転生 SnapScribe
Open source / Windows / Python + Rust

Snap Scribe

A note taker that runs in the background. Press one key and it listens, watches the screen, and writes the notes for you.

Get the source CtrlAltR to start listening
SnapScribe—▢✕
Listening
00:12:34
You
Them
Live transcript
On screen ● saved

A pricing page draft: three plans, the middle one highlighted, with a monthly or yearly toggle.

A sample meeting, transcribed live Screenshot every 10 s, if the screen changed
How it works

One key.
Four jobs.

It never records video unless you ask. Waiting in the tray it uses about 40 MB of memory and no processor to speak of.

transcript.md2 channels · nova-3

00:02:11You Can we push the launch to Thursday? The pricing page isn't done.

00:02:16Them Thursday works. Who's writing the release notes?

00:02:20You I'll take it. Mark this.

00:02:20★ marked · screenshot saved

00:02:25Them Great. Let's also cut the yearly discount to fifteen percent.

Ch.01

It listens to both sides

Your mic and your computer's audio, transcribed live as people talk, labelled You and Them. Mixed Hindi and English works too.

  • Live transcript
  • You / Them labels
  • Voice cues: "take a screenshot", "mark this"
  • Live subtitles, never captured
screenshots/every 10 s · if 2% changed
00:00:10
skipped · same
00:00:30
skipped · same
★ 00:02:20

AI · A pricing page draft with three plans; the middle one, "Team", is highlighted and a monthly / yearly toggle sits above.

Ch.02

It watches the screen, quietly

A screenshot every ten seconds, but only when the screen has actually changed. Each one gets a short AI description, so you can ask about it later.

  • Nothing saved if nothing changed
  • Ctrl Alt M marks a moment
  • Video only when you press Ctrl Alt V
  • Its own window is hidden from capture
notes.md14:03 · 38 min

Launch moved to Thursday

TL;DR: the launch slips two days for the pricing page; the yearly discount drops to 15%.

Decisions
  • Launch on Thursday, not Tuesday
  • Yearly discount 15%, down from 20%
Action items
  • You · release notes by Wednesday
  • Them · finish the pricing page
★ 00:02:20 release notes▶ 00:14:05 discount▶ 00:31:40 wrap-up
Ch.03

You stop. The notes write themselves.

Title, TL;DR, key points, decisions, action items, your starred moments and a timeline. A notification tells you when they're ready.

  • notes.md + record.html
  • Click any time to jump
  • Crash-safe: finished on next start
  • Rebuild with another model any time
AskCtrl Alt A
What did we decide about the yearly discount?
You agreed to drop it from 20% to 15%, and the launch moved to Thursday so the pricing page could catch up. ▶ Mon 14:03 · 00:02:25▶ Mon 14:03 · 00:14:05
Ask about anything you've recorded…↵
Ch.04

Ask it anything, later

A chat over every session you've recorded: notes, transcripts and what was on screen. Answers stream in and cite the exact moments as chips.

  • Click ▶ to open that second
  • Reads a session still running
  • Searches when history gets big
  • Also in the terminal
Hotkeys

Seven keys.
No window needed.

CtrlAltR

Start / stop listening

Live transcript, auto-screenshots, and notes at the end. The tray icon gets a red dot while it listens.

CtrlAltS

Screenshot now

CtrlAltM

★ Mark this moment

CtrlAltA

Ask your notes

CtrlAltT

Live subtitles

CtrlAltV

Screen video on / off

CtrlAltQ

Quit, after finishing

or just say it

"Take a screenshot" · "Mark this" · "Remember this"

Per session

Plain files.
Yours.

sessions/
  index.html          ← every session, searchable
  2026-11-08_14-03-22/
    record.html       ← notes + timeline + screenshots + audio
    notes.md          ← title, TL;DR, decisions, action items
    transcript.md     ← "You:" / "Them:" with timestamps
    audio.ogg         ← ~14 MB an hour
    screenshots/      ← each one described by AI
    timeline.jsonl    ← crash-safe event log
    video_01.mp4      ← only if you pressed Ctrl+Alt+V
Built light

Heavy only
while open.

40 MB

waiting in the tray

4 MB

the native window

Engine
Python: hotkeys, tray, audio, Deepgram, screenshots, notes, Ask. Audio libraries load only when you start listening.
Window
Rust, on tao and wry (what Tauri is built on), in WebView2. Closing it frees all of it; you're back in under half a second.
UI
Plain HTML, CSS and JS, no framework, no build step. Windows 11's own fonts and icons, live over server-sent events.
Setup

Three
commands.
Two keys.

You bring a Deepgram key for transcription and an OpenRouter key for the notes, descriptions and Ask. Settings live in config.toml; editing them needs no rebuild.

  1. 一

    Install

    pip install -r requirements.txt pyinstaller
  2. 二

    Add your keys

    copy .env.example .env   # DEEPGRAM_API_KEY, OPENROUTER_API_KEY
  3. 三

    Build

    build.bat   # the window + SnapScribe.exe + shortcuts

    Needs ffmpeg on your PATH, and Rust for the native window. Without Rust it opens in an Edge app window instead.

What it costs

A few cents an hour. Deepgram nova-3 bills per channel (set speakers = "mixed" for half). The AI is about $0.0003 a screenshot and $0.0005 a question.

Where your data goes

Sessions stay in a folder on your PC. Audio streams to Deepgram to be transcribed; text and screenshots go to the model you pick on OpenRouter. Nothing else leaves.