home  /  projects  /  magicnotesurface
live

magicnotesurface

Magicnote window in use: folder tree of a Vietnam trip on the left, a Mũi Né card in the middle with structured travel notes, and the assistant chat at the bottom showing the model thinking through the request before editing the card.

Magicnote is a notebook that listens, watches, and helps you think. You write in it the way you'd write in any notebook — pages inside folders — but it can also record audio, take photos, accept drawings, and talk to an assistant that reads and edits alongside you. This page walks through what it's like to use.

a look at the assistant working

It thinks out loud, then edits the page

The chat pane isn't a chatbot off to the side. It reads the card you're on, runs whatever capabilities it needs, and writes the result back into the page. A few snapshots of that happening.

Mũi Né card with the assistant chat below it. The user asks for an expanded section about boat tours. The assistant first replies 'I need to check what card is currently active and what destination it covers, so I can research and add accurate boat/island tour information for that specific location,' then 'Now I'll add an expanded section on boat/island tour information,' and a structured bullet list lands on the card.

Reads the card, then writes

You ask for an expanded section on boat tours. The assistant checks which card you're on, figures out what it's about, and writes the new bullets straight onto the page. Each step is visible in the thread, so when something goes sideways you can see where.

Assistant chat where the user asks to teach the Viet Vacation > Places folder a default card shape. The assistant proposes a New rule: cards start with a Location section (with map context or coordinates) before the Overview, then asks if it should make a test card for Tra Cu Vietnam.

Teaches the folder a card shape

You can also point the assistant at a folder and say "every card in here should start with a map section, then an overview, then the rest." It proposes the structure, you say go, and from then on every new card in that folder begins life with that skeleton.

new things since the last write-up

A few releases in

Live transcription dialog. Audio source can be the mic or the output stream of a running app. Save audio recording is on. Engine dropdown shows Whisper (offline) selected and Deepgram nova-3 as the alternative. Translate live to is off, with a Claude/Marian backend toggle.

Live transcription, your choice of engine

Hit LIVE and pick what you want. Whisper runs offline; Deepgram nova-3 runs in the cloud and is faster on long sessions. Same dialog lets you pick the audio source (mic, or the output of any running app on the laptop) and turn on per-paragraph translation with Claude or offline Marian as the backend.

Recover window listing soft-deleted items: a card under Places, a drawing artifact, a demo card with several artifacts (drawing, .ogg recording, transcript.txt, photo, video, map JSON, map PNG), more cards. Each row has a Restore button.

Delete + Recover

Cards, folders, artifacts, individual blocks — all soft-deleted. The Recover window shows what's gone and puts it back exactly where it came from. Mistakes aren't permanent, including the ones the assistant makes on your behalf.

Right-click context menu on a folder in the sidebar: New card here, New subfolder, Rename folder, Set transcribe defaults..., Export website..., Delete folder.

Export a folder as a site

Right-click any folder and pick Export website… You get a self-contained zip — clean HTML, all the assets, a manifest.json, and a README. It's both a browser-printable archive of your work and a handoff bundle the assistant (or any other LLM tool) can read end-to-end. The same menu is where folder-wide transcription defaults live.

Settings General tab. Microphone gain section with Mic Boost (0–3, 10 dB per step), Capture Volume (0–63, ~0.75 dB per step) and a combined dB readout. API keys section with masked fields for Anthropic, Deepgram, and Hugging Face. Translation defaults at the bottom (en → vi).

Settings grew a bit

Microphone gain on the Surface is two knobs you have to set together; magicnote shows you the combined dB so you stop guessing. API keys for Anthropic, Deepgram, and Hugging Face live in one place and are saved to a mode-0600 file. Translation source and target language sit alongside.

Settings Transcription tab: Defaults header noting that per-folder overrides win for cards inside that folder, with Model, Speaker mode, Language, and Initial prompt fields.

Per-folder transcription defaults

App-wide defaults for the transcribe action live here. But if you set different defaults on a folder — say, Vietnamese with a glossary prompt for your trip folder — those win for any card inside it. The right setting follows the work, not a global toggle.

Backup, three ways

One-click Back up notebook… from the hamburger menu writes a tarball using SQLite's online-backup API, so it's consistent even if you're mid-edit. A magicnotesurface-backup CLI does the same thing headlessly. A magicnotesurface-backup-s3 CLI pushes the same byte-identical bundle to MinIO, S3, B2, R2, or Wasabi and re-verifies it by SHA-256.

what it does for you

In plain language

Capture anything, in context

Write a note. Record the lecture you're sitting in. Photograph the whiteboard. Scribble on that photo with the stylus. Ask it to transcribe. Ask it to translate. Everything stays on the same page, in the order it happened.

An assistant that actually reads your notes

You're not chatting with a stranger. The assistant sees the page you're on, the folder you're in, and what you've been working on recently. "Summarise this" works. "What did I write about Kyoto last week?" works.

Yours, on your laptop

Your notes, recordings, photos, and drawings live on your own machine. No cloud sync you didn't turn on. You supply your own Anthropic key to talk to the assistant; if you don't, the notebook still captures, searches, and edits. It just stops being a chatty one.

Works offline when it matters

Transcription runs locally. Translation to English or Vietnamese has an offline fallback if the internet's gone. The parts that need the cloud say so. The parts that don't, don't.

first look

The window, from left to right

When you open magicnote, the window splits into three columns. On the left, a sidebar lists your folders and cards — this is the tree you browse. In the middle, the card you're currently on. On the right, the chat pane where you talk to the assistant.

If you rotate the Surface into portrait, the layout gracefully flips. The sidebar collapses into an overlay you can pull out, and the chat pane docks to the bottom.

Your first card

Cards are pages. Click the "new card" button in the sidebar. Give it a title. Start typing. Autosave kicks in after a second of idle — you never have to hit save.

A card can hold more than text. It can hold photos, voice memos, videos, drawings, maps, transcripts, translations, and the assistant's replies — all as blocks on the same page, in the order you added them. You drag to reorder. You right-click to operate on any block.

Writing, selection, formatting

Type like you'd type anywhere. Selecting text gives you a small popover with bold / italic / code / bullet list / numbered list. Headings are under the same popover. Nothing exotic — the goal is that if you've used any modern editor, you already know how.

Capturing what's happening

The toolbar at the bottom of the card has four capture buttons: camera, voice, video, and app audio. Each of them attaches an artifact to the card you're on, in the position you're looking at. Let's walk through what each one feels like.

Taking a photo

Tap the camera button. A viewfinder appears. Tap the shutter (or hit Space). Brief flash, and the photo lands on the card. If the photo's EXIF data has GPS coordinates and the card doesn't have a location yet, magicnote auto-stamps the card with where you were.

There's a small switcher for front/back camera. Autofocus happens before the shutter closes — you won't get a soft photo because you tapped fast.

Recording a voice memo

Tap the mic button. A slim recording bar slides in under the header with a timer and two choices: Stop & Save or Discard. You can keep typing. You can switch cards. You can take a photo. The recording keeps going.

Stop & Save writes the audio to the card you started recording on (even if you've navigated away). Discard just throws it out.

Recording a video

Tap the video button. A small preview window appears — it's non-modal, so it floats while you do other things. There's a 2.5 second warmup before recording starts, so the auto-exposure and auto-focus can settle. No blown-out first frame.

Capturing app audio

This one's a superpower. Tap the app-audio button and a dialog lists every application on your laptop that's making sound — Firefox playing a lecture, Zoom, mpv, anything. Pick one. Magicnote records its output cleanly and attaches the file to your card.

This means you can listen to a seminar, capture the slides' audio, and keep typing your own notes alongside — all on the same page. No virtual audio devices, no "stereo mix" hacks. It just works.

Drawing and annotating

Pick up your Surface Slim Pen and tap the pencil icon on any photo, or start a blank drawing from the toolbar. You get a palette of ten colours, a brush size slider, pen and eraser toggles, and undo/redo (Ctrl+Z / Ctrl+Y). The pen is pressure-sensitive.

Drawings are re-editable in place. Tap the pencil again later and you can keep refining — lines aren't flattened into the image until you want them to be.

Maps with annotation

Open the map dialog from the toolbar. Pan and zoom through a real world map (via OpenStreetMap tiles). Flip into drawing mode to pin a route, mark a destination, or annotate. The map preview lives on your card as a block; tap the pencil to edit it later.

Asking the assistant

Magicnote card on the left, chat pane on the right: the assistant is mid-stream answering a question about the card's contents with a tool-call trace visible.

The chat pane on the right is attached to the card you're on. Anything you type there is a question about this card, with full access to the card's contents. Some things to try:

The assistant streams its reply into the card as a block. You can edit it afterwards the same way you'd edit text you typed yourself.

Transcribing an audio block

Right-click any audio block and pick one of: Quick (fast, good enough), Accurate, Best (slower, better), or Plain text. The transcript appears as a new block right below the audio.

Under Custom… you can choose the model, turn on speaker separation (useful for a recorded conversation), set the language if the auto-detect guesses wrong, and optionally seed the transcriber with an initial prompt ("lecture on Vietnamese grammar, expect technical terms").

Live transcription

Magicnote card mid-LIVE session with the LIVE button highlighted, partial recognised text in a muted weight below committed paragraphs.

Press the LIVE button on the card footer and start speaking. Your words appear as they're recognised. Paragraphs form automatically when you pause or another speaker starts. When you're done, save the transcript to the card.

There's an optional "Translate live to:" dropdown. Turn it on and each finalised paragraph emits a translation alongside it. Useful if you're sitting in a talk in a language you're still learning.

Translating an existing block

Right-click any text, transcript, LLM reply, or summary block and pick Translate ▸. You'll see four options: English (Claude), Vietnamese (Claude), English offline (Marian), Vietnamese offline (Marian).

The Claude options give you the best quality. The offline Marian options work with no internet and no API key at all — useful when you're on a plane, on a train, or in a café with flaky WiFi. If Claude isn't reachable, magicnote will fall back to Marian on its own and tell you why.

Finding what you've written

There's a search box above the folder tree. Type a word and the sidebar swaps into a flat list of matching cards. It searches titles and contents (including derived transcripts). Hit Escape to drop back to the folder view.

Above the folders there's also a media tree — every audio file, every photo, every drawing, every video, every map, all grouped by kind. Click one and you jump straight to the card it lives on, with the block highlighted.

Organising with folders

Right-click any folder in the sidebar for: new subfolder, rename, delete (its cards reparent automatically), set default spell, set transcription defaults. Folders nest as deep as you want.

Spells — teaching it how you like things

Spells are rules you can attach to a card, a folder, or globally. Click ✦ Spells in the topbar to open the library. Examples:

Spells follow whatever you're working on. If two attached spells disagree, the assistant reconciles them or asks you, depending on the Spell conflict mode setting.

Settings

Magicnote settings modal: API key field, spell conflict mode dropdown, transcription defaults, mic gain slider, and an editable 'what the assistant remembers about you' block.

The ⚙ gear in the topbar gives you:

Things to try

The notebook is yours

Every recording, every photo, every drawing, every assistant reply lives in your notebook on your machine. If you want to move it to a different laptop, you copy one folder. If you want to back it up, you back up one folder. If you want to stop using it tomorrow, your notes are still plain files you can open.