Cursay turns your voice into text in the application you are already using. Hold the global shortcut, dictate for as long as you need, and release. Cursay transcribes the recording, cleans the text, verifies the clipboard, and pastes it at your cursor.
It is built for people who want a free, native Wispr Flow-style workflow without requiring an account or sending every thought through a subscription service. Cursay Pro is an optional managed-cloud path; it never replaces unlimited local dictation.
| Platform | Status | Start here |
|---|
| Ubuntu GNOME on Wayland | Stable public release: 1.2.1; 1.3.0 release candidate | Install on Ubuntu |
| Windows 10/11 x64 | Native release — build artifact | Windows guide |
| macOS 13+ | Native app — signed release pipeline | Mac build guide |
The whole interaction
- Focus any text field.
- Hold Ctrl + Space.
- Speak naturally—even with pauses.
- Release the keys.
- Your dictation appears at the cursor.
Silence does not end the recording. Releasing the shortcut does.
Why Cursay
- System-wide: dictate into browsers, chat apps, editors, terminals, and documents.
- Local by default: faster-whisper runs on your machine after a one-time model download.
- Optional Pro cloud: managed multilingual transcription and Smart Polish with a 1,500-minute monthly allowance and local fallback.
- Private history: searchable transcripts stay on your device.
- Dictation speed: track words per minute by session and over 14, 30, or 90 days in native Insights, including your existing history.
- Cost visibility: Insights estimates known provider charges from recorded audio duration and uses exact
provider-reported costs when available.
- Strict push-to-talk: recording follows the key press instead of guessing when you finished speaking.
- Safe automatic paste: Cursay confirms that the clipboard contains the new transcript before sending
Ctrl+V.
- Wayland-native shortcut: uses the Global Shortcuts portal instead of a keylogger.
- No Remote Desktop access: paste is performed through a private keyboard-only
uinput device.
- Useful cleanup: professional, casual, prompt, code, and raw modes; optional filler-word removal and AI rewriting.
- Comfortable appearance: follows your system light/dark preference, with a manual override.
- Backend-friendly: use the included local service or another OpenAI-compatible transcription endpoint.
Build on macOS
The native SwiftUI Mac app includes a menu-bar controller, configurable push-to-talk shortcuts, native microphone capture, automatic paste, private local history and cost insights, appearance controls, all five writing modes, Smart Polish, and the same Local Whisper, Cursay Cloud, and Custom provider choices as Ubuntu. Local Whisper is bundled and starts on demand. Tagged releases are built as hardened-runtime, Developer ID-signed, notarized, stapled DMGs with signed Sparkle updates.
On macOS 13 or newer with Xcode 15 or newer and Python 3.10 or newer:
cd macos
./scripts/build-app.sh
open dist/Cursay.app
See the Mac build guide for permissions, transcription-service setup, and distribution notes.
Install on Windows
Download and run Cursay-<version>-windows-x64-setup.exe from the matching release. The per-user installer includes the native app and private local speech service. Cursay lives in the notification area and uses Ctrl + Space for push-to-talk.
The first local dictation downloads the English base.en Whisper model. Windows may ask you to allow microphone access for desktop apps. See the Windows guide for build instructions, data locations, and current limitations.
Install on Ubuntu
git clone https://github.com/shadoprizm/cursay.git
cd cursay
./install.sh
The installer is per-user. It places the application in ~/.local/opt/cursay, installs user-level services, downloads wl-clipboard and ydotool from Ubuntu without installing them system-wide, and prepares the local Whisper runtime.
If GTK or the Python desktop bindings are missing, the installer will print the exact Ubuntu command needed. The local base.en model is roughly 150 MB and is downloaded once.
After installation, open Cursay from the app launcher. Ubuntu will show one Global Shortcut permission dialog. Approve the suggested shortcut or choose another one.
If the installer reports that /dev/uinput is protected, enable automatic paste with the included one-time system rule:
./scripts/enable-autopaste.sh
Ubuntu will request administrator authentication to install that narrow rule. Without it, transcription and clipboard copy still work; only the final automatic Ctrl+V is unavailable.
Installer options
./install.sh --no-start # install without starting Cursay
./install.sh --skip-model # download Whisper on first local dictation
./install.sh --skip-backend # use an existing compatible STT endpoint
Installed Ubuntu builds can check for updates from Settings → Application updates. Cursay downloads the
versioned release archive and checksum from GitHub, verifies the archive before installation, preserves settings
and history, and keeps your existing local-backend choice.
Ubuntu data
| Data | Default location | Policy |
|---|
| Settings | ~/.config/cursay/config.json | Private to your user |
| Dictation history | ~/.local/share/cursay/history.db | Stored locally |
| Whisper model | ~/.local/share/cursay/models/ | Downloaded once |
| Temporary audio | ~/.cache/cursay/recordings/ | Deleted after transcription |
| Preserved audio | ~/.local/share/cursay/recordings/ | Only when explicitly enabled |
Existing Local Flow settings and history are copied automatically the first time Cursay runs. The original data is left untouched as a recovery copy.
How it works
Global shortcut
↓
PipeWire microphone capture
↓
Local faster-whisper (or your compatible endpoint)
↓
Meaning-preserving text cleanup
↓
Wayland clipboard write + read-back verification
↓
Private keyboard-only Ctrl+V
The shortcut portal can observe only Cursay's approved shortcut. The paste device is launched with mouse support disabled and sends only the four key events needed for Ctrl+V.
Transcription choices
Choose Local Whisper, Cursay Cloud, or Custom in Settings. Local Whisper remains free, unlimited, accountless, and entirely on-device after the model download. Cursay Cloud is included in the optional $8/month or $80/year Pro plan, includes 1,500 minutes in each monthly service window, and falls back to Local Whisper when enabled.
Ubuntu stores Pro device tokens only in GNOME Secret Service, which requires secret-tool from the optional libsecret-tools package. macOS stores them only in Keychain. Cursay does not fall back to a plaintext token file.
The included backend uses faster-whisper with the English base.en model on CPU. To use a different model or language, edit the Cursay STT user service or the settings file.
Cursay can also call an OpenAI-compatible /v1/audio/transcriptions endpoint. Set stt_endpoint and stt_model in ~/.config/cursay/config.json. When you choose a remote endpoint, audio is governed by that provider's privacy policy; the included backend remains entirely local after its model is downloaded.
Native Insights → Dictation speed shows weighted period and all-time WPM averages, your latest session's
speed, and a daily trend for the last 14, 30, or 90 local calendar days. Speed is final transcript words divided
by recording minutes, including pauses. Transcription and Smart Polish wait time are excluded. Existing
history is included; sessions without words or timing do not affect speed averages. Private captures are never
added to history or analytics. These statistics are local to each device.
On Mac and Ubuntu, the Memory page links to the account workspace for saved thoughts, Ask Cursay,
reviews, projects, and confirmed vocabulary. Windows has a workspace action in Insights and Settings.
Desktop sync requires a linked account, separate workspace consent, and opting into future-capture sync.
The workspace reports account feature availability; see release status for rollout gates.
The Insights page shows a provider/model cost breakdown. xAI REST speech-to-text is estimated from locally
recorded duration at its published rate; unknown custom providers are left
unpriced rather than guessed. Provider invoices remain the final billing authority.
Writing styles
The style picker has two kinds of behavior:
- Professional, Casual, and Prompt always apply basic spacing, capitalization, and optional filler removal. Turn on Smart polish for a genuine rewrite through the OpenAI-compatible text endpoint configured by
polish_endpoint and polish_model.
- Prompt turns spoken intent into a ready-to-paste AI prompt. Simple requests stay concise; complex requests are organized by goal, context, requirements, constraints, and desired output when those details were actually spoken.
- Code converts spoken tokens such as “open paren” and “new line” without sending the result through Smart polish.
- Raw preserves the transcript apart from leading and trailing whitespace. It bypasses filler removal and Smart polish.
Smart polish is disabled by default. With Local Whisper or Custom, its privacy depends on the text endpoint you configure. With Cursay Cloud it uses the one-use polish grant from the immediately preceding cloud transcription. The Dictate screen shows whether it is active and reports a visible fallback when that service is unavailable.
Correcting dictations on Ubuntu
Click Correct beside Latest dictation or a History entry. Edit the text, or paste in the
corrected version from your email or another app. Click Review replacements to remember and select
only recurring recognition mistakes (for example, curse say → Cursay). Nothing is remembered until
you explicitly select a replacement and save. Save correction and copy updates the local history
entry and clipboard; paste that result into your destination app yourself.
The Corrections page lets you add or update a replacement, disable it, or delete it. Rules match
whole words and phrases without regard to capitalization, use your replacement's exact spelling,
and apply after basic cleanup and Smart polish. Raw mode bypasses them. Rules and corrected history
stay on this Ubuntu device, separately from account Memory sync. Editing an external text box alone
does not create a correction or a rule. This is explicit replacement behavior, not model retraining.
History shows the corrected text's word count, while session, daily, period, and all-time WPM keep
using the originally generated transcript. Typed additions or deletions never change voice speed.
Ubuntu basic filler cleanup removes um, uh, and erm, including elongated variants. It preserves
you know, I mean, ah, and er because these can carry meaning. The Insights filler counter totals
the logged basic-cleanup removals in the latest 500 history entries; older entries may include phrases
removed by earlier versions. Smart polish can independently remove fillers, and those edits are not
included in this counter. Raw mode bypasses both cleanup and polish.
Ubuntu installations built from source show the Git revision in the sidebar, a local changes
indicator when the source checkout had unpublished edits, and a fingerprint of the installed application
source. Hover over the build label for the full revision and fingerprint. Source archives without Git
metadata are labeled unversioned and still receive a source fingerprint.
Ubuntu diagnostics
~/.local/opt/cursay/bin/cursay --check
~/.local/opt/cursay/bin/cursay --record-test
systemctl --user status cursay.service cursay-input.service cursay-stt.service
journalctl --user -u cursay.service -n 100 --no-pager
Shortcut does nothing
Open Cursay, go to Settings → Change shortcut, and approve the shortcut in Ubuntu's dialog. The application must be installed before the portal can associate the permission with Cursay.
Text is copied but not pasted
Run ./scripts/enable-autopaste.sh, then check systemctl --user status cursay-input.service. Cursay will never paste an older clipboard value: if clipboard verification fails, it stops after copying the new text.
A Remote Desktop dialog appears
Cursay does not use the Remote Desktop portal. An old experimental Local Flow process may still be running; reinstalling Cursay disables and archives those old user services.
Development
The Ubuntu application intentionally relies mostly on Ubuntu's system Python packages. Run its tests with:
/usr/bin/python3 -m unittest discover -s tests -v
/usr/bin/python3 bin/cursay --smoke-test
On macOS, run the Swift tests and assemble a local app bundle with:
cd macos
./scripts/build-app.sh
On Windows, run the .NET tests and build the self-contained installer with:
.\windows\scripts\build.ps1
The account website and managed API live in website/. See CONTRIBUTING.md for its Next.js, Neon, Stripe, Clerk, and AI Gateway development checks.
The local STT service declares its direct dependencies in backend/requirements.txt and installs the fully pinned,
hash-verified dependency set from backend/requirements.lock.
Contributions are welcome. Read CONTRIBUTING.md before opening a pull request, see CHANGELOG.md for release notes, and use SECURITY.md for vulnerability reports.
Uninstall
./uninstall.sh
Uninstalling preserves settings and dictation history. The script prints the data directory if you also want to erase it manually.
License
Cursay is available under the MIT License.
Cursay Memory
Future dictations can sync to account Memory after separate consent at https://cursay.com/app and enabling sync on each device. Past local history is not imported. The browser workspace provides saved thoughts, search, Ask Cursay, reviews, projects, explicit vocabulary corrections and reviewable suggestions. Intelligence uses OpenAI GPT-6 Luna through Vercel AI Gateway; speech and Smart Polish keep their existing providers.
Private capture keeps a thought out of local history, retained audio, and Memory. App exclusions fail closed when application identity is unavailable. Only approved vocabulary can inform Local Whisper hints or opted-in Cloud Smart Polish, and Raw/Code modes remain unchanged. Account-bound offline queues apply server deletions before uploads; forgetting Memory invalidates old queued data. Windows uses DPAPI for cloud credentials and its Memory outbox.
The browser dictation workspace is live, and native capture sync is in current source. Production stored
Memory and paid credit top-ups remain disabled pending the owner and storage-readiness gates recorded in
release status. Desktop source and cloud source live in separate repositories;
see workspace alignment for their roles and verified locations.
