Screen Translator

A 4-in-1 Windows background utility that replaces Lightshot, Punto Switcher, TaskbarX, and QTranslate — in a single 3 MB executable with no installer, no ads, and no bundled telemetry.

Lives in the system tray and exposes everything through global hotkeys.

---

What it replaces

Replaces	With this feature
Lightshot	Region screenshot to clipboard (Ctrl+Alt+D)
QTranslate / Google Translate popup	Clipboard translation with popup (Ctrl+Alt+T) and region OCR + translation (Ctrl+Alt+S); auto-detects 18+ source languages, optional DeepSeek backend
Punto Switcher	Manual and automatic keyboard-layout correction across English, Russian and Ukrainian (привет ⇄ ghbdtn, привіт ⇄ ghbdsn)
TaskbarX / TaskbarCenter	Dynamically keeps taskbar icons centered

Everything is configurable through a single settings window and runs from the system tray with a ~20 MB memory footprint.

---

Features

• Clipboard translation — copies the current selection, translates it, and shows a popup next to the cursor. The source language is auto-detected from the text (Cyrillic → ru/uk, kana → ja, hangul → ko, Han → zh, Arabic / Hebrew / Greek / Thai / Devanagari, plus Latin variants de/es/fr/it/pt/pl/tr by distinctive characters); the target is the currently selected UI language.
• Two translation backends — by default uses the free MyMemory API (no key required); if a DeepSeek API key is set in settings, the much higher-quality deepseek-flash model is used instead, with automatic fallback to MyMemory on any error.
• Region OCR + translation — draw a rectangle with the mouse, the captured pixels are recognized via OCR.space (primary engine) with a fallback to the built-in Windows WinRT OCR, then translated.
• Region screenshot — capture a screen area directly to the clipboard (Lightshot-style), optionally also saved to a folder of your choice.
• Draw on the capture before you send it — a pencil and a rectangle sit in a strip beside the selection, both drawing in red. Whatever is drawn is baked into the saved or copied image. Ctrl+Z takes back the last shape; clicking the armed tool again disarms it and hands the resize handles back.
• Full-page (scrolling) screenshot — captures what a region would show if the window were tall enough. Pick the region, hit Full page, and the app scrolls it a step at a time, grabs a frame after each one, and stitches the frames into a single tall PNG.
  • Scrolls through UI Automation where it can. A ScrollPattern reports the exact scroll position, the share of the document on screen, and takes a target to move to — so there is no guessing at how far a wheel notch goes, no waiting out a fixed delay for smooth scrolling, and an honest signal for "this is the bottom". Chromium, Firefox, WPF, WinForms, UWP and Explorer all expose it.
  • Falls back to the mouse wheel for apps that paint their own scrolling without publishing the pattern. The wheel path calibrates itself on the first notch, since apps disagree wildly about how far one goes.
  • The frames decide where the seam is, either way: each new frame is matched against the last to measure how far the content really moved, so sticky headers, footers and the scrollbar don't throw the alignment off.
• Punto-style layout correction — words typed on the wrong keyboard layout are retyped on the right one as you type, and the layout is switched so the rest comes out right; clear cases are fixed within four or five keystrokes, the rest when the word ends.
  • Real dictionaries. A quarter of a million word forms per language (English, Russian, Ukrainian), ranked by frequency from subtitles and Wikipedia, plus a character model for words no list holds. Every word is judged by how likely each layout's reading of the same keys is, and what you typed gets a head start sized by how much it looks meant: a common word (руку, tv) is never rewritten into another one (here, ем), a typo stays a typo in its own language (кучча), and layout noise (ghbdtn, руддщ) loses to any real word.
  • Russian and Ukrainian told apart by the word. ghbdtn → привет, ghbdsn → привіт; Ukrainian typed on the Russian layout is fixed (привыт → привіт), and so is Russian on the Ukrainian one (єто → это). Spellings both languages share follow the language you have been writing in, and a correct word on the wrong Cyrillic layout (купувати on the Russian one) just has the layout switched under it.
  • Whole words, punctuation included. , . ; ' [ ] are letters on the Cyrillic layouts (б ю ж э х ъ), so ,jkmibvb becomes большими, not ,ольшими. Punctuation after a word is retyped as the target layout has it: ghbdtn/ becomes привет., руддщб becomes hello,.
  • Short words follow the sentence. Alone, z, ,s or ye could be anything; once the next word settles which layout you thought you were on, they are fixed along with it: z ,s gjikf → я бы пошла, ye lf → ну да.
  • Pause takes a correction back — restores what you typed, switches your layout back and remembers the word (punto_exceptions.txt next to the settings, one word per line, editable). On a word nobody touched it converts it instead, and pressing it again goes back. Ctrl+Alt+L does the same for the word just typed, or converts the selection when there is one. Backspace is just Backspace.
  • Enter and Tab act after the fix. The key that ends the word is held back until the correction lands, so Enter sends the corrected message.
  • Any installed layout variant — read from Windows itself, so "Russian (Ukraine)", Ukrainian (Enhanced) with ґ on its own key, or a UK English board all work. Classic password boxes are left alone.
• Ask the model (off by default — switch it on in Settings → Ask AI; while it's off, its hotkey isn't registered at all) — Ctrl+Tab drops a single input line beside the text you have selected, or in the middle of the screen when nothing is. Type, press Enter, and the answer unfolds underneath while the input stays where it was; the window height follows the reply, so a one-word answer doesn't leave a half-empty panel and a long one scrolls. No caption, no buttons — drag it anywhere by the panel around the fields, and dismiss it with Escape or the hotkey again. Clicking into another window leaves it alone, so an answer can stay on screen while you work underneath it. Whatever text is selected when the hotkey is pressed arrives already in the input, unsent, so a question can be asked about something without pasting it first. Follow-ups keep the thread, and it answers in the current UI language, since a hotkey leaves no room to ask for one. The conversation lives only as long as the window: reopening starts clean.
  • It has eyes. Paste a picture into it and the question is asked about that: a thumbnail appears beside the input, and clicking the thumbnail takes it back off. With a Claude API key set, the picture is sent to the model as a picture — it can read a chart, a diagram, or an interface. Without one, the window falls back to OCR: it reads the text out of the image locally and hands that to DeepSeek, which covers a screenshot of an error, a document or a page of code, but cannot describe what is drawn. Which backend answers is decided by the picture alone — a question with no image always goes to DeepSeek, whether or not the second key is set.
• Taskbar icon centering — dynamically repositions icons to the middle of the taskbar and keeps them centered as icons come and go.
• System tray — its own icon; left-click opens settings, right-click shows a dark "Settings / Exit" menu. The console window is hidden.
• Multi-language UI — settings, popup, and tray menu translated into 13 languages (en, ru, es, fr, de, pt, it, pl, tr, uk, zh, ja, ko).
• Modern dark interface — a frameless settings window with a navigation column (General / Shortcuts / Translation / Ask AI) instead of one long scroll; grouped cards, switches, keycap-style shortcut fields, and API keys masked until you click the eye, each with a live "works / rejected / unreachable" status. The translation popup and dropdowns are floating cards with antialiased corners and a soft shadow (per-pixel-alpha layered windows); the capture overlay keeps its tools and actions in one bar under the selection. Every rounded shape is antialiased — GDI has none of its own, and a stair-stepped corner is the one thing that gives a hand-drawn control away.
• Windows autostart — optional one-click toggle that writes to HKCU\Software\Microsoft\Windows\CurrentVersion\Run.

---

Default hotkeys

Hotkey	Action
Ctrl+Alt+T	Translate the currently selected text
Ctrl+Alt+S	Select a region → OCR → translate
Ctrl+Alt+D	Screenshot a region to the clipboard
Ctrl+Alt+L	Switch the layout of the word just typed, or of the selection
Pause	Take back the last automatic correction (or convert the last word)
Ctrl+Alt+A	Turn automatic layout correction on or off
Ctrl+Tab	Ask the model a question (only once turned on in settings)

Inside the capture overlay, once a region is selected:

Key	Action
Ctrl+C	Copy the region to the clipboard
Ctrl+S	Save the region to a file
Ctrl+Z	Undo the last drawn shape
Esc	Cancel

All hotkeys are remappable from the settings window (click the tray icon).

Once the ask window is on, Ctrl+Tab is registered globally, which takes it away from every other application while the app runs — browsers and editors included. Remap it in settings if you want tab switching back; with the ask window off it isn't taken at all.

---

Build

Requires Rust edition 2024 (stable 1.85+) and Windows 10 / 11.

cargo build --release

The binary is produced at target/release/screen-translator.exe and is fully self-contained — just copy it anywhere and run.

OCR.space API key

The default build uses the public demo key helloworld, which has very low rate limits. Get a free personal key at https://ocr.space/ocrapi and bake it in at build time:

# PowerShell
$env:OCR_SPACE_API_KEY = "your_key_here"; cargo build --release

# cmd
set OCR_SPACE_API_KEY=your_key_here && cargo build --release

# bash / Git Bash
OCR_SPACE_API_KEY=your_key_here cargo build --release

If no key is set and the demo quota is exhausted, OCR automatically falls back to the built-in Windows engine (requires the corresponding language packs to be installed in Windows).

---

Usage

1. Run screen-translator.exe. A tray icon appears next to the clock (possibly hidden under the "show hidden icons" arrow — drag it out to pin it).
2. Select any text in any application and press Ctrl+Alt+T. The translation pops up next to the cursor.
3. Press Ctrl+Alt+S, draw a rectangle on screen with the mouse — the recognized and translated text appears in the popup.
4. Press Ctrl+Alt+D to capture a region straight into the clipboard, ready to paste into any chat or document.
5. To capture a page that doesn't fit on screen, draw the region over the scrollable area and click Full page. Keep hands off the mouse and keyboard while it scrolls; when it stops, a save dialog offers the stitched PNG.
6. To mark something up, pick the pencil or the rectangle at the left of the bar under the selection and drag inside it, then save or copy as usual.
7. Turn the ask window on in Settings → Ask AI, then press Ctrl+Tab, type a question, press Enter. Select text first and it is waiting in the input when the window opens. Shift+Enter breaks the line instead of sending; Esc or the hotkey again closes the window. Ctrl+V attaches a picture from the clipboard instead of pasting text. Requires a DeepSeek key; a Claude key is what lets it see the picture rather than just read the text out of it.
8. Click the tray icon to open settings: remap hotkeys, pick a screenshot folder, toggle Punto / taskbar centering / autostart / the ask window. Tab moves between fields, Ctrl+Tab between pages, Enter saves, Esc closes.

---

Chrome extension: full-page screenshot, save image as PNG / JPG / WebP

The extension in chrome_extension/ does two things.

Screenshot the whole page — the toolbar button, Ctrl+Shift+S, or Screenshot the whole page in the page's right-click menu. The page is captured top to bottom in one go, wherever you are scrolled, without scrolling through it: Chrome renders the parts that are off screen itself (DevTools protocol, captureBeyondViewport). A long page takes about a second. The PNG goes straight to Downloads, named after the page title.

• While it runs (a second or so) Chrome shows its "… started debugging this browser" bar. That's how Chrome labels any use of the protocol, and it can't be switched off from an extension.
• Sticky headers come out once, at the top: the page is put at the top for the capture — instantly, not scrolled — and put back afterwards.
• Images marked loading="lazy" are loaded before the capture. Content a script only adds as you scroll (infinite feeds) isn't there to capture.
• Pages that scroll an inner panel instead of the page itself (web mail, chat apps) come out as one screen; the exe's Full page button handles those by scrolling.
• Pages taller than 32 000 px are cut there — the largest canvas Chrome will encode.
• chrome:// pages and the Web Store can't be captured by any extension.

Save image as PNG / JPG / WebP. Chrome's own Save image as… writes whatever the site served — mostly WebP or AVIF — and the type box in its dialog can't change that. The extension adds a Save image as → PNG… / JPG… / WebP… submenu to the right-click menu on images. It re-encodes the picture and then opens the ordinary Save As dialog with the right extension already in the file name.

It is separate from the exe; no program outside the browser can add entries to that menu, or render a page it hasn't painted.

Install (Chrome or Edge):

1. Open chrome://extensions (edge://extensions) and turn on Developer mode.
2. Click Load unpacked and pick the chrome_extension folder. Chrome loads it from there, so leave the folder where it is. After updating the folder, press the reload arrow on the extension's card.
3. If Ctrl+Shift+S is taken by something else, pick another key at chrome://extensions/shortcuts.

Behaviour worth knowing:

• JPG is written at quality 92; transparent areas become white, since JPEG has no alpha. PNG and WebP keep transparency.
• Animated GIF and WebP are saved as their first frame. SVG is rasterised at its own declared size.
• The image is fetched again by the extension, which is why it asks for access to all sites: pictures usually live on a CDN, not on the page's domain. Hotlink-protected images and blob: images are read from inside the page instead.
• If the image can't be fetched or decoded, the page shows an alert saying so.

---

Dependencies

• windows — Win32 API bindings
• arboard — clipboard access
• ureq — HTTP client
• serde + serde_json — settings persistence
• base64 — image encoding for OCR.space and for the vision API
• chrono — screenshot filename timestamps
• anyhow — error handling

External services used:

• MyMemory Translation API (free, no key required) — default translator
• DeepSeek Chat Completions API (paid, optional) — higher-quality alternative; configure the key in the settings window
• Anthropic Messages API (paid, optional) — what the ask window uses to look at an attached picture; without a key it falls back to OCR
• OCR.space Parse Image API (free, optional key)

---

Project layout

src/
├── main.rs            # entry point, main message loop, hotkey dispatch
├── settings.rs        # settings model, JSON load/save
├── settings_ui.rs     # frameless dark settings window with page navigation
├── tray.rs            # system tray icon (Shell_NotifyIcon)
├── autostart.rs       # Windows Registry Run key for autostart
├── autotype/          # Punto-style layout correction
│   ├── mod.rs         #   keyboard/mouse hooks on their own thread, SendInput
│   ├── engine.rs      #   keystrokes in, corrections out (no Win32; simulated in tests)
│   ├── decide.rs      #   which layout a word was meant for
│   ├── model.rs       #   dictionaries and character trigram models
│   ├── keymap.rs      #   what each key types on each installed layout
│   └── eval.rs        #   accuracy measurement on held-out word lists
├── taskbar_center.rs  # taskbar icon centering
├── capture.rs         # rectangular screen-region selector overlay
├── scroll_capture.rs  # scrolling "full page" capture and frame stitching
├── uia_scroll.rs      # UI Automation scroll driver (wheel is the fallback)
├── screenshot.rs      # pixel capture and PNG encoding
├── ocr.rs             # OCR.space + WinRT OCR
├── translate.rs       # translation policy: DeepSeek or MyMemory, language detection
├── deepseek.rs        # DeepSeek chat-completions client, shared by both callers
├── vision.rs          # Anthropic messages client — the ask window's eyes
├── ask.rs             # "ask the model" chat window
├── popup.rs           # translation result popup window
├── i18n.rs            # 13-language UI string table
├── paint.rs           # antialiased rounded rectangles and arbitrary shapes
├── button.rs          # macOS-style push buttons
├── theme.rs           # macOS dark-appearance colour palette
└── utils.rs           # UTF-16, urlencode, Win32 input helpers

data/punto/            # word models, built by tools/build_punto_dicts.py
tools/                 # the dictionary builder (downloads its corpora)

chrome_extension/      # "Save image as PNG / JPG / WebP" for Chrome and Edge
├── manifest.json
├── background.js      # context menu, fetch fallbacks, Save As dialog
├── offscreen.js       # decode and re-encode; owns the blob: URLs
├── offscreen.html
└── _locales/          # en, ru, uk

---

Settings location

Settings are persisted to %APPDATA%\screen-translator\settings.json. Delete the file to reset everything to defaults.

A log of what the app did is written next to it, at %APPDATA%\screen-translator\log.txt, and truncated on every start. It is the first thing to look at when a capture or a hotkey misbehaves — the scrolling capture in particular records which scroll driver it chose and how far each frame moved. Setting SCROLL_DEBUG_DIR to a folder additionally dumps every grabbed frame there as a PNG.

---

License

MIT