The interview is never
uploaded to be cut.
Drop in the full interview. It reads the whole thing, finds the moments worth cutting, and cuts them to 9:16 with captions on — all of it on your own machine. Nothing has to upload for any of that to happen. What does leave, and when, is listed further down this page by name — including the parts that leave automatically.
MP4 or MOV off your machine · the free scan runs on-device · no card
- Add the interviewIt stays connected — no re-adding it every session.Your machine
- Scan itFree on your own machine, or 2 credits for the deep read.Your machine
- Cut and dressTemplate, captions, bleeps, reframe. Preview is the export.Your machine
- Post itYouTube now or scheduled. Instagram and Facebook as the platforms clear us.Uploads the clip
- Rendered in your browser, not on a server
- The free scan runs on-device
- No subscription — credits never expire
- In the browser, transcription runs about realtime
Everyone else starts with “upload your video.”
That sentence is the whole product decision, and nobody writes about it.
Every clipper in this category begins the same way: hand over the whole hour, and we’ll tell you whether there was anything in it. If that hour is unreleased, unsigned, or has a name in it that hasn’t cleared, the upload is the decision. The clip was never the risk. The copy is.
This is not a claim that anyone is careless. It’s that upload gets treated as a law of physics here. It isn’t one. It’s an architecture choice, and it was made for a server bill.
Where each step runs
The whole list, in order. The rows that say anything other than “your machine” are the ones worth reading twice.
| Step | Runs on |
|---|---|
| Transcribe | Your machine, by default — the speech model ships inside the app and is served from the app itself. In a browser that runs at roughly realtime, so an hour of footage takes about an hour; there is a Mac build that does it on the GPU far faster, and it is not something you can download today. There is one exception to “your machine”, and it is in the list below: the background transcribe queue sends the audio to a transcription provider, and it spends credits. You choose that path; nothing takes it for you. |
| Find clips — free scan | Your machine. No model provider is called. Works with no connection. |
| Reframe to 9:16 | Your machine. The face model is a file in the app, not a service. |
| Captions, templates, bleeps, trims | Your machine — except the AI caption highlights, on by default, which send the clip’s caption lines to the model provider. |
| Render and export | Your browser. Canvas and WebCodecs. There is no server render path in this product. |
| AI scan — 2 credits, optional | Sends the transcript text to the model provider — and a reference image, if you attach one to a brand look. The scan never sends the video. |
| Post to IG, FB, YouTube or TikTok | Uploads the finished clip. That is what posting is. |
Scroll the table sideways for the rest
What leaves this machine
Open the Network tab while it works on an hour of footage. You will watch the project text — transcript included — sync to your account. You will not watch the hour upload, because nothing here needs it to.
Below is the complete list, automatic rows included. If it grows, this list changes.
| What | When | Where it goes |
|---|---|---|
| Your project — transcript, caption lines, the clip list with its hooks and reasons, guest names, quote text, social copy, your logo | Automatically, about a second and a half after any change, while you’re signed in | Your row in our database, so the project reopens on your other machine |
| Source files under 50 MB | Automatically, when you add one | A private bucket under your user id. An interview is far over the cap, so it stays put. A short clip is not. |
| Exported clips under 50 MB | Automatically, after a render | The same private bucket |
| A list of your library for your phone — each interview’s name and length, the clips found in it with their scores, titles and descriptions, what is already scheduled, one small poster picture per interview, and the requests you make from your phone | Only after you switch Phone link on in the Mac app, and only while that app is open and signed in | Your own folder in the private bucket. No video, no transcript. |
| A clip’s picture and sound, live as you watch it | Only while you press Watch live on your own phone, and only while it plays | Straight from your Mac to your phone, encrypted. The live stream is never stored, and does not pass through our servers. On a network that blocks a direct connection it goes through Cloudflare’s relay instead, still encrypted between your Mac and your phone, so the relay can’t see it either. If it still can’t connect, your phone asks for the preview in the next row. |
| A short private preview of one clip | Only when you press Watch on your phone and a live stream can’t connect — the one clip you chose, never the interview | Your own folder in the private bucket, opened by your phone with a link that expires within minutes. If your phone can’t connect directly, your Mac uploads a short private preview of just that clip, deleted within an hour. |
| A reference image you attach | Only when you generate a brand look from a screenshot | The model provider, with your description. It is never stored here. |
| A frame of the interview | When you make an AI-art thumbnail — by hand, or in a batch upload whose thumbnail style is AI art | OpenAI’s image model, with the video’s title, to redraw the guest into the story. It is never stored here. |
| The finished clip | When you post or schedule | The platform you post to. A short clip goes through a private bucket, and our server hands it on to Instagram, Facebook, YouTube or TikTok — or the Mac app does, for a TikTok it posts for you. A long YouTube cut uploads straight from your browser. |
| The transcript text | When you spend an AI scan, and whenever the AI works from it — writing titles, hooks and post copy, picking caption highlights, and describing the scene for a thumbnail picture it draws | The model provider. For a thumbnail picture, the short scene it describes then goes to OpenAI’s image model with the title |
| The audio | Only on the background transcribe queue, which spends credits | The transcription provider |
| Fifteen numbers and a yes/no per sentence | On an AI scan, to sharpen the on-device finder | Our pool. No words, no text, no media, and your address is hashed before it’s stored. |
Scroll the table sideways for the rest
“If footage sovereignty is the pitch, why does my transcript upload automatically with nothing to switch off? The words are the sensitive part. Nobody leaks an interview by leaking its file size.”
Because the claim is about the asset, not about zero egress, and we should say which. The ten-gigabyte file never moves — it is cut where it sits, by your own CPU, and no server has a copy to leak, subpoena, or train on. The derived text does move: signing in syncs your project so it reopens on your other machine, and the transcript is part of that project. It lands in a private row under your user id. Today there is no toggle for it. If you need zero text egress, this tool is not there yet, and no sentence on this page should let you believe otherwise.
Signed out, none of it moves. Read that list against any other clipper’s front page — the comparison is the argument.
It tells you when it came up short
“COMEDIC: this interview only carries 42s of usable beats — the cut will run short of 60s.”
Ask for a 60-second trailer from an interview that can’t carry one, and that is what you get — before it renders, with the real number. On the cut itself: “This cut runs 0:42, short of the 1:00 target — that is the real length, not a rounding.”
Two different things produce a short cut — a thin interview, or a mood that structurally tops out — and it tells you which one you have, because blaming the wrong one sends you hunting for a problem you don’t have.
It does the same with money. Take the free local pick and the chip reads Local pick — no AI credit. Pay for the deep read and get one mood filled in locally and it reads AI pick · 1 credit — this mood filled locally. It will not tell you it spent your money when it didn’t.
And there is a score — a 1–100 HEAT number on every clip, with the library sorted hottest first. It is a judgement about how strong a moment is, and the AI one is graded against what tends to travel on a channel like this. So it is a guess. Every other tool in this category stops there.
This one shows its work. Beside the number are its four axes — HOOK, FLOW, ENGAGE, TREND — the sentence that earned it, and the timecode. Open the moment and judge it yourself in ten seconds. A score you can check is worth more than a score you have to believe.
Everything the edit needs
From the raw file to the posted clip, in one place.
You buy credits. Not access.
There is no plan here. There is a balance. It goes down when you spend credits on AI, and it does nothing at all in between.
A credit is $1.00 in the $5 pack, $0.91 in the $10. The caption highlights share one charge with the rest of the AI writing — titles, hooks, post copy, a brand look: 1 credit for each 20-minute stretch of the clock in which any of it runs, so an hour of steady use is 3 or 4 credits. A trailer’s deep read is 1 credit. An AI-drawn thumbnail picture is 2 credits — 3 when its scene brief is the first AI writing in its 20-minute stretch. On-device mode costs zero.
When you buy credits it is a one-time Stripe checkout — your card never touches this app · no subscription · nothing to cancel · credits do not expire
30 AI scans a month is 60 credits — $55 in packs here. If you scan that much, price a flat monthly plan against it, and if it’s cheaper, go buy one.
Where this is the wrong tool
Three things you’d find out anyway.
Start with one interview.
3 free credits come with the account — enough for 1 AI scan. The on-device finder never runs out, so the tool keeps working after they’re gone.
Create an account