Launching Snapbak: capture, dictation, and redaction that stay on your Mac

Today Snapbak 1.0 goes out the door. It’s one macOS app for the capture work that used to take four or five: screenshots, screen recording, GIF export, voice dictation, AI notes, and bug reports. It installs at 10 MB, idles at 25 MB of RAM, and does everything locally unless you explicitly opt into a cloud feature.

If you use a recorder, a screenshot tool, and a dictation app today, you’re likely carrying over two gigabytes of apps whose job is to sit in the menu bar and wait for a hotkey — and the popular ones upload your captures to a server by default. We wrote about where that weight comes from; this post is about what shipping against it looks like.

What’s in 1.0

• Screenshots — capture to clipboard in about 150 ms (300 ms cap for a full‑screen grab), scrolling capture for long pages, OCR text grab, annotation with numbered callouts, and Beautify framing for public posts.
• Smart Redact — the screenshot is scanned on the device, before it goes anywhere: emails, card numbers, common API‑key formats, and faces get marked and blurred in one click. No network calls in the detection path — the sensitive image never uploads to be scanned.
• Screen recording — auto‑zoom that follows the cursor, cuts that land on word boundaries so the tracks stay in sync, filler removal with a review list first, and captions burned in or exported as SRT/VTT.
• GIF export — recordings to loops sized to fit under chat and GitHub limits.
• Offline dictation — English, on the Neural Engine, with a personal dictionary (“sounds like” hints) so names and jargon come out right. Works in any app; here’s how the models compare.
• AI notes — titles, summaries, and action items from recordings, in about half a minute for a ten‑minute capture. On‑device where the OS offers it, or your own API key (OpenAI, Anthropic, Google, Groq). A per‑capture switch turns it off entirely.
• Bug reports — repro steps assembled from the captured events, not guessed: sensitive headers redacted, password fields never captured, export as Markdown for GitHub, Linear, or Jira, or one offline HTML report.

The numbers, with the conditions

Everything below is a measurement, recorded with the machine it was taken on; the full table with methodology lives on the benchmarks page.

One honest footnote: dictation ties Whisper large‑v3‑turbo on accuracy but not on latency — 0.045 s to transcribe 3 seconds of speech, versus 0.68 s for Whisper. That gap, not accuracy, is why Snapbak ships Parakeet, and why key‑up‑to‑text is under half a second.

How it stays small

Four choices do most of the work:

1. Tauri v2 for the shell. The UI runs in macOS’s system WebView; no bundled Chromium. The web layer never touches a video frame or an audio sample.
2. Swift for the hot paths. Capture goes through ScreenCaptureKit, export through the hardware video encoder, OCR and dictation through Apple’s ML stacks. Nothing compresses on your CPU.
3. Models download on demand. The speech model is about a 540 MB first‑run fetch. If you never dictate, you never download it — the installer stays 10 MB because it doesn’t contain your ML workload until you ask for one.
4. The dictation helper quits. Transcription runs in a separate process that loads the model, does the work, and exits. The model weights are the memory, and they leave with it — which is why the app itself idles at 25 MB.

The local‑first line

Your screen, your voice, and your customers’ data are the most sensitive material a work machine holds, so the default is that none of it leaves. Redaction detection makes no network calls; dictation makes none; recordings stay as files on your disk and export when you say so. Where cloud features exist at all (Cloud AI, coming soon), the design rule is transcript‑only: text goes up, frames and voice never do — and bringing your own API key means the vendor in the middle is the one you picked. The redaction workflow this enables is the feature I’d point a new user at first.

What’s not in it yet

• Meeting minutes. The design is done — bot‑free capture from the meeting apps’ audio, a local transcript, minutes assembled after — but it isn’t in 1.0. Planned for a later release.
• Intel Macs. Apple Silicon and macOS 15+ only for now.
• A cold‑start number. We haven’t published one because the bench conditions aren’t pinned yet, and a number without its conditions is marketing. It goes on the benchmarks page when it’s measured — that page also lists what else we haven’t measured.

Pricing

Snapbak is free for the first 100 users. After that, Pro is planned at $69 one‑time — the core app is not a subscription — with optional Pro Cloud at $8/mo and Team at $12/user/mo.

That’s the launch. The numbers above are reproducible from the stated conditions, and the parts you care about are worth measuring yourself — download Snapbak and check us against the suite you’re running now.