On-device AI: why your footage should never leave your machine

Here’s the deal with most “AI video tools”: you record something, you hit upload, and your footage — raw, unedited, possibly containing your face, your screen, your voice, your proprietary product demo — travels to someone else’s server. There, a model transcribes it, cleans it up, and sends the result back. Convenient! Also: a privacy nightmare if you think about it for more than ten seconds.

Reframe does all of that locally. Transcription, voice cleanup, filler-word detection, zoom suggestions — all on your machine. Your footage never leaves your laptop.

Why we picked local

Three reasons, in order of how much we care about them:

1. Privacy is the feature

If you’re recording an unlisted product demo for an investor pitch, you don’t want it on a third-party server. If you’re recording a course module that’s part of a paid product, you don’t want it in someone else’s training set. If you’re recording a sensitive internal update, you really don’t want it anywhere but your machine.

“It’s private, trust us” isn’t good enough. “It never leaves your machine” is.

2. Speed

Cloud AI has a built-in bottleneck: upload. A ten-minute 4K screen recording is a big file. Uploading it takes time. The model takes time. Downloading the result takes time. Total round-trip: minutes, sometimes tens of minutes.

Local AI starts the instant you hit the button. Transcription of a ten-minute video finishes before you’ve finished your coffee. There’s no upload bar because there’s no upload.

3. It works on a plane

This sounds glib but it’s real. If your tool only works with an internet connection, it doesn’t work in a lot of places people actually want to make videos — planes, trains, coffee shops with bad Wi-Fi, the cabin your parents dragged you to for the weekend. Local-first means your tool works where you work.

The trade-offs (because there are some)

We’re not going to pretend local is free. It costs:

• Hardware requirements. Local models need RAM and GPU. Reframe runs great on Apple Silicon Macs and modern Windows machines. It will not run great on a 2014 MacBook Air.
• Model size. We ship smaller, more focused models instead of one giant one. That means transcription is excellent but it’s not GPT-4-grade reasoning over your script.
• No collaborative cloud features. You can’t share a project link with a teammate and have them edit in the cloud. We’re working on a way to do this without uploading your footage, but it’s hard.

Why we think it’s worth it

Because the alternative is a world where every video you make passes through someone else’s servers by default. Where the cost of using AI tools is giving up control of your raw footage. Where “the cloud” is just a euphemism for “a server you don’t control, run by a company that may or may not exist in three years.”

Local-first AI isn’t the easy choice. It’s the right one. And every year, as hardware gets faster and models get smaller, it gets easier too.

What this means for you, practically

When you hit “transcribe” in Reframe, the work happens on your CPU and GPU. When you click “remove filler words,” the detection runs locally. When you enable de-noise, the audio processing happens on your machine. Your raw recordings, your transcripts, your cleaned-up audio — they live in your project folder, on your disk, under your control.

You can copy that folder to an external drive. You can delete it. You can back it up. You can’t accidentally leak it to a third-party API because there is no third-party API.

That’s the whole pitch. It’s not flashy. It’s just how it should work.