Hold fn and talk. Let go, and the text is typed at your cursor, in any app. Esc throws a dictation away.
For macOS 14 and later
Szept is a light, fast Mac app for talking to AI agents like Claude Code, Codex and Cursor. Hold fn and talk, and the words land at your cursor. Point at the screen, and your agent gets the screenshot along with what you said.
Dictate, screenshot, record, hand off: the four things Szept does, in the order they chain together.
Light and fastno speech model on your Mac
60+languages, switching mid-sentence
~180 msfrom letting go to final text
$0.12per hour of speech, on your Soniox key
Features
Everything it does, on one screen.
Two shortcuts to remember, fn for words and ⌘ ⇧ A (Command-Shift-A) for the screen. Everything else is one click from there.
Removes the ums, the repeats and the false starts, fixes punctuation and keeps your words. It also checks which app the text is going to: an email gets paragraphs and your saved sign-off, a Slack message stays one line, and a terminal gets commands exactly as spoken.
Same sentence, three destinations
What you said
hi magda so um can we move the call to thursday no friday and uh I'll send you the deck before before then
What gets pasted ✦ Cleaning up…Hi Magda, Can we move the call to Friday? I'll send you the deck before then. Best, Mariusz
Press Command-Shift-A and the screen freezes. Pick a window or drag a region, then mark it up, blur, pin it on top, copy the text out of it or read a QR code.
AnnotateBlurPin to screenImage to textQR codeTranslateScrollshot
Press Option-Shift-R to record a region or the whole screen, with your Mac's own sound and your microphone. Add a camera bubble and click effects, trim it, and save as MP4 or GIF.
Send it to your agent
fn
With a screenshot or a recording on screen, hold fn and say what you want. Szept pastes your words and the file into the app you came from, such as Claude Code, Cursor or a chat. You press return.
- Every dictation is kept on your Mac and searchable in any language. Typing “wdrozenia” finds “wdrożenia”.
- Screenshots and recordings sit in the same timeline, and in a Media tab laid out like a contact sheet.
- ↑ ↓ to move, ↵ to copy. Open it from any app and paste an old sentence again.
HistoryMediaSettings
Today
Refactor the upload handler before Friday11:47 · EN
Plan wdrożenia na poniedziałek, potem demo10:05 · PL
Screenshot 14.02.1109:58 · PNG
Yesterday
Hey Anna, thanks for the notes on the deck18:32 · EN
Das Meeting ist um drei, zgadza się?16:03 · DE
1 of 214 · 38 min this month · stored on this Mac
Languages and words
60+ languages
Switch languages mid-sentence and Szept keeps up. Add names and jargon to your dictionary so they come out spelled right.
PolskiEnglishDeutsch EspañolFrançaisУкраїнська Italiano日本語Português+51 more
Light, quiet, cheap to run
No model on your disk. No meter running for us.
There is no Whisper build inside and no Electron around it. Speech recognition runs in the cloud on your own Soniox key, so the app stays small, the Mac stays quiet, and an hour of talking costs cents. Every number here was measured on the shipping build.
An hour of live speech
Szept, on your Soniox key$0.12
ElevenLabs Scribe v2 Realtime$0.39
Google, live~$0.54
Deepgram Nova-3, multilingual$0.55
Azure, real-time~$1.00
OpenAI, live$1.02
Real-time list prices of each speech API, per hour of audio. Szept charges nothing on top.
What a local engine carries
Szept4.9 MB
Whisper tiny, quantised31 MB
Whisper base142 MB
Whisper small466 MB
Whisper large-v32.9 GB
Bars are to scale, and Szept's is a sliver on purpose: it keeps no model on disk and none in memory. Model sizes from whisper.cpp's model repository.
- 0.0%
- CPU while idle
- 24–32 MB
- memory at rest
- 0.1 / s
- idle wakeups (Finder: 38 / s)
- ~180 ms
- from key-up to final text
Measured with footprint and top: 0.0% of a core across five one-second samples, and 82 idle wakeups in fifteen minutes against Finder's 34,389. The 180 ms runs from letting go of fn to Soniox's final text, against the live service.
Watch
The whole app in 49 seconds.
Dictation, cleanup, screenshots, recording and the handoff to your agent, one after another, in the order you would use them.
What you get
One purchase. The app, the code, and the context.
You get a finished Mac app to use today, plus everything an AI agent needs to change it for you tomorrow.
The app
A signed, drag-to-install DMG for macOS 14 or later. It is signed with Developer ID and notarized by Apple, so it opens like any other download.
The source
The complete Swift project in a private GitHub repository: the app, its tests, and the scripts that build, sign and package it.
The agent kit
The context that makes Claude Code, Cursor or Codex useful on this codebase from day one.
- 78
- pitfalls written down in CLAUDE.md
- 638
- tests, all passing
- 12
- implementation plans
- 5
- design specs
A year of updates
Every release for twelve months, as a new DMG and in the repository. After that, everything you have stays yours.
Why the kit matters: things an agent can't guess
An agent can write a dictation app in an afternoon. What it can't know is what only breaks on someone else's Mac. A few of the 78:
- The permission switch that isn't there. A sandboxed Mac app can't add itself to the Accessibility list, so onboarding points at a switch nobody can find.
- The crash that only happens on other Macs. SwiftPM's
Bundle.modulefinds its resources on the machine that built the app and crashes everywhere else. - On in System Settings, off in the app. An ad-hoc signature changes on every build, so macOS shows Accessibility granted while the app is still refused.
- The frame that never repeats. ScreenCaptureKit doesn't resend an unchanged frame, so "wait for three identical frames" never fires on a live stream.
License
Yours to change. Not ours to give away.
The Szept Source License 1.0 is based on the PolyForm Internal Use License 1.0.0. Here it is in plain words; the full text is the LICENSE file that comes with the source. Personal covers one person, Team up to ten people in one company.
You can
- Use it on all your Macs, for yourself or for your work
- With Team: share it, the source and your builds with the ten people it covers
- Change anything: features, the speech engine, the look
- Build and run your own versions
- Keep your changes private, forever
You can't
- Share the source or your builds
- Publish or sell a fork, on GitHub, a website or any app store
- Use the Szept name or icon for your builds
Built something other people should have? Write to [email protected].
Pricing
Pay once. The app and its source.
Founding price · first 100 buyers
After the first 100 it is $25 for the same thing. Updates renew at $9.99 a year, or not at all; nothing you have stops working.
Checkout by Polar, taxes included. You'll be asked for your GitHub username.
- The Szept app for Mac, signed and notarized
- Private GitHub repository with the full Swift source
- The agent kit: CLAUDE.md, specs, plans and tests
- Twelve months of updates, then renew for $9.99 a year or don't
- 14-day refund, no questions asked
Speech recognition runs on Soniox, on your own key. Real-time speech costs about $0.12 per hour of talking, billed by Soniox. Szept never charges for it.
Team
$199once · up to 10 people
- Everything in Personal, for ten named people in one company
- One invoice, taxes handled by Polar
- MDM deployment kit: permissions profile, preset settings, install notes
- A place can move to someone new when a person leaves
Company
Let's talkmore than 10 people
- A written licence for the whole company, or the part that needs it
- Help rolling Szept out through Jamf, Kandji or Intune
- Your own Soniox project and your own cleanup endpoint, set centrally
Questions
Questions about Szept
Szept at a glance
| Platform | macOS 14 Sonoma or later |
|---|---|
| Size | 4.9 MB installed, no speech model on disk |
| At rest | 24–32 MB of memory, 0.0% CPU |
| Languages | 60+, switching mid-sentence |
| Speech engine | Soniox stt-rt-v5, on your own API key |
| Speech cost | About $0.12 per hour of talking, billed by Soniox |
| AI cleanup | Optional; DeepSeek by default, or any OpenAI-compatible endpoint |
| Shortcuts | fn dictate · fn + ⌘ one-off cleanup · ⌘ ⇧ A screenshot · ⌥ ⇧ R record · ⌘ ⇧ H history · Esc cancel |
| Price | Personal: $9.99 once for the first 100 buyers, then $25 · Team (10 people): $199 once · Company: on request; no subscription |
| Included | The app, the full Swift source, the agent kit, 12 months of updates |
| License | Szept Source License 1.0, based on PolyForm Internal Use 1.0.0 |
| Refund | 14 days, full amount |
What is Szept?
Szept is a Mac app for voice dictation, AI text cleanup, screenshots and screen recording, built for working with AI coding agents. Hold fn and speak, and the text is typed where your cursor is, in any app. It costs $9.99 once and includes the source code.
How is it different from the dictation built into macOS?
Szept adds AI cleanup that removes filler words and shapes the text for the app it lands in, recognises more than 60 languages with switching mid-sentence, keeps a searchable history, and includes screenshots, screen recording and a handoff to AI agents like Claude Code.
Does it work in every app?
Yes. Szept types at the cursor in any app that accepts text: Mail, Slack, Notes, a browser, VS Code, Xcode, a terminal. If an app refuses the paste, the text stays on the clipboard so you can press Command-V.
Which languages does it support?
More than 60, including Polish, English, German, Spanish, French, Ukrainian, Italian, Portuguese and Japanese. You can switch languages in the middle of a sentence.
Does it work offline?
No. Speech recognition runs in the cloud on Soniox, which is why the app stays light and needs no model on disk. Screenshots, recording and your history work without a connection.
How much does speech recognition cost?
About $0.12 per hour of talking, billed by Soniox to your own account. Szept never charges for speech and runs no servers of its own.
Do I need to be a developer?
No. The app works as soon as it is installed. The source is there for the day you want your agent to change something.
Where does my data go?
Audio goes to Soniox under your own account. AI cleanup, if you turn it on, sends text to the provider you pick. History and captures stay on your Mac. The details, including regions and certifications, are in the next answers and in the privacy policy.
Where is my voice processed, and is it stored?
Szept streams your audio over TLS 1.2+ to Soniox's United States region. In real-time mode Soniox does not store audio or transcripts, and never uses them to train its models; it keeps only usage metadata such as duration, without audio or text. Szept itself runs no servers, so nothing reaches us.
Which security certifications cover the speech service?
Soniox states it holds SOC 2 Type 2 and ISO/IEC 27001:2022, and meets GDPR and HIPAA requirements; the reports are available in the Soniox Console. Szept has no servers of its own, so there is no Szept backend to certify.
Can my speech data stay in the EU?
Soniox offers EU, Japan and India regions for a project, but Szept currently connects to Soniox's default United States endpoint. Because you get the source, switching is a change of two addresses to the EU endpoints plus an API key from an EU project; ask your agent to do it.
Where does AI cleanup send my text?
Only if you turn cleanup on. The default provider, DeepSeek, states that it stores data in the People's Republic of China and may use it to improve its models unless you opt out. For sensitive text, point cleanup at another OpenAI-compatible provider you trust, or leave it off: dictation works without it.
Which AI does the cleanup use?
DeepSeek by default, or any OpenAI-compatible endpoint you set. If it takes longer than 3 seconds, your words are pasted as dictated.
Can I send screenshots to Claude Code or Cursor?
Yes. With a screenshot or recording on screen, hold fn and say what you want. Szept pastes your words and the file path into the app you started from. Nothing is sent until you press return.
What are the keyboard shortcuts?
Hold fn to dictate, fn with Command for a one-off AI cleanup, Command-Shift-A for a screenshot, Option-Shift-R to start and stop a recording, Command-Shift-H for history, and Esc to cancel.
Which Macs does it run on?
Any Mac with macOS 14 Sonoma or later.
Will macOS let me open it?
Yes. It is signed with a Developer ID and notarized by Apple, so it opens like any other download.
Is there a subscription?
No. You pay once. Twelve months of updates are included; after that you can renew updates for $9.99 a year, or keep what you have.
Do you sell licences for teams?
Yes. Team is $199 once for up to ten people in one company, with the MDM deployment kit: a privacy-permissions profile, settings IT can preset, and install notes. For more people, write to [email protected] for a Company licence.
Can I get a refund?
Yes. Write to [email protected] within 14 days of buying and you get the full amount back.
Made by
Mariusz Szyma
I built Szept for myself. ChatGPT and Cursor understood my Polish perfectly, but only inside ChatGPT and Cursor; Claude Code, Gmail and Slack did not. I wanted that dictation in every app, light enough to keep open all day, with screenshots I can hand straight to an agent. By day I work in product, across crypto, regulated fintech and AI, and after hours I build my own AI apps.
I'm open to new roles, especially Head of Product.