Documentation / Diagon

Release notes

Last updated: 17 November 2026

What changed in each version of Diagon, newest first. Windows and AppImage installs update themselves; a .deb install needs the new package from the download page. Licensing & updates explains how updates arrive.

1.6.0

A GIF export: a picture of your diagram whose arrows keep moving, in slides, chat and e-mail too.

The reason. A flow arrow says that data is moving, and a picture of the diagram should say so as well. The SVG export has kept the dashes moving since 1.5.0, but only where a web browser draws the file. In slides, chat and e-mail an SVG stands still or is not accepted at all, and a PNG never moves.

  • Export → GIF (and File → Export GIF…) saves the diagram as a GIF in which the dashes of flow lines keep moving. It is twice the screen size, like the PNG, and repeats for ever without a jump.
  • The file stays small, because from one picture to the next only the dashes are stored again: the Arrow styles example makes a file of about 26 KB.
  • A GIF holds 256 colours at most. That is plenty for a diagram: large even areas keep their exact colour.
  • A diagram without flow lines makes a GIF that stands still, and the status bar says so. A very large diagram is saved smaller than twice the screen size.
  • The Export menu now says what each file is good for: SVG — vector, stays sharp, arrows keep moving; PNG — still picture, 2x; GIF — picture, arrows keep moving, 2x.
  • On a Confluence page the SVG is still the better file: its arrows move there as well, on the page and when you click the picture to see it in full, and it stays sharp at any zoom. The editor says which export to pick.

1.5.0

Line styles for arrows, pictures from more places in a diagram an assistant wrote, and a refreshed window in which Ctrl+Z reaches further.

The reasons, one for each part. An arrow could only say "this connects to that", and a diagram often has to say more: this link is optional, this is the main path, data is moving here. An assistant could take icons from one library only. And Ctrl+Z gave up the moment Diagon changed your text for you: a moved box, Auto-arrange or a choice in the toolbar ended the undo history.

Line styles

  • Write style: in the braces at the end of an arrow's line — api -> queue: publish { style: dashed }. There are six: solid (the default), dashed, dotted, thick, double and flow. The words dash, dot, bold and animated are read as well.
  • flow is a faint line with small bright dashes that travel toward the arrowhead, for data that is moving. The dashes stand still in an exported PNG, in a very large diagram, when more than 200 lines flow, when your computer is set to show less motion, and in every picture handed to an AI assistant. An exported SVG keeps them moving when you open it in a web browser.
  • stroke: sets the colour as before, and the two go together: { style: flow, stroke: #54d18c }. A style works with every kind of arrow, and on a chain every arrow of that line gets it.
  • A style Diagon doesn't have is an error with its line number, and the arrow is drawn as a plain line.
  • The suggestion list offers style: and its values inside an arrow's braces, and Help → Examples → Arrow styles shows all six. An assistant connected over MCP is told about the styles and what each is for. The language has the table.

Pictures in a diagram an assistant wrote

  • If you never connect an assistant, nothing changes: an icon address you type yourself loads from any link, as always. (An address to a website that an assistant once wrote into one of your diagrams on this computer is treated as the assistant's; AI assistants explains.)
  • An assistant connected over MCP could add icons from Simple Icons only. Now three sources load straight away: Simple Icons, Iconify, and picture files in public GitHub repositories (raw.githubusercontent.com) — the cloud icons for Azure, AWS, Google Cloud and Kubernetes come from there.
  • For a picture from any other public website Diagon asks you first, in a bar at the bottom left: This diagram wants to show pictures from example.com. — Allow or Not now. Until you allow it, the icon is a dashed placeholder. Diagon remembers the websites you allow and, while the connection is switched on, lists them under Pictures in Help → Connect an AI assistant…; one switch there stops the question for good.
  • An address that is not an https:// link, or that leads to your own computer or a machine on your network, is never loaded for an assistant. Diagon looks up where a website name really leads before every request. Such a picture stays a placeholder with a "no entry" mark; point at it and it says Not loaded — click to see why. The notice shows the address and where it leads. If it is an ordinary web address and it is yours after all, It is mine — load it takes it as your own.
  • This is a safety rule, and it has nothing to do with licences: a diagram an assistant wrote must not make your computer call addresses that only your computer can reach, or send things out packed into an address.
  • To tell an assistant's addresses from yours, Diagon keeps a fingerprint of the text assistants wrote — not the text itself — in its own data folder on your computer. Every open Diagon window shares it. AI assistants has the details and the limits.

The window

  • The toolbar is one row of buttons with icons: New, Open, Save, Export, Fit, Auto-arrange, Direction, Routing, Auto-color, a moon (or sun) button that switches the picture between dark and light, and Connect AI. In a narrow window some buttons show only their icon; rest the mouse on one to see its name. The dot on Connect AI is grey while the connection is off, amber while Diagon waits for an assistant, and green once one is connected.
  • Export is one button with a menu of two: SVG and PNG.
  • A status bar runs along the bottom: on the left the count of nodes and edges and the drawing time, on the right the line and column of the text cursor and the zoom of the picture.
  • The preview has zoom buttons in its bottom right corner — − and +, the percentage between them (click it to go back to 100%) and Fit. An empty diagram shows a small card with a line to try.
  • In the editor a faint band marks the line the text cursor is on, and the number of a line with a mistake turns red. That works inside boxes too: errors in a diagram with boxes now say which line they are on.
  • Help → Examples opens four diagrams that come with Diagon: Login flow, System architecture, Brand icons and Arrow styles.
  • The editor font is Cascadia Mono where it is installed, and the canvas keeps its grid.

Ctrl+Z and the suggestion list

  • New and Help → Examples replace the diagram on screen, and Ctrl+Z now brings your previous text back. If that text had unsaved changes, a notice says so. The text comes back as an untitled diagram, so Save asks where to put it.
  • A moved box, a choice of Direction or Routing in the toolbar, and Auto-arrange are undo steps, like anything you type. (In a diagram with boxes Diagon does not write these into your text, so Ctrl+Z has nothing to take back there.) Only a very large rewrite is different: when the first line Auto-arrange changes and the last are more than 400 lines apart — as a rule the first box you placed and the last — it is not an undo step, and the undo history starts anew there.
  • Open ends the undo history: Ctrl+Z cannot go back past opening a file.
  • While a dialog is open, Ctrl+Z leaves your diagram alone; the step is still there once the dialog is closed.
  • The suggestion list that appears while you type opens above the line when there is no room under it, and stays inside the editor.

What changes for diagrams you already have

Nothing: every file opens and draws as it did. One thing to know — style: in an arrow's braces used to be ignored and is read now, so an arrow that already carried one is drawn in that style, or reported if Diagon has no such style. The subscription and the licence are unchanged. If you use an assistant, restart it once after the update so that it connects to the new version.

1.4.0

A layout update: diagrams come out tighter and symmetric, icons come in three sizes (bigger by default), and lines behave at the closer spacing.

The reason for all of it: a diagram that looks fine on a big monitor turned into confetti once a PowerPoint slide on a laptop shrank it to fit. Most of what the slide was shrinking was empty space.

Layout

  • Rows sit about one box-height apart (48 px; it was 88), and boxes within a row 24 px apart. A row grows only where something needs the room — an edge label between two rows, or a wide fan of children — so a long label in a left-to-right diagram is no longer cut off by the boxes at its ends, and a plain flowchart is 25-30 % shorter.
  • A parent now sits centred over its children and a chain stays straight. Before, every arranging pass shoved the lower half of a diagram a little further to one side, so a fan of children ended up beside its parent, chains became staircases, and the shipped architecture example had one long diagonal line across an empty region. That example is now 60 % narrower, and fitted to a laptop slide its labels are 9.5 px instead of 7.
  • Over the shipped examples the share of the drawing that is actually boxes went from 22 % to 31 %.

Boxes

  • Unconnected things inside a box form rows of a few — seven become 4 + 3, not a 3 × 3 grid with holes — and each row is as tall as its tallest member. In a left-to-right diagram they form columns the same way.
  • A box that holds a small flow plus loose children puts the loose ones in a block beside the flow instead of one long row across the top; this applies at every level, the diagram itself included.
  • A box is never narrower than its title or the labels on its own lines, and box frames lose a third of their air. A 67-node, 13-box architecture of the kind customers draw comes out half the size it was and twice as full.

Icons

  • Icons come in three sizes — small (32 px), medium (48 px) and large (64 px) — chosen in the text with one word, icon-size: on a line at the top for the whole diagram, inside a box for that box, or in a node's { … } for that node. The nearest setting wins; a number of pixels (16-96) also works. The language has the details.
  • A wrong icon-size value is an error with its line number, wherever it is written.
  • Medium (48 px) is the new default, so every diagram with icons gets icons 41 % bigger without changing anything. icon-size: small is close to the old size (32 px; it was 34). For a diagram that will go on a slide, icon-size: large on the first line.
  • An icon on a cylinder sits under the rim instead of across it. An assistant connected over MCP is told about the sizes and the rule for slides.

Lines

  • In a diagram without boxes, routing: orthogonal now goes around the other boxes; it used to draw straight through them.
  • Arrows leave by the bottom of a box and enter by the top (right and left in a left-to-right diagram). A hub with more lines than its bottom edge can hold hands the most sideways ones to its side edges, so arrowheads no longer merge.
  • A line through a gap runs down the middle instead of hugging a box, parallel lines sharing a gap are spread apart, and the small stair-steps where a line left a box are gone.
  • routing: curved in a diagram with boxes is a smooth rounded route that never cuts a box. It used to loop, double back and cross boxes — 26 lines through boxes in the customer-style architecture; now none.
  • A line never runs through a box's title, and an edge label never sits on a box.
  • A line keeps off box frames.
  • The careful routing costs time on very large boxed diagrams: one with 600 lines takes about 0.7 s to redraw instead of a tenth of a second. If a diagram is ever too large to route in time, the status bar says how many lines were left straight. A hub with very many lines can still have arrowheads that touch.

What changes for diagrams you already have

Every diagram gets tighter and moves. A box you pinned by hand (x:/y: in the text, or a drag) keeps its absolute spot while the automatic part shrinks around it, so it may look farther away than before — Auto-arrange once fixes that. A decision diamond now sits centred over its two outcomes, so a chain of decisions steps sideways a little instead of running dead straight. A very large diagram (dozens of boxes) is still too much for one laptop slide even at this spacing: its labels come out at about 3 px once the whole thing is fitted. Split it, or show it on a page you can zoom.

1.3.1

A maintenance update: a newer engine underneath, and updates that no longer close a window you were working in.

The engine

Diagon is built on Electron, which carries the Chromium browser engine. 1.3.1 moves it from Electron 43.0.0 to 43.7.9 — nineteen maintenance releases newer — and with it the security fixes made to Electron and Chromium since June. Diagon shows pictures fetched from the web (the icons in a diagram), and the part that does that should not be months behind on fixes. Your diagrams, files and settings are untouched.

On Linux this Electron has a known fault with the add-ons that KDE Plasma and some other desktops load into every program: it crashes at start-up where a global menu is used (Plasma's, and some Xfce, MATE and Budgie panels), and on any KDE Plasma desktop it crashes when the colour scheme changes while the program is open. Diagon works around both, so it starts and stays open there as 1.3.0 did — and from this version on, every Linux package is started on such a desktop, with the colour scheme changed under it, before it is released.

Updates

  • An update no longer closes another open Diagon window. On Windows, with Diagon open twice, quitting one installed the waiting update, and the installer closed the other window as well — unsaved work in it was lost. An update now installs only when Diagon is closed with no other Diagon window open, and Restart now says so instead of restarting. On a PC where Diagon is installed for all users, the same holds while Diagon runs in another signed-in account.
  • Restart now restarts quietly. On Windows it used to open the setup wizard. It now installs in the background and starts the new version. It warns first if the diagram has unsaved changes, and if the update cannot be started it says so and fetches it again.

Both are in 1.3.1 itself, so you will notice them from the next update on: 1.3.1 is still installed by the version you have now.

Smaller things

  • Windows names Soverain s.r.o. as the company in the program's properties and the list of installed apps (it said "Haustorius").
  • The user guide and the licence text in the Linux packages carried Windows line endings.

1.3.0

The built-in AI is gone. Your own assistant connects instead.

Up to 1.2.1, Diagon shipped with a small AI model inside the app and a ✨ Generate panel that turned a sentence into a diagram. In 1.3.0 that model, the panel and the button are removed. In their place, Diagon can be connected to the AI assistant you already use — Claude, Cursor, Copilot in VS Code, Codex and others — through the Model Context Protocol (MCP).

What you can do now

  • Ask your assistant, in its own chat, to draw or change a diagram, and it appears in your Diagon window.
  • The assistant learns Diagon's language from Diagon itself, has its text checked by Diagon's real parser, and can look at the finished picture and tidy it — or look at a draft without touching what's in your window.
  • An assistant's change is one Ctrl+Z away from undone, and nothing is saved until you save it.
  • Set it up in Help → Connect an AI assistant…. It's off until you switch it on, and it's included in the trial and in every subscription.

AI assistants has the setup for each assistant and the details.

Why

A model small enough to bundle and run on any laptop can only do so much, and it added about 1.6 GB to every download. So Diagon no longer ships one. If you use an AI assistant, Diagon now works with that; how well it draws depends on the assistant, not on Diagon.

What this changes for you

  • The download is about 100 MB (130 MB for the Linux AppImage) instead of about 2 GB, and the 8 GB memory recommendation is gone.
  • Diagon itself contains no AI and still sends your diagrams nowhere. The difference is on your assistant's side: what it reads from your diagram goes to its provider, under your agreement with them. With the old built-in model, that text never left your machine. If that matters for your diagrams, leave the connection switched off — everything else in Diagon works exactly as before.
  • If you used ✨ Generate, you now need an assistant of your own to get the same help. Diagon doesn't supply one.
  • If you never used it, nothing changes except the size.

If you subscribed for the built-in AI and this isn't what you want, write to hello@soverain.cz. A charge from the last 14 days is refunded in full on request, and the refund policy says what applies when Diagon no longer matches what we described.

Also in this version

  • Exports no longer cut off lines. SVG and PNG exports were cropped to the outermost boxes, so a loop, a line detouring around a pinned node, or a label on a long bend could be clipped. They're now cropped to everything that's drawn.
  • A web address written where a colour belongs is no longer fetched. A diagram could put an address in place of a colour (fill: url(…)), and Diagon would request it while drawing. Only real colours are applied now, whoever wrote the diagram.
  • A { on a line of its own froze the preview. A block with no name in front of it made the picture stop following the text, with no message. Diagon now says the block needs a name and draws the rest.
  • Help → Open full User Guide did nothing in an installed copy. It opens the guide now.
  • An icon that failed to download once stayed missing from exports until Diagon was restarted. It is tried again after a minute.
  • The icons example and the in-app guide pointed at an icon that no longer exists; both are fixed.

Earlier versions

Notes for 1.2.1 and earlier are on the releases page.