SayKuku
简体中文 · Download · Setup guide (Chinese) · Privacy (Chinese)
Promotional illustration. Actual app screenshots are below.
SayKuku is a free, native macOS voice app. Press Fn to dictate at the cursor; press Fn twice to tell Voice Agent what you need. Ask a question, draft text, or follow up without selecting anything; select text when you want it rewritten or translated. It uses Qwen cloud models, so you bring your own Alibaba Cloud Model Studio API key and pay any model charges on your account.
Two ways to speak
| Gesture | What happens |
|---|
| Fn | Voice Input transcribes speech and tries to insert the result at the original cursor. Choose hold-to-talk or tap-to-start/tap-to-stop. Adjust language, number formatting, and spoken-word cleanup in Settings. |
| Fn Fn | Voice Agent uses your request and the context you allow to draft text, answer a question, or follow up without a selection, or rewrite and translate selected text. It checks the original target before writing. Running a Shortcut requires confirmation; opening a link or searching the web requires confirmation when untrusted context is involved. |
You can also use configurable global shortcuts. The defaults are ⌃⌘V for Voice Input and ⌃⌘A for Voice Agent. A single modifier can start Voice Input, and Voice Agent can use a double tap of that same key. Writing into another app depends on that app's text controls; check important text before sending it.
A look at the app
Home screen from a signed development build. The interface is shown in Simplified Chinese.
Privacy settings from a development build. These are real UI screenshots; the image notes describe their source and limits.
Get started
You need macOS 14 or newer, an internet connection for Qwen processing, and your own Qwen API key. Model usage may incur charges.
- Download the signed, notarized DMG and drag SayKuku into Applications.
- Open the app and allow Microphone and Accessibility access when guided. Accessibility supports Fn and single-modifier gestures, reading an allowed selection, and writing to another app.
- Choose the Qwen region that matches your API key, enter the key, and add a Workspace ID if your account uses one. The key is stored in macOS Keychain.
- Put the cursor in a text field and press Fn. Press Fn twice and ask a question to try Voice Agent—no selection needed. To rewrite text, select it first and say what to change.
If Fn conflicts with a macOS keyboard setting, follow the in-app guidance or use the configurable global shortcuts. The setup guide (Chinese) covers permissions and Qwen configuration.
More than dictation
- Memory: add names, projects, organizations, and terms; import suggestions from pasted text; or approve a correction after dictation. Review or edit saved entries in the Memory page.
- History: search, replay saved recordings, retry a failed dictation when audio exists, star entries, and delete records. Voice Agent's recent conversation stays in memory for up to 30 minutes and is cleared on quit.
- Usage: view daily, weekly, and monthly dictated characters, dictation counts, speaking time, active days, input trends, top apps, and Voice Agent uses. Only verified dictation insertions count as input; confirmed Undo adjusts the totals. Open with ⌘4.
- Control: change input style, overlay position, language, shortcuts, automatic writing, context sources, history retention, and analytics in Settings. The interface supports English and Simplified Chinese.
Privacy and data
| Data or permission | What SayKuku does |
|---|
| Microphone | Captures audio only for Voice Input, Voice Agent, or a microphone level test. Voice audio goes to the selected Qwen region; the level test is neither saved nor uploaded. |
| Voice Agent context | Selected text, current app, window title, and visible text in the current window are allowed by default. Clipboard text and the Safari/Chrome page URL are off by default. Each source has a switch in Settings → Privacy. Visible text is limited to 2,000 characters and is sent only with that request, without being stored in History or logs. Voice Input does not read screen text. |
| Voice Input context | With Use text near the cursor on (the default), Voice Input sends up to 800 characters before and 200 after the cursor in the focused text field, leaving out any selection, so names, terms, and similar-sounding words match what you are writing. It goes only with that request and is not stored in History or logs. Turn it off in Settings → Voice Input. Password fields, secure input, and password managers are never read. |
| Voice Input cleanup | Automatic cleanup removes fillers, repeats, and self-corrections during recognition. Automatic paragraphs and lists is on by default; turning it off retains explicitly dictated formatting. Cleanup preference → Prefer quality sends the transcript, never the audio or nearby cursor text, to a Qwen text model in the same region for one extra check. Long text uses a compact prompt and waits up to 12 seconds; short text waits up to 2 seconds. Failures keep the first result. Prefer speed is the default for users who have not chosen a preference and skips this check. Keep spoken wording preserves fillers and repeats. Each app can override these choices in Settings → Voice Input → Per-app settings. |
| Local data | The API key is stored in this Mac’s Keychain. History, Memory, correction suggestions, and local usage totals are saved in ~/Library/Application Support/SayKuku/store.json; saved recordings are in ~/Library/Application Support/SayKuku/Audio/*.wav. |
| History defaults | New history is kept for 30 days and recording storage is on by default. Starred entries are exempt from automatic deletion. Settings → History → Don’t keep stops new history and recordings; it does not delete older entries. |
| Local usage statistics | Daily totals contain counts, speaking time, app names, and task categories, without text or audio. Collection starts with this feature; older history is not backfilled. Days follow the local time zone at collection and weeks follow regional preferences. Comparisons use equal numbers of complete days with uninterrupted collection. History deletion does not clear totals. Pause or clear them separately in . Turning off omits app names from new statistics. |
SayKuku blocks recording and content access in secure text fields and known password managers. Private browsing windows are not detected separately. It does not request Input Monitoring permission or record keystrokes. Read the full privacy and permissions explanation (Chinese) before enabling optional context sources.
Build and contribute
The project uses Swift 6, SwiftUI, AppKit, and Swift Package Manager, with no third-party Swift package dependencies. It targets macOS 14+; packaging requires Xcode 26 or newer for the macOS 26 SDK.
swift test
Scripts/package-app.sh debug
open Build/SayKuku.app
The debug package requires a stable Apple Development signing identity. swift run SayKuku is useful for source-level work, but its process identity and resources do not represent a signed App Bundle; use the packaged app to verify permissions, Keychain access, menu bar behavior, and resources. Default tests do not need a live Qwen API, network, microphone, or Accessibility authorization.
Release builds are signed, notarized, and packaged locally and published to the existing R2 bucket and website through Scripts/release.sh --notes FILE; the repository does not use GitHub Actions for macOS release validation. Release builds use Sparkle 2 for background downloads and install on normal Quit without relaunching; automatic checks and downloads default to on. Follow the local release guide rather than running a Release package step on its own.
License
SayKuku source code is licensed under Apache License 2.0. The Lucide-derived icon assets carry their own notices.
