What's actually running

RUNNINGWriting on a Boox Go 10.3 Lumi. Native app, raw Onyx pen strokes with pressure. Ink latency 30 ms → 5 ms. Measured 31 Aug 2026.
RUNNINGReading real handwriting. 87.5% of words exactly right on a real handwritten line. Diacritic errors 7.1%, and the index strips diacritics, so accent-free typing still finds the page.
RUNNINGBacklog import. A real archive run: 11 pages, 27 regions indexed. Typing speed up, really build and UI UX each returned a page that had been unfindable by text before.
RUNNINGreMarkable file import. .rm v6 read directly, the strokes, not a flattened image. Verified against four real files from the community test suite.
RUNNINGCross-notebook search. SQLite + FTS5, one index across every notebook and every device. Click a hit and it lands on the exact strokes.
RUNNINGMCP server. 16 tests, and a real MCP client connects to it.
RUNNINGDesktop app. Electron; library, page redrawn from strokes, search, recognition, correct-and-confirm. 12 store tests that run without opening a window.
RUNNINGBrowser writing surface. Runs in the tablet's own browser on any e-ink device, nothing to install, ~30 ms latency instead of 5.
NOT YETDiagram rebuild. The stroke classifier, box vs arrow vs circle vs text, is the next engineering task and the first attempt scored zero. The screens exist. The rebuild does not. Honest state: not working.
NOT YETScrawled pages and diagrams. Recognition is good on writing laid out in lines (~97% of regions). On rough scratch pages it is not reliably measured, and we won't quote a number we don't have.
NOT YETSupernote and Kindle Scribe. Not read.
NOT YETPhotographing paper notebooks. The biggest expansion of who this is for, and not started.
RUNNINGDevice pairing. The tablet asks, your computer shows you which device is asking, and only after you approve does it hand back a 32-character key that every later request has to carry. Six endpoints check it, with no exception — the browser writing surface asks for its own key the same way a tablet does. Two limits are real and worth saying: until you pair the first device the server accepts whatever the address filter lets through, and the traffic is plain HTTP inside your own network, so the key is readable on the wire. Both are written out on the security page.
RUNNINGInstallers get built. macOS (.dmg, both chips), Windows (.exe) and Linux (.AppImage), rebuilt on demand, each one on its own machine. The Linux build is launched in a virtual display afterwards to check it actually opens, because a build tool reports success when it finishes writing a file, not when the app runs.
NOT YETSigned installers. No certificate is held, so macOS and Windows would warn on first launch. That is why the buttons on get the app are still greyed out.
RUNNINGCommitment extraction. A line struck through on paper marks it done. Deadlines are read out of the words beside it and anchored to when you wrote them, so “by Friday” written three weeks ago is overdue, not this Friday. Who you promised it to is pulled from the same line, and every name shows the words it came from, because a wrong name reads perfectly. All three land in the app and in the Markdown export.
NOT YETMeeting-audio index. The other half of slice three: handwriting as the index into a recording. Not started.
NOT YETCertifications. None held. The honest table is on the security page.
NOT YETHardware. Nothing has been manufactured. Four written conditions gate it, see the device.

What comes next

Three slices, in order.

Each one has to reach a real person's hands before the next one starts. No horizontal layers, no half-built platform waiting for the rest of itself.

SLICE 01 · RUNNING

“Pour the old archive in, and find things”

Import raw strokes from reMarkable and Boox, not PNGs. Local index, full-text search across every notebook and device, batch processing of the backlog. Even the rough version beats what the whole industry ships.

SLICE 02 · PART RUNNING

“A scrawled map, cleaned up”

Export runs. A map read off your strokes leaves as a Mermaid block that renders in GitHub or Obsidian, and every page leaves as SVG, both inside the Markdown export. The hard half doesn't. Stroke-level semantics — which shape is a node, which is an edge — is the classifier whose first attempt scored zero, and re-layout, OPML and the benchmark all wait behind it.

SLICE 03 · PART RUNNING

“Don't forget what you promised”

Commitments run. Struck through on paper means done; the deadline is read from the words beside it and anchored to when you wrote them; who you promised it to comes from the same line. The audio half hasn't started — handwriting as the index into a meeting recording. One user has already built that by hand, which is the clearest signal there is.

Where this is heading, stated once and not sold as a slogan: the notebook is one input. An audio recorder is the second, with evidence behind it. The immutable archive underneath, and the MCP door on top of it, are what stay constant. A second input does not open until the first one works end to end.

Software pricing

Draft, and we'll say so.

Two pieces of evidence point in opposite directions and both are real: some of the people who want this most are furious about subscriptions, and yet 568,000 reMarkable owners pay a monthly fee. Which is true depends on who turns out to want this, which is exactly what the wishlist is for.

Tier	Draft price	Who	What's in it
Free	$0	Anyone	Local import and search, one device, 200 pages
Personal	$9/month	Knowledge workers	Unlimited, every device, diagram rebuild, MCP
Pro	$29/month	Consultants, solo lawyers	Commitment extraction, meeting-audio index, batch export
Firm	$49/seat	Practices, 5 seats up	Encryption, MDM/Intune, SSO, remote wipe, data residency, BAA, SOC 2

Two things we won't do: charge a subscription purely to unlock sync, that's the thing reMarkable customers resent most, and inheriting that resentment is a gift, and lock your data in. The 200-page free tier is a limit on volume, not on days, because the moment this product earns its keep is when you pour your old archive into it.

For scale: break-even is around 730 paying customers. Not a million users. reMarkable alone has roughly 600,000 people already paying for note software, 730 is 0.12% of that.

See the devices