Say the rough
thing first.

Hold your global shortcut (F8 by default), speak, and release to insert text at the cursor. Change the shortcut in Settings. Multiline output is typed with native key events; there is no separate document to copy from.

Use your own Groq or OpenAI key for speech recognition. Choose Groq, OpenAI, OpenRouter, Anthropic or Gemini independently for text refinement. Browse available models with pricing labels. Provider accounts, model availability, internet access and API charges are separate from the app.

Balanced, Casual, Formal and Developer presets, plus individual controls for filler removal, grammar and punctuation, code casing, emojis and Markdown lists. An illustrative rule simulator previews a fixed example locally. Select an output translation language or turn translation off to keep the original. Normal dictation transcribes questions instead of answering them; provider failures can return the raw transcript.

Optional Vibe Coding compresses spoken instructions into a dense, keyword-led prompt. It is a deliberate text transformation, not a coding agent or a guarantee that every nuance will survive compression.

Save incorrect-to-correct word pairs for names, specialist terms and preferred spellings. These corrections guide text refinement; they do not train a personal speech model.

Create, edit and delete trigger phrases for reusable emails, signatures and other text blocks. Matching triggers expand locally after refinement, including multiline content.

Opt in to Clipboard AI and choose its trigger words. A triggered spoken command passes the clipboard text to your selected refinement provider, for example to rewrite or translate it. The clipboard is context only when this mode and its trigger are used.

Assign additional shortcuts and spoken toggle phrases for Vibe Coding, grammar, filler removal, emojis, lists, Clipboard AI and optional Enter after insertion. An exact spoken toggle changes the setting rather than inserting the command as dictation. Auto-Enter can submit text in the target app, so enable it consciously.

Inspect timestamps, the raw transcript and final text, copy an entry and replay its saved audio. Delete an entry or clear the history, and choose Off, 100, 1,000, 10,000 or unlimited entries. Turning history off does not erase existing entries. History and recordings are stored locally; unlimited history uses disk space and local storage is not a promise of encryption.

Choose Wave, Bars, Pulse or Line recording visualizers with a live preview. Use the skippable onboarding tour, collapsible settings, system tray and single-instance behavior. German and English interface selection is separate from dictation and voice languages; some labels remain untranslated. The support panel provides a support-email link, a feedback email draft and a donation link; drafting does not send the email.

Enable the QWERTZ keyboard in Android Settings, then select it as your input method. Use letters, numbers, symbols, punctuation, Shift and Caps Lock. Choose Dark, Light or AMOLED, three keyboard heights and an optional number row, with live layout previews.

Long-press letters for umlauts, ß and €. Move the cursor by character or word, delete words, or swipe on Backspace to preview a word range before deleting it. Optional glide typing ranks words locally from the traced keys, context and history; it is a lightweight deterministic decoder, not Gboard’s proprietary model.

Record only during an explicit hold, with a live microphone visualizer. With a saved OpenRouter key, the keyboard sends a bounded in-memory audio request to a compatible audio-to-text model. A downloaded local Whisper package is an optional offline fallback. A missing backend shows a status message without pulling you out of the editor.

The same key is used for cloud dictation, deliberate writing assistance and conversation voice output. Account metadata discovery chooses compatible text and audio models where available, and you can configure the text model. No automatic switch to a different cloud provider.

Choose Casual, Formal, Developer or Vibe style and Original, German, English, Spanish, French or Italian output. Individually control hesitation-word removal, grammar, spoken punctuation, sentence capitalization, dictionary spelling, snippet expansion, spoken lists, code casing and optional emojis. Original-language dictation preserves the original language and script rather than answering the dictated request.

Save exact spellings and multiline shortcuts such as /mail or /bug. Edit these in Settings and clear the local dictation history. Android history is limited to 30 entries in app-private storage; API keys are encrypted, but the history itself is not encrypted.

Open the assistance panel only when you want to rewrite, shorten, translate or draft a reply. Selected text takes priority; otherwise the paragraph before the cursor is used, bounded to 2,000 characters. Type an explicit instruction inside the panel, review the proposal, apply or discard it, and undo while the editor still matches. No automatic message sending.

The assistance panel also offers Read / Stop for selected text using an installed Android voice that reports offline in the device language. This separate selection reader has no silent cloud fallback; engine behavior remains device-dependent. It is not the conversation mode’s OpenRouter voice route.

Open the full-screen conversation from Settings or the keyboard’s translation entry. Choose a pair from the 59-language ML Kit text catalog (German ↔ Thai by default), speak or type, edit the recognized original, and review the translation. Switching text or language invalidates stale output. Text-model downloads require consent and Wi-Fi; installed models can translate offline. Non-English pairs may translate through English.

Discover compatible OpenRouter voices and explicitly play the exact reviewed target text with your existing key. Voice availability and declared language coverage depend on the account’s model catalog; the 59 text languages do not promise 59 available voices. Conversation playback has no Android system-voice fallback and needs a compatible online audio-output model.

Start from selected text in a non-private field, review the translated result, then return it to the keyboard. A second explicit tap inserts it only if the same app and original selection still match. It never sends the message for you.

Password, private and no-personalized-learning fields disable dictation, assistance, history and contextual suggestions. Asynchronous results are tied to the original editor, text and selection. Android cloud backup and device transfer are excluded. Changing fields cannot retract a cloud upload that has already started.