macOS · Apple Silicon AI on-device Working MVP · pre-launch
A desktop video editor whose AI runs entirely on the creator’s own Mac: subtitles, voiceover, auto-editing and generation. No cloud, no accounts, no credits.
Packaged app v0.1.0 Full workflow tested offline Pre-revenue
Network log · 0 requests
Auto-captions, voiceover and silence cuts are now expected. In most editors they mean an upload, a sign-in and a credit balance.
In mainstream consumer editors, auto-captions, voiceover, translation and generation run on the vendor’s servers. They need a login, an upload and paid credits or a subscription.
Unreleased client work, interviews, medical or legal footage, minors and internal corporate video often cannot leave the device. Agencies, journalists, educators and companies are bound by NDAs and data rules.
Multi-gigabyte uploads, queues and outages hurt on weak or metered connections, and nothing works offline: travel, on set, field work.
Professional editors are complex. AI apps are single-purpose web tools that force round-trips between apps.
Popular consumer editors send speech, voice and generation jobs to their servers, behind sign-ins and credits. Kadr Studio does that work on the user’s own machine.
Media library, multitrack timeline, live preview, inspector and AI panel — in one window.
Every screenshot on this page is the real app. The interface ships in English and Russian: it follows the system language and switches in the app (top bar or View → Language) without a restart.
Thumbnails, duration, resolution, fps and size for every import.
Live preview with frame-accurate timecode; a 16:9 / 9:16 / 1:1 switch in the top bar.
Transform, speed, volume, fades, colour and transitions per clip.
Video, audio, text and subtitle tracks with snapping, markers and 200-step undo.
60 unedited seconds of the packaged app on an M4, in its English interface: playback, a split and undo, an image generated on-device, and Auto Reframe to 9:16 and back. One continuous take, no network.
Screen recording of the packaged app Apple M4, 16 GB all AI ran locally no audio or cursor
Video, audio, text and subtitle tracks with the tools editors expect. A 99-clip, 4-track project stays at 60 fps.
The playhead moves to 4 seconds, S splits the aerial coast clip in two, Q trims one second from the head of the left part, and the AI voiceover clip snaps to the playhead.
Split 13–15 ms on 99 clips60 fps playback200-step undoSnappingMute / lock / hide
Transcribe locally, get a caption track, fix any word, export SRT, VTT or TXT — or import an existing SRT.
As the playhead crosses each speech peak, the next word of the subtitle appears, ending with the full line: Everything runs on this Mac, even offline.
Transcript · auto-generated, editable
Language
Export
Font, weight, colour, background, outline, shadow and position — set once and applied to every subtitle on the track, then burned in on export. Titles get the same controls plus five presets.
Roadmap: word-level captions
Ten on-device features behind nine provider interfaces. Partial features are labelled — never hidden.
Transcription language in the app: Auto, Russian or English.
A speech waveform is converted into a row of transcribed words.
Transcribes locally with voice-activity filtering to suppress phantom phrases, then builds an editable caption track.
4–6× realtime
Animations are illustrative; the captures beside them are the real app timings: Apple M4, 16 GB RAM
The demo teaser in 16:9, reframed to 9:16 and to 1:1 by Auto Reframe — face-aware where it finds a face, centred where it doesn’t.
Real app captures, preview area demo project made with the app’s own AI
The Model Manager installs optional models only when the user asks, shows size, RAM, licence and commercial terms for each, and marks what is built in. Nothing downloads automatically. Listed download sizes are 0.2–3.6 GB per model — about 8.5 GB for every model this page describes.
“Used: 18.0 GB” in the capture is the test Mac’s disk: about 15 GB for this page’s models as installed, plus two Whisper variants the demo does not use.
Heavy sources play from lightweight proxies automatically. Exports always use the originals.
A 6-second 4K60 HDR original gets a 720p proxy with GPU tone-mapping in about 6.5 seconds on an Apple M4; editing plays the proxy at 60 fps; export renders from the original.
4K, HEVC, HDR and ProRes get 720p edit proxies HDR is tone-mapped to SDR (macOS) no HDR export
FIG 09 Real app the 20 s demo timeline exported in 3.8 s
Autosave within 1.5 seconds of every change, atomic writes and an automatic backup copy.
Six more full captures from the running app, each with English notes on what it shows.
FR 01/06
Network access is blocked at three layers, and every request is logged where the user can see it. The only traffic is a model download the user explicitly starts.
Illustrative · the enforcement layers, summarised
Illustrative summary: the interface, the main process and both AI workers are offline, the full workflow runs, and the network request count stays at zero.
Verified end to endA test runs the whole workflow with networking denied at the OS kernel level. The dialog on the left is what the app reported after generating the images, video, voiceover and subtitles used on this page.
Uploading footage and buying credits made sense when local models were weak. That is no longer true.
Whisper-class speech recognition, open multilingual text-to-speech (Chatterbox, MIT) and open image and video models now run on consumer hardware.
Every Mac sold since 2020 has Apple Silicon with unified memory, a GPU and a Neural Engine, and AI PCs ship with NPUs on Windows. On-device inference is a mass-market capability, not a lab demo.
Credit systems and rising subscriptions make “AI included, offline” a clear value proposition — and a local product carries no per-use inference cost.
Companies, agencies and public institutions restrict uploading media to third-party AI services.
Start with Mac creators who need captions, silence cuts and voiceover most; widen to professionals who cannot upload.
Video editing software, 2026
≈ $3.75 B
Growing to ≈ $4.99 B by 2031
Source: Mordor Intelligence · third-party estimate
Creator economy by 2027
≈ $480 B
~50 M creators worldwide
Source: Goldman Sachs, 2023 · third-party estimate
Market figures are third-party estimates. Kadr Studio has no market traction yet.
Where the AI runs decides what a creator has to give up: footage, an account, or a credit balance.
| Product | AI runs | Account / upload | Pricing model | Editor depth |
|---|---|---|---|---|
| CapCut desktop | Cloud | Required for AI | Subscription + credits | Consumer editor |
| Descript | Cloud | Required | Subscription | Transcript-first editor |
| Premiere Pro / Final Cut Pro / DaVinci Resolve | Mix; growing on-device features | Varies | Subscription or one-time | Professional editor |
| Single-purpose AI web tools | Cloud | Required | Credits | None |
| Kadr Studio | On-device | None | Planned: free core + Pro licence | Multitrack editor + AI suite |
Honest note: incumbents are adding on-device features, and Kadr Studio is early and smaller in features. The bet is a product that is local-first by design across the whole AI suite, not a cloud product with a few offline features.
Speech-to-text, voiceover, auto-editing, background removal and generation. No credits, no uploads, no account.
An in-app network log, and an automated test that runs the full workflow with networking blocked at the OS level.
Every AI feature sits behind a typed provider interface, so a better open model can replace an old one without UI changes.
Compute runs on the user’s hardware, so gross margin does not shrink with usage.
Every model and native component was audited for commercial use; non-commercial models were deliberately excluded.
A sandboxed Electron shell, a pure TypeScript editing core, an LGPL FFmpeg build and out-of-process AI workers — each layer replaceable.
FFmpeg, whisper.cpp and Python workers run as child processes. The UI stays at 60 fps during export, proxies and AI jobs.
Nine typed provider interfaces. Replacing a model means replacing a provider; preview and export share one frame model.
Auto Cut, highlights, reframing and silence cuts are unit-tested functions that propose a change.
A bundled, relocatable Python 3.12 with two isolated AI environments. The user installs nothing.
The benchmarks below come from recorded runs on an Apple M4 with 16 GB RAM (04.10.2026). Per-feature AI timings elsewhere on the page are stated measurements on the same machine.
Export · 3:06 timeline → 1080p
28.6s
6.5× realtime · 99 clips · 4 tracks
bar = render time vs. timeline length · Recorded run 04.10.2026
UI during playback, export, proxies
60fps
p95 frame time 18 ms
bar = 60 fps target · Recorded run 04.10.2026
Split on a 99-clip project
13ms
13–15 ms measured
bar = one frame at 60 fps (16.7 ms) · Recorded run 04.10.2026
Memory, all processes
<900MB
845 MB source · 889 MB installed app
bar = share of 16 GB · Recorded run 04.10.2026
Automated tests
63
48 unit/integration · 7 worker · 8 e2e
run 06.10.2026 · longest model test run on its own · Recorded run 04.10.2026
All green on 04.10.2026 — from source and against the installed app with every development dependency hidden.
19 issues found and fixed during verification and packaging (11 in QA, 8 in packaging).
| Partial + roadmap | Status | Note |
|---|---|---|
| Scene detection + highlights | Partial | Cut-mapping defect (quick fix); highlights are markers |
| Auto Cut proposals | Partial | Silence rules; scene rules wait on scene fix |
| Auto Reframe | Partial | Beta: one framing per clip |
| Audio clean-up | Partial | Denoise is DSP, not neural |
| Colour + speed | Partial | Basic: no curves/LUTs, constant speed |
| Video generation | Partial | Experimental: 512×288, 2–3 s, 16 GB |
| HDR sources | Partial | Tone-mapped to SDR, macOS |
| Voice cloning | Partial | Engine only, no UI |
| Windows | Partial | Platform layer prepared, not yet run on Windows |
| Keyframes + animations | Roadmap | Unlocks continuous reframe |
| Word-level captions + transcript editing | Roadmap | Whisper segments exist |
| Masks, chroma key, stabilization | Roadmap | Opportunity |
Compute runs on the user’s own hardware, so there is no inference bill to pass on — and none to absorb.
Planned · not live · no revenue yet
Free for distribution, with Kadr Pro as a paid licence (one-time or annual) that unlocks advanced AI tools and model packs. Price points will be validated in the beta.
For agencies and organisations that need on-device processing: volume licensing and offline deployment.
Revenue is mostly gross margin; the main costs are R&D and distribution.
Comparable anchors: Final Cut Pro and DaVinci Resolve Studio sell one-time licences; CapCut and Descript sell subscriptions.
Planned channels
Where the product stands today, and what comes next. The roadmap is a plan, not a promise.
Kadr Studio started as a test of how much of a modern AI video workflow can run fully offline on an ordinary Mac. The answer turned out to be most of it: captions, voiceover, auto-editing, background removal and generation, inside a real multitrack editor. Next: a full local AI video platform, where an agentic “edit for me” workflow runs over models on the creator’s own machine.
Built by a solo technical founder.
Designed and built end to end by the founder: the editor, the media pipeline, the AI-worker runtime, the packaging and the test suites.
Solo founder today. Funding would hire 1–2 engineers.
Straight answers, including the parts that are not finished.
They will add some of it. But their businesses are built on cloud services, accounts and credits, and making the whole AI suite local cuts against those revenue models. Kadr Studio is local-first by design.
By selling licences (Pro, Team) with no per-user inference cost, which gives software-like margins. Pricing is unvalidated today and will be tested during the beta.
For speech-to-text (Whisper large-v3) and voiceover (Chatterbox), yes: these are state-of-the-art open models. Image generation is solid. Local video generation is still experimental on 16 GB machines, and we say so. The provider architecture lets the product adopt better models as they appear.
The platform layer is written but has not been run yet. Windows is on the roadmap at 6–12 months, targeting AI PCs.
None yet. The product is built and tested but not publicly released. The first milestone after funding is a signed public beta.
A solo founder today. Funding would hire 1–2 engineers.
Every model and native component was audited, and non-commercial models are excluded. The OpenRAIL use restrictions for Stable Diffusion 1.5 and AnimateDiff will be passed to users in the EULA. FFmpeg is built without GPL or nonfree components.
No. There are no cloud AI APIs, accounts or telemetry. The network is used only for user-initiated model downloads, which are logged in the app. After models are installed the whole workflow runs offline, verified by an automated test with networking blocked at the OS level.
It is ad-hoc signed, not notarized: there is no Apple Developer ID yet. Signing and notarization come first on the roadmap, before the public beta.
A working, packaged macOS video editor whose AI runs entirely on the creator’s computer — now heading for a public beta.
To request the deck, reply through the channel this page was shared in.
