Chat with an AI about your handwritten notes and documents — right on your Supernote, privacy first.
SmartNote AI is a plugin for Supernote e-ink tablets (Manta, Nomad…). It puts a Mistral AI assistant on your device: ask about the page you are writing, lasso a corner of it for a quick question, or build a searchable library of your own notebooks. Everything runs with your own Mistral API key, and your transcripts never leave the tablet.
SmartNote AI is a personal project built by a Supernote user, for Supernote users. It is not an official product of Supernote or Mistral AI, just a plugin that loves them both.
Pick either — or both. Nothing here is mandatory.
- Just ask. Open the floating assistant on any note and ask the AI about the current page — or lasso part of the page and ask about that. Nothing has to be transcribed or stored first; it just works.
- Build a library (READ). Optionally transcribe your notes and PDFs into
a local library on the tablet. This is never required — it simply opens up
more: powerful SEARCH across everything you ever wrote, AI AGENTS
that always know a chosen set of documents, and EXPORT to
.md/.txt.
Transcribing is an opportunity, not an obligation.
The Open floating assistant button opens a movable, resizable panel that floats over your note — you keep writing on the page while you talk to the AI about it.
- Brain dropdown (the name + ▾ in the header): switch between the built-in CHAT and any of your agents. It shows the model, how many documents and pages are loaded, a New chat action and your History.
- "Chat context:" tags just under the header: the current page (or a range, or the whole note), extra pages you added from the Library or Search, and any lasso images. The page tag follows you as you turn pages.
- Quick actions write a ready-made prompt into the input field (summarize, translate, to-do list…); you send it yourself with the Mistral M button.
- One-shot Web search: arm Web and it applies to your next message only, with cited sources.
- Answer style: precise / balanced / creative.
- History: every chat is saved locally, resumable and deletable.
Follow-up questions cost a fraction of the first, thanks to Mistral's prompt caching.
Lassoing part of a note attaches it to the chat as an image. It is not a separate tool — it rides on whatever chat or agent is active.
- The image becomes a persistent "🖼 Lasso" chip in the Chat context row; add several if you like, remove any with ✕.
- An editable "Lasso — image reading" directive (in 2 · READ) tells the AI there is an image to read and act on. Leave it empty to send no instruction.
- Up to three lasso quick actions (default: About this selection, in 3 · CHAT & AGENTS) appear ahead of the normal ones, but only while an image is in the context.
READ is the engine underneath SEARCH, AGENTS and EXPORT. Handwritten pages are read by Mistral OCR 4 and then by a Vision model (Ministral); the OCR text rides along as a hint and its word-confidence feeds a glossary of your own vocabulary — your biggest accuracy lever. Printed PDFs are read by OCR alone (strongest and cheapest on print); Vision is applied only where it earns its keep — a schema or figure the OCR misses, and any page carrying handwritten annotations.
Every file or folder has a sync mode, set with one tap on its chip in the Library:
- Off — never sent to the AI. The privacy switch.
- Manual — read only when you press a Sync button (or ask the chat about it). Since v0.88 Sync now is a standing order: if the sync is interrupted (big backlog, offline, wrong app open) it finishes by itself on the next passes, Vision included — no need to tap again.
- Auto — read in the background, page by page, while the plugin is open.
The sync adapts to what is on screen by itself: note pages are read while a note can be rendered, PDF Vision runs while a PDF reader is available, and the SYNC STATUS panel shows which one is active ("Rendering via: …"). No page is ever left with OCR text but no Vision — the missing Vision passes run automatically as soon as the right app is open.
The plugin detects exactly which pages changed, so an unchanged page is never paid for twice. Setting a chip costs nothing — you only pay when pages are actually read.
The handwritten ink you add on top of a PDF (the Supernote .mark layer) is now
read too. The plugin notices changes to the .mark layer, so annotated pages
are picked up automatically at the next sync — and you can force a Vision read
yourself:
- 🔍 Read with Vision (schemas & annotations) on a single PDF page.
- 🔍 Read all with Vision on the whole document — or select pages first and read just those.
Once a PDF has been read, editing or adding an annotation later re-reads only that page, with Vision — the rest of the document is never re-charged.
The Library is your local transcript store and the cockpit for syncing.
- A tree of your Note and Document folders (SD card included), each row showing its mode chip, page count and how many pages are still to read.
- A SYNCHRONISATION panel with side-by-side AUTO and MANUAL columns, an always-visible SYNC STATUS progress bar, and a Keep Supernote awake during Sync toggle so a big backlog can drain without the screen sleeping.
- Whole-library actions: Export all (
.md/.txt), Back up / Restore the library, Clear all transcripts, and Clear transcripts of Off notes. - Open a document to see its pages; open a page to see the transcript beside the original page image and the Supernote live OCR. Correct anything by hand — a hand-fixed page is marked manual and never overwritten by a later sync — or tap an underlined "unsure" word to fix just that one. Clear transcript wipes a single document.
- Each row also shows a badge for every agent that knows it; tap it to remove it from that agent.
An agent is a custom assistant with its own name, icon, persona, model, answer style, quick actions — and a set of Library documents it always keeps in mind.
Examples: a Meeting notes agent over your work notebooks; a Recipes agent over your kitchen notebook; a Thesis agent that knows a long PDF and quotes it.
- Create up to 8 custom agents in 3 · CHAT & AGENTS (plus the built-in CHAT).
- Four ready-made starter presets (Extractor, Writer, Tutor, Brainstorm) can be added in one tap. A fresh install ships with no agents.
- Build an agent's context from the Library ("+ Add to ▾ → this agent"); the screen shows how many pages are read and what reading the rest would cost.
- An agent's documents ride in a stable order, so Mistral's prompt cache keeps follow-up costs low. If some of its pages aren't read yet, the plugin offers to read them first — with the price — or to chat with what's already there.
- Download the latest
smartnoteai-x.y.z.snplg(or build it:bash buildPlugin.sh, with Node ≥ 18, JDK ≥ 19, Android SDK 35). - Copy it to
MyStyle/on the Supernote (USB oradb push). - On the device: Settings → Apps → Plugins → Add Plugin.
- Open the plugin from the notes toolbar and paste your Mistral API key in 1 · API key, privacy, backup & appearance.
SmartNote AI has no middleman server and no account of ours — you bring your own key. It takes about a minute:
1. Sign in or create an account at console.mistral.ai (email, or Google / Apple / Microsoft).
2. Open the console (Mistral Studio) and go to API Keys in the left menu (or the Create an API key quickstart link).
3. Open My API Keys and click Add a new key.
4. Fill in the Create API Key dialog: give the key a name (e.g.
SmartNote AI); you can leave Expiration, Connector access scope and
Workspace at their defaults.
5. Name it and confirm with New key.
6. Copy the key now — it is shown only once.
7. The key now appears in your list (masked). Paste the copied value into the plugin, door 1 · API key…. It is stored encrypted in the plugin's private directory, never in a cloud-synced folder.
Free tier note. The free tier lets you use every feature, but only a paid plan guarantees Mistral never trains on your data. Field-tested in July 2026, the free tier allows roughly ~1300 pages of transcription before you hit the monthly limit.
Every field in the plugin has Paste / Clear / Reset-to-default, and the text/button-size settings apply everywhere, floating panel included. The user guide ships as a PDF inside the plugin and can be restored from the Home page.
(Still: do not share confidential information you would not type into any cloud service.)
- Fully open source — audit this repository.
- Mistral AI only — a European company on European infrastructure. Your data stays under EU jurisdiction (GDPR). On a paid plan your requests are never used to train Mistral's models.
- Bring your own key, stored encrypted on the device.
- Transcripts and conversations are stored locally only, never synced. Files set to Off are never sent to the AI (a one-time per-conversation consent excepted, and that read is never saved).
I built SmartNote AI to talk to exactly one AI company: Mistral AI, based in France. That's my deliberate choice as the author, not a missing feature. And since you may already use ChatGPT, Claude, Gemini, or a Chinese model like DeepSeek or Qwen, I owe you the reasoning.
Your handwritten notes are some of the most personal data you own: journals, health, money, half-formed ideas, the names of the people around you. So the real question isn't just "is the model smart?" but "under whose laws does this text land, and who can force access to it?"
How SmartNote AI already limits what leaves your device. The design keeps exposure small before Mistral is ever involved:
- Your memory lives on the tablet. Every transcript, conversation and setting (and your encrypted API key) is stored on-device only, never in a cloud folder. Mistral is not your memory; your tablet is.
- Each request is transient. A request carries only what that one question needs; Mistral answers and keeps no memory of you. The single piece of short-lived server state is a prompt cache (the reason a follow-up question costs about 10% of the first), and it expires. Between requests, Mistral holds nothing about you.
- Off means off. A file set to Off is never sent at all (a one-time, per-conversation consent aside, and that read is never stored).
What the GDPR adds on top. Because Mistral is an EU company, those transient requests are handled under Europe's General Data Protection Regulation:
- No training on your data on a paid plan, and this matters more than a checkbox (see below).
- Purpose limitation: your text may be used only to answer your request, not quietly repurposed.
- Security and minimization duties: the company is legally required to protect your data and to keep no more of it than it needs.
- Real rights, real teeth: access and deletion are rights, enforced by independent regulators with fines up to 4% of worldwide revenue. Not optional PR.
The problem with US providers. Wherever the servers physically sit, US law can compel a US company to hand over data, through the CLOUD Act, and through FISA §702, a program that openly authorizes the surveillance of non-US persons' data held by US companies. If you're not American, that program is aimed squarely at you. And European courts have repeatedly ruled that US surveillance fails to meet European privacy standards, striking down two successive EU-US data agreements.
The "just turn off training" objection. Yes, several US providers now offer a no-training opt-out. But an opt-out is only as good as your trust that it's honored: you can't verify it, and no independent regulator stands behind it. With an EU provider, the no-training commitment isn't a toggle you take on faith; it rests on an enforceable legal framework.
The problem with Chinese providers. National-intelligence and data-security laws oblige companies to cooperate with the state on request, and data handling is opaque.
"I'm in the US or in China, does an EU model even help me?" Honestly, maybe more. Two things protect you here, and neither depends on where you live:
- Handling. Because Mistral is bound by the GDPR, your notes get European-grade treatment (no training on a paid plan, genuine security duties) that a domestic consumer AI usually won't give you.
- Jurisdiction. Your notes sit with a European company on European infrastructure, outside the easy reach of your own country's data-access laws. US authorities can't point the CLOUD Act at a French company; Chinese data rules don't reach EU servers. To obtain your data, a government would have to go the slow, formal, cross-border route, not a quiet domestic request.
No provider makes you invisible to a determined state, and you still shouldn't put something that would be genuinely dangerous for you into any cloud. But for everyday personal notes, moving them out of your own jurisdiction and under EU law is a real privacy gain, wherever you are.
"Is the quality as good?" Honest answer: for this job, reading your handwriting and answering questions about your own pages, the bottleneck is how legible your writing is and how well you've filled the glossary, not frontier reasoning. Mistral's OCR and vision models are genuinely strong here, and I chose them on measured results, not branding. On the very hardest open-ended reasoning a frontier US model may still edge ahead, and that's a trade I make on purpose. For a personal notebook, I think keeping the data in Europe is worth it.
You pay Mistral directly, per page read:
- A handwritten page costs roughly half a euro-cent (OCR + Vision); printed PDF pages cost less (OCR only).
- Unchanged pages are never re-paid. Every paid action shows its estimated price before you confirm.
- Chat questions are billed by token; follow-ups benefit from prompt caching (~10% of the first). The console at console.mistral.ai shows your live usage — you can set a monthly budget cap there to stay fully in control.
A few honest notes so the plugin meets your expectations:
- You bring your own Mistral key. There is no server of mine in the middle and no subscription to me — you pay Mistral directly for what you read, and only that. Creating a key takes two minutes (see above); a few euros of credit last a long time, and every paid action shows its price before you confirm.
- Auto reads only while the plugin is open. Background transcription runs as long as the SmartNote AI panel is open on the device (the Keep awake during Sync toggle stops the screen sleeping mid-backlog). Close the plugin and Auto simply pauses, then resumes where it left off. Manual and on-demand reads are unaffected.
- It is not instant. A page takes roughly 8–30 seconds to read on e-ink (render + OCR + Vision over your connection). A first sync of a whole notebook is a "start it and let it run" job, not a tap-and-wait.
- Accuracy depends on your handwriting — and the glossary is your lever. OCR and Vision handle most handwriting well, but messy pages, unusual names and jargon trip any AI. Feed those words into the glossary (door 2) and accuracy jumps noticeably.
- The interface is in English for now.
- Privacy, one honest caveat. Everything is stored only on the device and Off files never leave it. A paid Mistral plan is never trained on your data; the free tier may be — so keep anything sensitive on a paid plan, or set those notebooks to Off.
- A Supernote running PluginHost (the plugin framework). Built and tested on the Manta and the A5 X; other current models (Nomad, …) use the same framework and should work, but are less tested — please report what you find.
- Because a plugin rides on device internals, a Supernote firmware update could briefly break it until the plugin is updated. If something stops working right after a firmware update, check here for a new release first.
- Install is by side-loading the
.snplg(copy toMyStyle/, then Settings → Apps → Plugins → Add Plugin) — there is no app store.
If you enjoy this plugin, please consider sponsoring a few tokens ;-) My time and skills are free; the AI tokens behind this plugin are not! Thank you for your support ☕
- Project & issues: https://github.com/AgP42/SN-Plugin-SmartNoteAI
- Built for Supernote's PluginHost (React Native 0.79.2,
sn-plugin-lib). SPEC-v0.20.md/SPEC-UI-v0.20.md: architecture & UI specs.HANDOVER.md: engineering handover (version history, gotchas).docs/UI-TEXTS.md: every user-facing string, for review.
Tests: npx jest · Type-check: npx tsc --noEmit