatever app you are in. No copying into a browser tab, no switching windows: select the text, press the shortcut, and either replace it with the translation or read it in a popup.

Commands

To English

Inline⌥T

Type in your language, get natural English. Keeps your tone, uses simple words.

Use To English when you think or draft in your native language and need the result in natural-sounding English. It detects the source language automatically and produces fluent English that matches your original tone, whether casual, formal, or frustrated. Unlike machine translation, the output reads like a native English speaker wrote it from scratch.

TranslationEnglishLanguageInline

That difference matters most when the text is embedded in something you are already doing - a message you are replying to, a PDF you are reading in Preview, a comment on a pull request. A browser translator makes you leave the task; a shortcut does not.

Because the translation is done by a general-purpose language model rather than a phrase-matching engine, it keeps register and intent: a formal email stays formal, an informal message does not come back stilted. For a specific language pair, see the per-language guides such as translating Arabic to English or Spanish to English on Mac.

The To English command is deliberately direction-agnostic - it detects the source language rather than asking you to pick one, which is what makes it usable on text you cannot read. If you routinely work in one pair, duplicate the command and pin the languages in its prompt so the model never has to guess.

Translation is one of the cases where the choice of model matters most. Larger models handle idiom, register and low-resource languages noticeably better than small ones, and the per-language guides name the model that performed best for each pair in our testing. If your text cannot leave your machine, a local model through Ollama will handle common European pairs acceptably, though quality drops on languages with less training data.