- A custom keyboard that transcribes your voice on-device and optionally polishes the result with the LLM you configure. Works in WeChat, Notes, Mail, ChatGPT — everywhere you type. Up to 60 seconds per take, with a long-lived Flow session so you never bounce back to the host app. + A custom keyboard that transcribes your voice on-device, polishes with AI, and optionally translates — in WeChat, Notes, Mail, ChatGPT, and everywhere you type. Pick a local engine (built-in polish) or cloud engine (your API). Up to 3.5 minutes per take, with an 8-minute Flow session so you never bounce back to the host app.
Native SwiftUI, privacy-first architecture, and a Typeless-style push-to-talk experience.
+Native SwiftUI, on-device transcription, configurable polish intensity, optional translation, and a personal dictionary — all in a tap-to-toggle Flow session.
Tap the mic to start, speak naturally, tap again to stop. A live countdown keeps you honest at 60 seconds per take.
+Tap the mic to start, speak naturally, tap again to stop. A live countdown tracks each take — up to 210 seconds. Delete, space, and return keys sit below the mic.
Speech stays on your iPhone. Apple's on-device speech stack handles transcription — audio never leaves the device.
+Speech stays on your iPhone. iOS 26 SpeechAnalyzer + DictationTranscriber handle transcription — audio never leaves the device.
Optional cloud pass adds punctuation, structure, and clarity. Only the transcript is sent — not your voice.
+Light, Medium, or Heavy polish adds punctuation, structure, and clarity. Ultra-short utterances skip the LLM and insert directly. Only the transcript is sent — never audio.
Works with any OpenAI-compatible endpoint: OpenAI, DeepSeek, Qwen DashScope, or your own server.
+Local: on-device ASR + built-in DeepSeek polish — no API key needed. Cloud: on-device ASR + your OpenAI-compatible API (OpenAI, Qwen, Zhipu, Moonshot, or self-hosted).
No ordinary keystroke logging. You choose when cloud polish runs and which provider receives text.
+ +After polish, optionally translate into English, 简体中文, 日本語, 한국어, and more. Toggle from the keyboard chip or Settings — both engines supported.
Pure Swift 6 and SwiftUI — no CocoaPods, no Carthage, no third-party packages.
+ +Add names, product terms, and jargon in the Dictionary tab. The LLM preserves your entries; aliases help correct common ASR misrecognitions.
A long-lived host-app session keeps the recorder warm. Record many takes without bouncing back to the app.
+A long-lived host-app session (8 minutes) keeps the recorder warm. Record many takes without bouncing back to the app — tap Start on the keyboard if the session expires.
Local mode is ASR-only. Enable cloud polish to route the transcript through your chosen LLM when iOS speech recognition isn't strong enough.
+ +Drag the pads beside the mic to move the caret precisely — horizontal or vertical. Toggle on/off in Settings → Preferences. Left- and right-handed layouts supported.
Three gestures between thought and sent message.
+From first launch to polished text in any app — about five minutes.
Build from source and launch. The 5-step onboarding starts the Flow session automatically once permissions are granted.
+Build from source and launch. The 5-step onboarding walks you through permissions, keyboard setup, and engine choice. Flow session starts automatically when ready.
Grant microphone and on-device speech recognition access when prompted. Both stay on-device.
+Grant microphone and on-device speech recognition when prompted. You can also finish most steps from the keyboard's own setup overlay.
In iOS Settings, add OSGKeyboard and allow Full Access so it can use the mic and your API.
+In iOS Settings, add OSGKeyboard and allow Full Access — required for the mic, API key, and Flow session bridge.
Choose local ASR or cloud. For cloud polish, paste any OpenAI-compatible key (DeepSeek works out of the box).
+Local: on-device ASR + built-in polish, no API key. Cloud: on-device ASR + your OpenAI-compatible key (OpenAI, Qwen, Zhipu, Moonshot, or custom).
Switch to OSGKeyboard in any text field, tap the mic, speak (up to 60s), tap again. Polished text lands at your cursor.
+Set polish intensity, translation target, and personal dictionary in Settings. Switch to OSGKeyboard, tap the mic, speak (up to 3.5 min), tap again — text lands at your cursor.