AiTranslator: translate and improve any text with a single key

Five AI engines to choose from, 28 languages, three keyboard shortcuts, translation of entire files, reading text from an image, microphone dictation and audio playback, on Windows, macOS and Linux, in x64 as well as ARM64

Select text anywhere, press F9, it's translated. There you go, you know the software. You can stop reading here if you're in a hurry.

Try it for free, keep it for €3 per month

Online AI engines are paid for separately, through their provider, with your own API key. Ollama, on the other hand, is free and runs on your machine.

Do you use several of my software products? The PRO account at €12 per month includes them all, AiTranslator included.

AiTranslator in detail

For those who want to know everything before installing: I wrote AiTranslator because I couldn't take the obstacle course anymore. Copy the text, open a tab, paste, wait, copy the result, go back to my email. And when it wasn't about translating but about making a message a little more presentable, that was yet another tool. Three services for a four-line paragraph, honestly, no.

AiTranslator stays in the system tray and wakes up with a keyboard shortcut. It translates into 28 languages, it rewrites a text in its own language, and it lets you choose which AI does the work: OpenAI, Anthropic, Google, xAI, or Ollama which runs on your own machine without sending anything anywhere.

How it works

You highlight text in whatever application you want: an email, a web page, a document, a conversation. You press F9 (that's the default shortcut, you can change it in the preferences if you already use that key for something else). The window appears with your selection already in it, and since 1.8.2 the translation starts on its own. Before, you still had to click Translate, which was one gesture too many for software that promises a single key. If you prefer to proofread before starting, you can uncheck the box in Preferences.

Two buttons in the window. Translate, you get the idea. Improve, that's the button I use most often: it keeps your language and rewrites the text to make it clearer, more polite, more professional. Handy for the email you type at 6:50 p.m. while being vaguely annoyed.

The response is displayed as a stream, meaning word by word while the model writes, without waiting for it to finish its sentence. With a slightly long paragraph, it changes everything: you read the beginning while the end is being made. One click on Copy and the result goes to the clipboard.

This is what it looks like on my end, in dark mode, with an English message translated into French:

The little blue badge en in the top left is the source language detected by the AI. It has been clickable for a little while now: if detection gets it wrong, with a short text or a mixture of two languages, you impose the right language on it and don't think about it anymore. At the bottom of each area, the character and word counter. On the right, the microphone and the speaker, which I talk about just below.

Three shortcuts, including one that writes in your place

For a long time there was only one shortcut. There are three now, one for each action, and each one can be configured in Preferences:

• Translate, the historic F9: it opens the window and translates the selection.
• Improve: it opens the window and rewrites the selection in its own language. The button already existed, it was missing its key.
• Translate and replace: this one doesn't open anything at all. The selected text is replaced by its translation, right where it is.

The two new ones don't have a default shortcut, and that's intentional. A global shortcut takes its combination away from all other applications, and on a French keyboard Ctrl+Alt plus a letter is already AltGr. I prefer you to choose. On my machine it's Ctrl+F9 for Improve and F10 for Translate and replace. Preferences reject a duplicate between two shortcuts, as well as Ctrl+C and Ctrl+V, which the application needs for itself.

Translate and replace is the action I've wanted from the beginning. I'm replying to an email in French, I select my reply, I press the shortcut, and it's in English in the email. AiTranslator copies the selection, translates it without showing itself, pastes the result in its place, then puts back in the clipboard whatever was there before. If something goes wrong (nothing selected, license expired, engine unreachable, window changed during translation), nothing is pasted and the window opens to say why. And if you change your mind along the way, a second press on any of the three shortcuts cancels.

Two honest details. Nothing is displayed between pressing and pasting: a second or two with an online engine, quite a bit longer with a large Ollama model on a modest processor. And in a Linux terminal, where Ctrl+C means something completely different from copying, this shortcut is useless.

Translate an entire file

You click the paperclip in the upper area to choose a file, drag it onto it, or copy it in the file explorer and press Ctrl+V in it, and its text is loaded then translated like any selection. The accepted formats:

• plain text and its cousins: .txt, .md, .log, and the subtitles .srt and .vtt ;
• documents: .docx (Word), .odt (LibreOffice) and .html ;
• PDFs, provided they contain real text.

And if your file has an extension I didn't plan for, that's no longer a problem: any file that contains text is accepted. Data files such as .json, .xml or .csv work too, but they're translated as plain text, including field names. AiTranslator warns you beforehand, because a configuration file translated from start to finish doesn't always restart.

A scanned PDF, that is to say a series of photos of pages, contains no text to read: AiTranslator says so clearly instead of translating nothing. Same thing for a password-protected PDF or a format it doesn't know, with a list of the ones it accepts. Images now have their own path, I talk about it just below.

Only the text is retained. The formatting, tables and images in the document are left at the door, and on a multi-column PDF the reading order can be surprising. For a twelve-page contract or the documentation for a tool, that's exactly what's needed. For a heavily laid-out brochure, less so.

That leaves the question nobody likes: length. An entire document costs more than a sentence, so beyond two pages AiTranslator stops and spells it out: number of pages, characters, words, and how many calls to the engine will be needed. Nothing is sent until you have clicked Continue. The document is then sent in several pieces, cut cleanly between two paragraphs, with progress like “3/9” and a Cancel button that keeps what has already been translated. The limit is about 200,000 characters, or a good sixty pages.

Once the translation is finished, the new Save button, next to Copy, writes it to a file whose name is already suggested based on the original one. A translation of more than 20,000 characters isn't kept in the history, either, so it doesn't make it swell: this button is what you use to keep it.

While I was at it, I removed two silent truncations, and those concern everyone, whether there's a file or not. Text that was too long was being cut off without a word, and is now rejected with an explanation. And when an engine stopped halfway through because it had reached its response limit, the truncated translation was displayed as if nothing had happened. It is now announced as incomplete.

In the same family, I came across a sneakier bug: English texts that came back in English, looking like a successful translation. The model had simply returned the original text. Since 1.15.1, AiTranslator detects it and announces it as an error, and this false result is neither copied to the clipboard nor kept in the cache.

Read text from an image

I had announced this one for another version. Here it is. An error message that can't be selected, a screenshot sent by a colleague, a photo of a menu: there's text, but it's impossible to highlight it. AiTranslator can now read the text in the image, put it in the top area, then translate it like any other selection.

Three ways to give it an image:

• you take a screenshot and paste it into the window with Ctrl+V ;
• you drop a .png, .jpg or .webp file onto the top area ;
• you press F9 without having selected anything, while an image is waiting in the clipboard.

An image is more indiscreet than a sentence: you don't always know everything a screenshot shows. Before sending it to an online engine, AiTranslator therefore displays a thumbnail and asks for your agreement. With the shortcut, the question comes back every time you press it. You click Cancel, nothing is sent. With Ollama on your machine, the question isn't asked, since the image doesn't leave your computer. And in every case it is neither kept on disk, cached, nor written to a log.

All five engines know how to read an image, provided you choose a model that has eyes. If your Ollama model only reads text, AiTranslator offers to download one that can read, and uses it only for reading: your translation model doesn't change. An image more than 2048 pixels on a side is reduced before sending, and a message tells you so. If there is no text on it, you get a message too, and your previous text stays in place.

The limitations, because there are some. GIF and BMP formats are not recognized, and scanned PDFs will have to wait for a future version. An image also costs more to have read than a paragraph to translate. And it's brand new: I don't yet have enough experience to tell you which engine reads a blurry screenshot or handwriting best.

Five AI engines, and you choose

No engine is imposed. You switch from one to the other in the dropdown menu of the main window, and each one offers several models, from the fast and cheap one to the more polished one:

• OpenAI: gpt-6-luna, the newest in the family and also the cheapest. I ran on gpt-5.6-luna for a long time, which translates very well, and I don't yet have enough experience with the 6 to say whether it does better.
• Claude, from Anthropic: claude-haiku-5-5.
• Gemini, from Google: gemini-3.5-flash-lite, the budget version. I haven't tried it in translation yet, so its quality remains to be judged on the evidence. Its big brother gemini-3.8-flash costs more than twice as much.
• Grok, from xAI: grok-4.3. Haven't tried it either. Grok 4.7, which is newer, is significantly more expensive.
• Ollama: Gemma 3, the one I recommend (I explain why just below), or any other model you have installed.

As for prices, taken from the official pages in October 2026 and given in dollars per million tokens (a token is roughly three quarters of a word), for input and then output: 0.10 and 0.50 for gpt-6-luna as for claude-haiku-5-5, 0.30 and 2.50 for gemini-3.5-flash-lite, 1.25 and 2.50 for grok-4.3. Suffice it to say that translating an email costs a fraction of a cent, whatever the engine.

These names age quickly, one model chases out the other every three months. AiTranslator therefore doesn't freeze the list: in Settings, under the AI Providers tab, the Load Models button loads the models your key is allowed to use, and the field next to it accepts any name typed in by hand. On the first installation, each engine starts with the model mentioned just above. A model you chose yourself is not replaced by an update.

Ollama deserves a mention. It's an AI server that runs on your machine, free of charge and without an account. No API key to pay for, no data leaving your home, no volume limit. AiTranslator detects it and starts it automatically when you select it as the provider, and if it isn't installed yet, it offers to do so with one click (I go into a bit more detail about that further down). An old laptop won't perform miracles with a large model, but for everyday translation with a medium-sized model, it works really well.

If I had to recommend just one, it would be Gemma 3 from Google, in the 4B or 12B version depending on your machine's memory. It translates cleanly into French, and starting with the 4B it can also read images (the small 1B only does text). Since AiTranslator now reads the text in images, this is the right model to have on hand.

What I really use every day

The rest accumulated version after version, generally because I needed it myself.

• Translation in both directions. I work in French and English all day, and I used to spend my time changing the target language. Now I never change it anymore: an English text goes into French, a French text goes into English, and a Spanish text goes into my usual language. The second language in the pair is chosen in Preferences, and a small badge shows the language actually selected when it differs from the one I chose.
• The result already in the clipboard. Since 1.7.0, the completed translation lands there all by itself. Concretely, I select, F9, go back to my email, Ctrl+V. That's it. Before, I still had to click Copy, and I forgot it one time out of two. You can uncheck it in Preferences if it bothers you.
• Dictation through the microphone. One click on the microphone, you speak, your voice is transcribed by Whisper (OpenAI's speech recognition model) and the text lands in the source area, ready to be translated. To dictate three sentences to translate during a meeting, it's noticeably faster than typing them.
• Audio playback. The speaker reads the translation out loud, with a real synthetic voice and not the 2004 robot. The Auto checkbox starts playback by itself after each translation. And since version 1.6.3, listening again to an already played translation is instant and doesn't charge you again: the audio is kept in memory.
• Tone. Neutral, Formal, Casual, Technical or Literary. The same text doesn't come out the same depending on the slider, and it also works in Improve mode, not just in translation.
• Glossary. You impose your own translations for your in-house terms, your product names, your professional jargon. The AI respects them instead of reinventing a different word on every call.
• History. The last 200 translations, with an instant search. You reload an old translation with one click, copy it, delete it.
• Cache. Translating the same text twice doesn't call the API again: instant result, and not a cent spent. It now survives a restart: up to 2,000 translations are kept on your disk, without the original text, and nothing goes anywhere else. A small badge indicates a translation that came from the cache, and clicking it requests a real one from the engine again. The cache is cleared in Preferences, under the Backup tab. In the same spirit, in 1.7.0 I fixed an ugly bug that I had eventually noticed: a single press on the shortcut would sometimes trigger two complete translations, so twice the bill. Only one call is sent now, and the display no longer starts from zero halfway through.
• Usage counter. With your own API key, you always end up wondering how much it costs. The Usage window answers that: the tokens sent and received, and the estimated cost in dollars, by engine, for today, the week or the month. It also counts the calls that the cache saved you. When an engine doesn't provide its exact figures, the total is marked as estimated, and Ollama is shown for what it is: local and free. These figures stay on your machine.
• Theme. Light, dark, or automatic according to your system setting.
• Your API keys stay with you. They are encrypted locally, using DPAPI on Windows (the system vault, linked to your user account) and AES-256 on macOS and Linux. Nothing goes to the cloud except the text you ask to translate.
• A backup in .jsonl (a text file with one translation per line), with automatic rotation, to find what has gone through the application.
• System tray. The application minimizes to the tray and can launch at startup. It waits for its shortcut without cluttering your screen with a window.

An honest note about audio: the microphone and read-aloud playback work on Windows and macOS. On Linux, it's not connected yet, it's on my list. Everything else works on all three systems. And since transcription and voice go through OpenAI's audio models, these two features require an OpenAI key even if you usually translate with Claude or Gemini.

As for languages: 28 translation languages, from French to Hindi via Arabic, Japanese, Ukrainian and Vietnamese. The interface itself is translated into 29 languages, so there's a good chance it speaks yours. On first launch, it actually uses your system language, and the language list shows them by name (Deutsch, Español, Українська) instead of a two-letter code that nobody remembers.

Installation, three minutes flat

Download the archive that corresponds to your system at the bottom of this page, and unzip it wherever you want, for example in C:\AiTranslator\ on Windows or in /Applications/AiTranslator/ on Mac.

On Linux and macOS, remember to make the binary executable before the first launch:

chmod +x AiTranslator
./AiTranslator

When you launch it for the first time, a wizard opens by itself. You choose your engine, it gives you the link to the page where you can create the API key, you paste it in, and it tests it with a real little translation: if the key is rejected, you know right away, not three days later in the middle of an email. You can also choose Ollama and not paste any key at all, the banner will guide you if something is missing. Don't feel like doing it now? Close the wizard, a banner in the window will bring you back to it whenever you want, and everything can still be adjusted in Preferences. Then, select some text somewhere or drop a file into the window, F9, and off you go.

Ollama, free AI, in one click

A Windows 11 user wrote to me, a little dejected: he had chosen Ollama from the list, started a translation, and got an incomprehensible red line talking about “retry”. He was right to complain. Ollama is a separate piece of software, which had to be installed by hand, and nothing told him that. Since 1.8.0, that is sorted: AiTranslator takes care of it.

Choose Ollama from the drop-down list in the main window. If something is missing, a banner appears just above the translation, with the right button:

• Install Ollama when Ollama is not on your computer. On Windows, AiTranslator downloads the official installer while displaying the percentage, runs it, and all you have to do is follow the wizard.
• Start Ollama when it is installed but stopped.
• Download the model when the AI model is missing, the “brain” that does the translation, a large file of a few gigabytes, downloaded only once.

One click takes care of everything that is missing, with a Cancel button if you change your mind. At the end, the banner disappears and your translation starts by itself. On Mac, AiTranslator opens the Ollama download page: you install it, then click Retry. On Linux, installation requires administrator privileges, so the banner displays the command to paste into a terminal.

For the model, AiTranslator chooses according to your PC's memory, from Google's Gemma 3 models, which know more than 140 languages: the 1B version (0.8 GB) with 8 GB of memory, the 4B (3.3 GB) up to 16 GB, and the 12B (8 GB) beyond that. Nothing stops you from choosing another one in Settings, AI Providers tab, Ollama card: that choice is never replaced. And the model is loaded into memory as soon as the application starts, so the first translation no longer makes you wait.

A trap I've seen come up: Ollama does not install with npm. If you run into an error like Cannot find module npm-cli.js, it's Node.js complaining, not Ollama. Forget it, the banner button is enough.

And if you like doing everything by hand, that is still possible: the installer is on the Ollama download page, and the models can be downloaded from the Ollama card in Settings, with the Download button.

Free trial and license

30-day free trial, all features included.

Download AiTranslator and enjoy 30 full days with no restrictions at all. Test all AI providers, all languages, all features. No credit card required.

After the trial period, two plans to choose from. The AiTranslator subscription, at €3 a month or €25 a year, for this software alone: one license covers 3 devices. Or the PRO account at €12 a month, which gives access to all my software.

Everything is handled from the application's license window. You log in to it, subscribe, and when you return from payment the I've paid, verify button activates the license without logging you out. The My account link opens your subscription page on the site. If a license is not issued, the application gives the exact reason: for example, that your 3 devices are already in use, with the figures. In that case, Free this device frees up a spot for another computer, or you get one more license. So that you can recognize your machines in this list, the computer's name is sent to the site when you log in.

The online version is 1.15.2, available in six versions: Windows x64 and ARM64, Linux x64 and ARM64, macOS Intel and macOS Apple Silicon. If you have a recent laptop with a Snapdragon processor, or a Raspberry Pi, the ARM64 version is the one you need. Compared with 1.8.3, it brings text reading from an image, the first-launch assistant, the cache that survives a restart, the usage counter and the new license window. 1.15.0 added the button for opening a file and a discreet reminder when an update is available. 1.15.2 starts in your system's language, displays the remaining trial days even when you are not logged in, and arranges the buttons on two lines when the window is narrow, instead of piling them on top of one another. An account suspended on the site is also announced as such when you log in. The details of what has changed can be found in the release notes, just below the download buttons.