Development
Built in the open, including the parts that are broken.
This page is the working notebook: what is being verified, what is known to be wrong, and what is deliberately not being built. Nothing on it is a promise about the version you can download today — the home page is where capabilities live, and every claim there is pinned to a source.
Last updated 4 October 2026 · version 1.1.5
How the work is judged
Four rules that decide what ships.
Nothing is claimed until it is proved on the saved file
A feature that looks right on screen is not evidence. The proof we accept is the file itself — open the saved document as an archive and find the value in the XML. It is slower, and it is the only check that cannot be fooled by the application agreeing with itself.
The defect list is published, not buried
Where verification finds something broken, it goes on the record before it is fixed. A known limitation you were told about is a different thing from one you discovered on your own work.
The AI never interrupts
No floating buttons, no modals, no upsell, no assistant volunteering itself over the document you are working on. It answers when you summon it, and until you do it writes nothing and asks you for nothing. A docked panel sitting in its own column is deliberately not counted as an interruption — available is a different thing from pushed at you. This is a design rule with a test behind it, not a preference: the test opens the app, opens a document, waits for anything delayed to surface, and fails if an AI overlay, modal, floating button, nag or upsell has appeared, if any dialog on screen is about the assistant, if the copy tries to sell AI, or if the assistant has produced a single word on its own.
Fixes travel by download
The application does not update itself, so improvements described here reach you only when you choose to install a newer version. That is the trade for a program that does not phone anywhere.
On the bench
What is being worked on, and why.
In order. Several of these exist because a review in August 2026 checked the suite against its own documentation and found the documentation too generous.
-
In progress
One verification harness, covering every published claim
A fixed corpus of real-world documents, opened and re-saved, with assertions that read the resulting file rather than the screen. It runs on every change. The rule it enforces is symmetrical: nothing is said on this site that the harness does not prove, and nothing is proved that the site does not say.
-
Done
1.1.5: setting up the AI, redaction, text recognition, signing, form fields, version history, the Reviewing Pane and Check Style
In the 1.1.5 downloads for Apple silicon Macs, Windows and Linux. The first time the app starts with AI on and nothing to run it, a Set up AI assistant checks the computer, helps install and start Ollama (one click on Windows), downloads a model sized to the machine and checks it answers; skipping it switches AI off rather than leaving every request to fail. A spreadsheet can be saved as CSV with a choice of encoding (UTF-8, UTF-8 with a byte-order mark for older Excel, or Western-European Windows) and of separator (comma, semicolon or tab); it writes the sheet you are on, as displayed. Western-European Windows CSV files open with their accented characters intact, and so do they when attached to the assistant. The Linux AppImage no longer needs the FUSE 2 library, and the .deb and .rpm now declare the sound and graphics libraries a minimal system was missing. The PDF editor can redact: drag a box, and on save the text, annotations and picture pixels under it are deleted from the file rather than covered, scanned pages included. It can recognise the text on scanned pages, in English and on this computer, and save a searchable copy; add text fields and checkboxes; and sign a copy with a certificate made on this computer or imported. On Windows and Linux the word processor checks spelling offline, in English. It keeps the last 20 saved versions of a document, which you can open as a copy, restore or delete under File, Browse Versions; lists every tracked change in a Reviewing Pane you can filter and accept or reject from; and, with AI on, suggests clearer wording a paragraph at a time under Review, Check Style. Slides can loop a show until Esc or run it as a kiosk. The limitations page says where each of these stops.
-
Done
Slicers survive being saved
A slicer — the button panel that filters a pivot table — used to exist only while the workbook was open; the parts of the file that would carry it were never written. It is now written, read back, removable and retogglable, and deleting a sheet takes its slicers with it. The first fix failed its review twice — once for a stray rectangle it left on the grid, once for a remove button that silently did nothing — and shipped only after a third review found nothing left. That sequence, not the feature, is the reason this page exists.
-
Next
Signed installers
The current builds are unsigned, so the operating system warns about them on first run. For software asking to be trusted with documents that never leave the machine, a warning at the front door is the wrong first sentence. Signing and notarisation come before any wider release.
-
Done
Three fixes people have asked the incumbents for, for years
Centre text across several columns without merging the cells, so sorting and selection keep working. Paste data that is left exactly as typed — no silent conversion of part numbers or gene names into dates. Open a CSV in UTF-8, UTF-16 or a common Chinese, Japanese or Korean encoding without the characters turning to rubble. Small, unglamorous, and the reason people swear at spreadsheets. The first two are on a Paste Special entry and the alignment menu; the third applies wherever a CSV is read. Western-European Windows files were the one case still wrong in 1.1.4; 1.1.5 fixes them.
-
Done
Compressing a PDF, without ever making it bigger
Compressing is the most common thing people do to a PDF and this suite could not do it at all. Over-resolution images are re-rendered and re-encoded, then the file structure is repacked — but every step is only a proposal, and the original is kept unless a candidate is genuinely smaller. Transparent images and one-bit stencils are left alone, because re-encoding those visibly damages them.
-
Done
A first run that does not send you to a terminal
Someone with no local model was previously told to run a command in a terminal, with no model named and no hint that the suite works completely without any of it. Both status messages now say what is actually needed — a model that supports tools, because only those can edit a document — where to get one, and that the assistant is the only part that needs it.
-
Investigating
Showing you the network, not just describing it
The privacy notice describes where the application does and does not connect. The stronger version is a panel that shows it, plus instructions for checking with your own tools rather than taking ours. Before any of that is published it has to survive a hostile packet capture of the packaged application, on all three platforms. Until it does, the wording on this site stays as it is.
Deliberately not building
The gaps we are choosing to keep.
A short list of things this suite is not going to do. They are here so you can rule the software out quickly if one of them matters to you — which is a better outcome for both of us than finding out after you have moved your work across.
Macros and the automation language
Decades of recorded macros exist in the world, and a half-working imitation is worse than an honest gap. If you depend on them, that is a real reason to keep the software you have.
The data-model and query tooling
The same argument. It is the genuine strength of what people already pay for, and pretending otherwise would waste your time and ours.
A phone version
The whole design rests on a model running on your own machine, and phones cannot run one. A convincing phone app would have to send your documents somewhere, which is the one thing this suite exists not to do.
Optical character recognition, for now
Turning a scanned page back into editable text is a large piece of work to do properly and offline. Where a PDF is images rather than text, the application should say so plainly instead of pretending to read it.
Where this came from
A fork, and what was taken out of it.
This suite began as a fork of an open-source project whose editors were free to modify but whose intelligence was not: the assistant ran as a hosted service, behind an account, spending credits. That arrangement is the single thing people consistently object to about the original, and it is what this fork removed — the vendor client is deleted from the source, not switched off, and replaced with a model running wherever you point it, which by default is your own computer.
The engineering that followed was mostly subtraction, and subtraction is harder than it sounds. Removing a cloud feature leaves behind the instructions that told the assistant to use it — a sentence in a prompt, a tool's own description, a guard that redirects to the thing you just deleted. Those fail quietly and plausibly: the application looks fine and the assistant answers politely, having been told to do something impossible. Finding them took driving the real application end to end, repeatedly, and reading the files it saved.
That experience is why this page is written the way it is. The failures worth publishing are the ones that do not announce themselves.