New Bring your own storage. Keep your diagrams in your own private GitHub repository, on Pro and Team.Bring your own storage on GitHub, for Pro and Team. Learn more about bring your own storage

changelog

New features, improvements, and fixes to the Zindex platform.

Week of · 76 changes · Every week

new

Workflow: Pools and lanes lay themselves out

Workflow: a pool, a frame with containerType: "pool" whose children are lane frames listing their tasks, events and gateways, needs no coordinates. The engine stacks the lanes across the flow in the order the pool lists them, places each node in the lane that lists it, and shares columns across every lane and pool, so the flow reads as time from one lane to the next. Each lane is as deep as its contents need, at least 60 px and enough for its title, and several pools stack 40 px apart in document order, all as long as the longest.

new

Long node labels wrap on their own

textStyle.wrap takes "auto", the new default for nodes: a label stays on one line up to 250 px of text, the node widening to hold it (a diamond up to 400 px wide), and a longer label wraps at spaces into lines of even length, the node growing taller instead of wider. Diagrams saved before this store wrap: "none" on their nodes and draw as before through later edits, Mermaid edits included, and updateNode or setText with "auto" wraps their labels, while a node an edit adds gets "auto". Text elements keep "none" as their default, and "auto" on one wraps by word only when the text is wider than its box.

new

Edits add text and images to saved diagrams

The ops createText and createImage add a caption, legend, title or picture to a saved diagram in an edit, through diagram_edit or POST /v1/diagrams/{id}/revisions, where before text and image elements could only be made when a diagram was created, and createNode with kind: "text" or "image" is refused with the op to use. An image shows an entry of the scene's assets, which updateScene now adds or replaces by ID in the same edit or an earlier one, best as a data: URI, since a remote one gets ASSET_URI_POLICY_VIOLATION and is left out of PNGs. The published JSON Schemas, at https://zindex.ai/schemas/ and in the open-source spec repository tagged v8.8.0, have both ops.

new

Edits add nodes to frames that already exist

createNode takes an optional frame, the ID of a frame that exists when the op runs, and adds the new node to that frame, which is how an edit puts a task in a lane already there. updateNode with frame moves a node into that frame and out of every other frame and group that lists it, such as a task to another lane, where before both took deleting the frames and creating them again with the node among their children. A frame naming an ID that doesn't exist yet is refused with OP_TARGET_NOT_FOUND, and one naming a group or a node with OP_TARGET_KIND_MISMATCH.

new

Diagrams can be renamed on their page

On a diagram's page in the dashboard, click the title, or choose More, then Rename, and type the new one: Enter saves it and Escape cancels. A rename saves a new version, so History shows who renamed the diagram, with a message naming the old and new titles, and if someone else saved since the page loaded, the rename applies to the latest version, unless they renamed it too, when the page shows their title and Rename again keeps yours. Owners and members can rename a diagram and viewers can't, while agents rename one as before, with updateScene title in diagram_edit.

new

The Diagrams list is a grid of thumbnails

The dashboard's Diagrams list shows a grid of cards by default, each with a 16:9 thumbnail of the diagram on its own canvas color, its type, its number of versions, when it last changed and, in a team workspace, who saved it, with Present beside them. List keeps the table, and the page remembers which view you chose in each browser.

new

Stars mark the diagrams you come back to

Star a diagram from its card or row on the Diagrams list, or beside its title on its own page, and Starred lists only the diagrams you starred. Stars are yours alone, and anyone who can read a diagram, a viewer or a read key included, can star it. diagram_find and GET /v1/diagrams mark yours with starred: true, and PUT and DELETE /v1/diagrams/{id}/star star and unstar a diagram over REST, while no MCP tool does.

new

The Diagrams list filters by diagram type

A Type select on the Diagrams list shows one family (Architecture, Workflow, Entity-relationship, Sequence, Network, Org chart or UI flow) or the diagrams with no type, together with search and Starred, and Clear filters removes them all. The search and filters sit in the page's address, so Back and reload keep them. Over REST, GET /v1/diagrams takes diagramFamily (a family, or none for diagrams without one) and starred=true.

new

A thumbnail image for every saved version

GET /v1/images/{id}@{revision}.thumbnail.png returns a 640 x 360 PNG of one version on its canvas color, without its address line, as the Diagrams list shows it. It can be cached for a year. Thumbnails count against a rate limit of their own, so a page of them doesn't use up the API limit your agents work within.

new

A warning names the node an edge crosses

A straight or curved edge whose line or curve runs through a node it doesn't join gets the warning EDGE_CROSSES_NODE, one per edge, naming the node. Such an edge keeps the shape you set and isn't routed around nodes, so the engine reports the crossing rather than moving the edge; it measures on the node's outline, and a line that only passes a round or slanted node's empty corner isn't reported. The fix sets router: "orthogonal" with updateEdge, or leaves router off, so the engine routes the edge around nodes, or moves the node off the line with updateNode.

new

Workflow: Warnings name nodes and lanes out of place

Workflow: POOL_NODE_OUTSIDE_LANE names a node that is in a pool but in none of its lanes, and its fix says to move it into a lane with updateNode's frame field, since until then it is drawn in an untitled band after the lanes. LANE_OUTSIDE_POOL names a lane that no pool lists, and its fix says to list it in a pool's children, in an edit by deleting the pool and creating it again with the lane among them, since until then it is drawn as a pool of one lane. Both come back in a save's warnings, as other validation warnings do.

new

Zindex is listed in Claude's connector directory

Zindex is in Claude's connector directory, with the Community label, at claude.ai/directory/connectors/zindex, for claude.ai and the Claude apps. Choose Connect to Claude and sign in, and your agent finds, saves, edits and renders your diagrams; on Claude's Team and Enterprise plans an Owner adds Zindex for the organization first, and Team members can request it. The connect page and the tools guide start Claude users at the listing, and a custom connector with https://api.zindex.ai/mcp still works.

new

Claude pages show a diagram live through Zindex

A Claude page that calls the Zindex connector's diagram_render with a share link or a published diagram's zindex.ai/d/ address, and preview: true, shows the diagram live. The call needs no sign-in, works from any workspace, doesn't count against the owner's plan and, through a share link, never returns the title. Viewers without Zindex connected see no picture, and neither does anyone who opens a public link signed out of Claude or from outside the page's organization.

new

knownRevision keeps a page's refresh cheap

diagram_render takes knownRevision, a revision you already have, and GET /v1/diagrams/{id}/render takes ?knownRevision=. When that is still the version the call would render, the answer is { id, revision, unchanged: true }, with no links or image, and isn't counted as a render. A page that refreshes its picture gets the PNG again only once a new version is saved.

new

Shared diagram links preview with their picture

The link-preview crawlers of X, Facebook, LinkedIn, Slack and Discord may fetch a shared or published diagram's zindex.ai/d/ page and image, so its link shows its picture. Every other crawler is still asked to skip /d/, and its pages answer with noindex, which keeps them out of search. Each /d/ page gives its image's size and type in og:image tags, and GET /v1/public/diagrams/{ref} returns width and height, so a preview lays out its card before the image loads.

new

Links to zindex.ai show a picture on X

A zindex.ai page without its own picture, such as the homepage or a docs page, previews on X and other sites as a 1200 x 630 card. The card shows the Zindex logo, the tagline "Immutable agent diagrams" and a diagram from the homepage, where before X showed a blank placeholder. X keeps a link's card for 7 days, so a link posted there before today may keep its old card until then.

new

Each example page previews with its own card

A link to one of the seven example pages previews on X, Slack, LinkedIn and other sites as its own 1200 x 630 card, the example's title beside its diagram as the page draws it. Before, every example page previewed with the default Zindex card.

new

The Zindex plugin installs in Antigravity CLI

In Antigravity CLI, agy plugin install https://github.com/zindexai/zindex installs the Zindex plugin, the Zindex skill and the MCP server: sign in when it asks, and your agent finds, reads, edits and renders your diagrams. Antigravity CLI takes Gemini CLI's place among the apps named in the plugin's README and the tools guide, since Google stopped serving Gemini CLI to individual Google accounts on June 18, 2026. The plugin still works in Gemini CLI with a Gemini API key or a Gemini Code Assist license.

new

The sign-up page runs Cloudflare's Turnstile check

Opening the sign-up page runs Cloudflare's Turnstile check, which most people never see: it asks for a click only when Cloudflare isn't sure, and if it can't load or finish, the page says what to do. The privacy policy, updated 2026-10-28, lists it under Cloudflare, with what it reads when you open the page: your IP address and details of your browser and connection.

new

Policy changes are emailed 15 days ahead

The privacy policy and terms each end with a dated list of their changes, newest first. Before a material change to either takes effect, or a sub-processor is added or replaced, Zindex emails every account holder at least 15 days ahead, with no need to ask for it. The first entries, dated 2026-10-29, took effect at once, since they describe what was already true or add to what Zindex promises.

improved

Decision branches leave by their own corners

Each branch out of a decision or BPMN gateway leaves by a different corner of the diamond: the branch whose target lies ahead along the flow takes the corner facing it (the bottom, in a top-to-bottom diagram), and each other branch the side corner facing its target, where before branches could share one line out of the bottom corner and a branch could cross a node on its way out. Yes and no labels sit beside their decision, at the nearest clear place along the branch within 200 px of it, on plain decision nodes, such as a Mermaid import makes, as well as on gateways. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

Edges into a decision join with one arrowhead

Two or more edges entering one side of a flowchart decision join on a short trunk into its corner, with a dot where their lines meet and a single arrowhead, as a flowchart draws one flow line into a decision, and each keeps its own line until the junction. An entry drawn differently from the rest, such as a dashed edge among solid ones, takes a free corner of its own, and BPMN gateways and workflow.decision keep an arrowhead per sequence flow. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the joins.

improved

Extra decision branches split from a shared corner

A decision or BPMN gateway with more branches than free corners, such as four in a top-to-bottom diagram, sends the extra branch out of the corner facing its target together with the branch already there, as one line that splits toward both targets so each enters its target's facing side, where before the extra branch left from one of the diamond's sloped sides. A flowchart decision draws a dot at the split, which the computed path lists in junctions, while a workflow.* diamond splits without one, and each branch's label sits on its own arm past the split. Only branches drawn alike share a corner, so a dashed branch among solid ones still leaves by a side, as it does when an anchored branch, a loop back or a gateway's label holds the corner, and saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

Edges enter the side that faces their source

An orthogonal edge enters its target by a side facing where it comes from whenever a route there is clear, so an edge from above enters the top, and it leaves its source straight out instead of running along the source's border first. When a node in between blocks the straight line, the edge runs straight past it, 16 px clear, instead of detouring around to another side, and edges converging on one side take their places in the order they arrive from, so they don't cross on the way in. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

Workflow: Pools and lanes draw BPMN header strips

Workflow: each pool and lane draws its title in a 30 px header strip, as the BPMN 2.0.2 specification does, down the left of a horizontal pool with the title reading bottom to top, or along the top of a vertical one, with a line dividing a pool's strip from its lanes and none beside a lane's unless the lane holds lanes, where before a 24 px band crossed the top of every pool. Lane borders stay unbroken where flows cross them, where before they broke for 8 px at each crossing, and a lane title that wouldn't read on the lane's fill takes the nearest shade that does. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the strips.

improved

Workflow: New tasks grow their lane

Workflow: a task an edit adds to a lane without a layout takes a free place in its column inside the lane, and when there is none the lane grows and everything past its far border, later lanes and pools included, moves across the flow by the growth, where before a task added later could stack under another or land behind the task it follows. A task added between two placed columns of a pool opens a column, moving that column and everything past it along the flow in every lane, and a task updateNode moves to another lane is placed again in its new lane, in its column. Other placed nodes keep their positions, and the edit's changed lists every node these moves place.

improved

Workflow: Flows cross lanes on direct routes

Workflow: a sequence flow crosses the lanes of its pool on a direct route, where before each lane holding neither end was an obstacle that the flow went around, or crossed with EDGE_OBSTACLE_GRAZED. A message flow crosses the lanes of the pools it connects and keeps clear of other pools, a route crosses a lane border rather than running along it, and edges keep off pool and lane header strips as they keep off other frames' titles. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

Workflow: Edges keep off event and gateway captions

Workflow: an edge ends on a face of a BPMN event or gateway that its caption leaves free whenever one gives a clean route, so in a left-to-right flow, where captions sit below their shapes, a flow to or from a lower row meets an event at its side instead of running through its caption. A gateway branch whose own side corner the caption covers leaves by the corner facing along the flow, sharing it when another branch is there, instead of turning back across the diamond from the opposite corner. When no other face is clear an edge still crosses the caption, and saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

Edges no longer run along another edge's end

A route keeps at least 8 px from the first or last stretch of another edge wherever a clear route allows, where before the two could run side by side and draw as one line, as two message flows climbing one column into nodes stacked in it did. A branch of a shared trunk whose target sits straight ahead of the trunk runs straight on through the split while the others branch off it, instead of stepping off its line and back. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

improved

diagram_find results tell similar diagrams apart

Each result from MCP diagram_find and GET /v1/diagrams carries the diagram's diagramFamily and brief, the first line of its brief cut at a word to 120 characters, so an agent can tell two similarly titled diagrams apart without opening them. A search lists title matches first, then key matches, then brief matches, each latest version first, where before every match was listed latest version first. Both fields are left out when a diagram has none, and diagram_get returns the whole brief.

improved

Render results give the image's size and date

diagram_render and GET /v1/diagrams/{id}/render return the image's width and height in pixels, the same for the SVG and the PNG, so a page can make room before it loads. They also return savedAt, when the version was saved, and latest, the diagram's latest revision. title comes for a caller in the diagram's workspace or a published diagram, never for anyone else through a share link, and is left out when there is none.

improved

Agents look for a diagram before making one

The MCP server's instructions and the Zindex skill tell an agent to start with diagram_find, or the ID in a repository's .zindex.json, and to edit the diagram the person means rather than make a copy, creating one only when they want another diagram. They also ask for a key and a brief on every diagram, with the brief's first line saying what it shows, since that line is what diagram_find shows the next agent.

improved

Agents enlarge a placed node where it stands

The MCP server's instructions, the Zindex skill, the operations reference and the ops schema tell an agent making a placed node bigger to give resizeElement the side its flow enters, anchor: "left" in a left-to-right diagram (the default) or "top" in a top-to-bottom one or an org chart, so the node keeps its center across the flow and edges that ran straight into it still do. Without anchor the node still grows right and down from its top-left corner, as before, moving its center by half the growth, and an edge into a diamond's corner or a BPMN sequence flow can take a step.

improved

Straight edges point at the nodes they join

An edge with router: "straight" is the line between its two nodes' centers, ending exactly where it crosses each node's outline, whether ellipse, diamond, cylinder, hexagon, cloud, pill or box, as other diagramming tools draw it, where before it ran between the middles of two box sides and could enter a node's bottom while pointing past it. Where the nodes line up, the line still runs exactly horizontal or vertical, an end that would meet an outline at a glancing angle moves along it, around a box's corner, to where the line meets it steeply, so the arrowhead stays whole, and two straight edges between one pair of nodes run as parallel lines 12 px apart. Saved versions keep their stored images, and a diagram's next version, a preview and ?engine=current draw the new lines.

improved

Curved edges are one smooth arc between outlines

An edge with router: "curved" is one smooth arc the engine draws itself, ending on both nodes' outlines and bowing to whichever side, and by however much, keeps it clear of other nodes where it can, where before the curve always bowed to the left of its direction of travel, so it could bow into its own nodes, hide its arrowhead under its target or leave the image. Curves running both ways between one pair of nodes form a lens, a curved self-loop is rounded, and a curve's label sits beside or on the curve it draws. A curved edge's computed path carries control, the curve's control point, and saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new curves.

improved

New nodes make room for their edge labels

When an edit adds a node next to saved ones and its edge label needs more room than the column (or, top to bottom, the row) it joins, so the label would cross a frame's border or cover a node or arrowhead, that column and everything beyond it move along just far enough for the label to sit clear, carrying the frames past it and growing a frame across it. Only nodes an edit adds do this: nodes the engine places on a first save or for autoLayout with elements move nothing, so coordinates you write on a first save are kept, and a pool, lane or fragment with its own layout across the column keeps everything where it is. The save's changed and a compare list the nodes that moved, so an edit can report saved nodes it didn't name as changed.

improved

Thumbnails of tall diagrams show their top

A diagram more than about 1.1 times as tall as it is wide fills its thumbnail's width and shows its top: the start of a top-down flow, a tree's root or an ER model's first tables. Before, it was fitted whole into a thin strip too small to read. This holds on the Diagrams list and for GET /v1/images/{id}@{revision}.thumbnail.png, and wide and ordinary diagrams are still fitted whole.

improved

Unverified accounts are deleted after 24 hours

An account whose email isn't verified is deleted once 24 hours have passed since sign-up, where before it was kept for 7 days, so if someone signs up with your address, you can sign up with it yourself a day later. The privacy policy and terms, updated 2026-10-28, and the email with the verification code say the same. GitHub and Google sign-ups are verified at once, so only email and password sign-ups are affected.

improved

Pro checkout needs a verified email address

Start Pro, on the pricing page or the Billing page, opens checkout only once the account's email is verified, as creating a team workspace already did, and before then the page says to verify it at zindex.ai/verify-email. GitHub and Google sign-ups are verified at once, and an email sign-up that began at Start Pro already verifies the address before checkout opens, so only an account left unverified meets this.

fixed

Cursor signs you in when it connects

Cursor opens the Zindex sign-in page as it connects to https://api.zindex.ai/mcp, where before it added the server without ever asking you to sign in, so saving and reading diagrams failed with "Unauthorized". Cursor starts sign-in only when it connects, not when a tool first needs your account, so in Cursor diagram_validate and unsaved previews need an account too. Other apps, and Cursor set up with an API key, are unchanged.

fixed

Start Pro with GitHub or Google opens checkout

Signing up with GitHub or Google from Start Pro on the pricing page goes on to checkout, monthly or yearly as you chose, where before it brought you back to the pricing page signed in, to press Start Pro again. If the GitHub or Google sign-up fails, trying again from the sign-in page still goes on to checkout.

fixed

Sign-in errors name the provider you used

An error from signing in or up with GitHub or Google, such as an expired sign-in or an unverified email, names the provider you used, where before a GitHub sign-in's errors said Google. The message shows once, so it no longer comes back when you reload the sign-in page, and it goes when the email form shows an error of its own.

fixed

ER diagram: Relationships keep to their column rows

ER diagram: a relationship anchored to column rows runs from row to row, where before a long route around other tables could become a loop under the diagram into both tables' bottom edges. Relationships into one row from the same side, drawn alike, share one line into it, with one glyph at the row and a dot where each joins, instead of a line each a few pixels apart. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

fixed

Edges that share a route draw one line

Edges that come the same way to one point, as a fan-out does, stay on one line until their routes part, where before one could be pushed 8 px aside into a parallel line. A stretch several edges drawn alike share, such as a fan-out's stub or a joined trunk, is stroked once, so it keeps a single line's weight instead of darkening with each edge on it. The homepage's examples, the example pages' pictures and the default Zindex card are redrawn with the current engine.

fixed

Sequence diagram: Messages cross bars, fragment tabs clear lifelines

Sequence diagram: messages draw over the lifelines, actors and activation bars they cross, whatever their zIndex, so an async message across a bar no longer reads as ending at it. A combined fragment reaches past the lifelines its messages touch, so its operator tab sits beside the first lifeline and its bar instead of on top of them. Only fragment boxes widen, and a guard can cross the first lifeline, drawn over it on its own backdrop.

fixed

Architecture: Auto-grouped bands no longer overlap

Architecture: with extensions.autoGroup: "byNodeType", the engine lays each band's nodes out as one region apart from the other bands, as it does a frame's. A band whose nodes span several columns no longer wraps another type's node or overlaps the next band. Where kept positions would make bands overlap, as a saved diagram's or an added node's can, the diagram draws without bands until autoLayout without elements lays them out again.

fixed

Edges keep clear of captions and rounded corners

An edge keeps out of its own event's or gateway's caption except where it meets the shape, and edge labels keep a line's height, 16 px, from those captions. An arrowhead or other end glyph on a node with rounded corners meets its side at least 14 px from the corners, so it sits along the straight part instead of half beside the curve. A straight or curved edge's end that meets an orthogonal edge's at one point moves along the outline, so the orthogonal flow keeps its place.

fixed

Group titles no longer draw stray text

A group element's title is written into the SVG as the group's <title> and isn't drawn, where before it was drawn as text above the canvas and never showed. For a box with a visible title around elements, use a frame.

fixed

Workflow: A pool's box is a minimum

Workflow: a pool's or lane's layout is a minimum, so a pool grows right and down from its corner to hold its lanes and their tasks, where before a pool given a box split it into equal lanes, and tasks without a position landed in another lane or outside the pool with NODE_IN_WRONG_FRAME and NODE_OUTSIDE_FRAME. Pools and lanes no longer get NODE_OUTSIDE_FRAME, a lane's x and y are ignored since its pool places it, and pools given overlapping boxes no longer draw one over the other. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw each pool grown to hold its tasks.

fixed

Workflow: A pool's lane direction sets the flow

Workflow: lanes always run along the flow and stack across it, so when layoutStrategy.direction runs along a pool's lanes, such as laneDirection: "horizontal" (lanes side by side) with LR, the pool's laneDirection wins and the diagram flows across its lanes, and a save or edit returns the warning POOL_FLOW_DIRECTION for that pool, whose fix names the laneDirection that keeps the flow you asked for, or says to leave it out. Before, each lane's tasks slid along the flow as a block, so steps could run out of order with flows pointing backward. Without laneDirection, lanes stack across the scene's direction; the first pool in document order with one sets the flow for the whole diagram, and lanes along the flow, such as phases or milestones, aren't supported.

fixed

Wrapped text stays inside a diamond

A diamond's wrapped label lies in the largest rectangle inside the diamond, half its width and height, so its lines stay clear of the sloped sides, where before they could run past them, and with "auto" the diamond grows to hold them. A diamond whose label fits on one line keeps its size, and saved versions keep their stored images.

fixed

Node labels follow align and verticalAlign

A node's label is placed by its textStyle.align and verticalAlign whether or not it wraps, at the center or half the font size in from a side, the top or the bottom, and the node's icon moves with it, where before a label that didn't wrap was always centered. Centered labels draw as before, and nodes that draw their own text, such as org chart cards, UI flow screens, ER tables with columns and sequence nodes, still ignore both. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current place the labels.

fixed

Text elements follow their vertical alignment and weight

A text element is placed by its textStyle.verticalAlign, at the top of its box when it sets none, and drawn in its style.fontWeight, where before text that didn't wrap sat in the middle of its box whatever it asked for, and every text element was drawn in regular weight. Text of several lines is placed as one block at its alignment. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the text where it asks to be.

fixed

Unstarring a diagram in Starred offers Undo

In the Diagrams list's Starred view, unstarring a diagram turns its card or row, in place, into a note that it was removed from Starred, with an Undo button that stars it again and puts it back. Before, the card or row stayed unchanged until the page loaded again, which now leaves the diagram out.

fixed

Edge labels read as their own edge's

Where edges share a trunk and then fork, as a fan-out from one node or a fan-in to one does, a label sits on the stretch only its own edge draws, past the fork, so it reads as that edge's and not a sibling's. Other edges' lines stay at least 12 px from a label, twice the 6 px its own line keeps, no other edge's line runs along the top or bottom of a label beside a vertical line within 30 px, and a label no longer covers its own line at a bend. Where an edge has no such place, as on a short branch past a fork, its label can still sit near another line, as the Layout Engine docs page describes; sequence message labels keep their rows above their arrows, and saved versions keep their stored images.

fixed

Edge labels keep clear of same-label edges

An edge label keeps at least 12 px from another edge's line even when that edge carries the same label text, as two edges from one node both labeled "retry" do, so each label reads as its own edge's. Before, a label could sit beside the line of a sibling that drew its own copy of the same text. Only a fan-in drawn with a single label (EDGE_LABEL_SUPPRESSED_FANIN) leaves its edges' lines out, since that one label names them all, and saved versions keep their stored images.

fixed

Edge labels keep clear of arrowheads in every theme

An edge label's backdrop keeps clear of the arrowheads and ER cardinality glyphs at both ends of every edge, sized for the thickest line any theme draws (the sketch theme's 2 px) or for an edge's own wider strokeWidth, so a place that is clear in one theme is clear in all of them. Before, a label could cover part of an arrowhead or a crow's foot, most often in the sketch theme. Where an edge has no clear place, its label still keeps off nodes, other labels and other edges' lines first and then covers as little of a glyph as it can, and saved versions keep their stored images.

fixed

New layouts leave room between ranks for labels

A new layout spaces its ranks to hold each edge label with its backdrop, the arrowheads or ER glyphs at both ends and both nodes' outlines, and keeps 6 px clear on either side of the border of a frame laid out with its members where it runs between the ranks, so labels stop crossing frame borders in new layouts. A node an edit adds without a position gets the same room when it starts a column of its own, and when it joins a column the saved layout already has, that column keeps its gap unless the new node's edge label needs more room, and then moves along just far enough. Saved nodes otherwise keep their positions, so a label between them can still cross a border where the gap is too narrow, as on a new edge between two saved nodes or a longer label on a saved edge, until autoLayout without elements lays every node out again, moving the nodes you positioned, and saved versions keep their stored images.

fixed

Colored diffs keep arrowheads at their usual size

A visual diff, from Color what changed on a diagram's page or in Present or from a ?from= image, draws a changed edge's arrowheads and ER cardinality glyphs at their usual size, where before they grew with the heavier line to about twice their size, could wrap around a bend and were cut into by a nearby label's backdrop. A stretch of line several edges share, such as the stub a node's edges leave by or a fan-in's trunk, is colored only when it changed, so a new edge's color starts where it forks off and a shared line that is still there isn't colored, and what didn't change still fades, but its text keeps 4.5:1 contrast and its lines 3:1. A removed edge's ghost is dotted, unless its edge type is already drawn dashed, so a removed call no longer looks like a reply, and the legend lists only the kinds of change in the picture (Added, Removed, Changed, or No changes) and names the version the picture is compared with, as in "Since version 2".

fixed

Orthogonal edges stay level or vertical

An orthogonal edge drawn as one straight line between two boxes whose facing sides overlap only a little stays level or vertical when the engine spreads the edges on those sides apart, where before it could be pushed past the overlap and drawn slanted to a box's corner, and Zindex's routing tests check that no orthogonal route slants. A line running beside the stretch where another edge leaves or enters its node, as an edge leaving a decision's side corner can, keeps the usual 8 px from it where there is room, where before the two could run 4 px apart and read as one line. Saved versions keep their stored images, while a diagram's next version, a preview and ?engine=current draw the new routes.

fixed

Straight and curved back-edges keep their shape

A straight or curved edge that runs back against the direction of an auto-laid-out flow stays one line or one curve, where before the engine redrew it as an orthogonal loop. For a loop back, curved bows around the nodes between its ends where it can, while a straight one runs through them and gets EDGE_CROSSES_NODE.

fixed

Edges meet a pill's rounded ends

A pill shape's rounded ends are its outline, so an edge that ends off the pill's middle meets the curve instead of stopping up to 4 px short of it, beside the cap. A pill taller than it is wide is an ellipse, as it is drawn, and saved versions keep their stored images.

fixed

@zindexai/mcp 1.0.7 starts when Zindex refuses its key

@zindexai/mcp 1.0.7, the package Claude Desktop, Claude Code, Cursor and other hosts run with npx -y @zindexai/mcp, starts even when Zindex refuses the ZINDEX_API_KEY in its config because the key is revoked, expired or mistyped. The host lists the Zindex tools, every tool call answers AUTH_REQUIRED saying Zindex refused the key and to create a new one on the API keys page in the dashboard and restart, and the start-up line in the host's MCP log says the same. Before, the package stopped at start and the host showed only "Connection closed", with no hint of why.

fixed

ChatGPT connects through Plugins, without developer mode

The connect page and the tools guide give ChatGPT's current steps: on the web, open Plugins at chatgpt.com/plugins, choose Add custom MCP server, enter https://api.zindex.ai/mcp with OAuth sign-in, and create it as a plugin, with no developer mode to turn on. The tools guide also says that claude.ai and the Claude apps can send a Zindex API key instead of signing in when the organization has Claude's request headers, a beta.

fixed

Versions saved from VS Code name their client

A version saved from VS Code records its client as Visual Studio Code with its version, such as Visual Studio Code/1.105.1, where it recorded only Visual, and so does any client whose User-Agent names a product with a space in it. The client shows in the diagram's History, in diagram_history and in the outline's recent: line. Versions already saved keep the client they recorded.

fixed

Billing errors, READ_ONLY and outlines match the docs

Every error from the /v1/billing endpoints the dashboard calls carries a code, as the workspace docs say, with every status as before, among them AUTH_REQUIRED when you aren't signed in, BILLING_UNAVAILABLE when billing can't be reached, and CHECKOUT_REQUIRED, SUBSCRIPTION_EXISTS, NO_SUBSCRIPTION and STALE_BILLING_STATE; a 404 NOT_FOUND carries a reason (account, workspace or checkout). READ_ONLY's message says what a read key or a viewer can't do as the docs do: change or share diagrams, or report a problem with them. diagram_get's outline indents a brief's later lines two spaces, so they no longer read as header fields of their own between brief: and recent:.

fixed

Docs list every case that moves a saved node

Every page that says when a later edit moves a node from its saved position, from the tools and API guides to the schema reference and the MCP instructions, lists all six cases: a node that grows, a new node's edge label, a new task in a pool, a node moved to another lane, a frame added or deleted around placed nodes, and a new order constraint. The rendering reference lists every warning-level code, POOL_FLOW_DIRECTION and DRAW_TIMEOUT among them, and gives DRAW_TIMEOUT its own section.

fixed

Closing an account says to download invoices first

When your personal workspace has a subscription, the Delete account dialog on the Account settings page, the confirmation page and the email with the confirmation link say to download any invoices you need first, from Manage subscription on its Billing page, since its billing history can't be opened once the account is closed. Deleting a team workspace already said the same.

fixed

The privacy policy says where the API runs

The privacy policy, updated 2026-10-29, says the API runs in Ashburn, Virginia, and Chicago, and that Zindex processes personal data in the United States, apart from the copies of public pages and images Cloudflare's edge cache keeps near their visitors for up to an hour, and billing data Stripe handles under its own privacy policy. It also says scene content is used only to answer requests for it: in its workspace, through a public or share link someone with write access made, and in a problem report or feedback sent with it.

fixed

Emails write dates month first

Dates in Zindex's emails, such as a team invitation's expiry and the deletion date in an unused account's warnings, read month first, as in November 8, 2026, where they read 8 November 2026.

fixed

Sign-in forms match the rest of the site

The sign-in, sign-up, password reset, contact and email verification pages draw their fields, labels, error lines and password rules as the site's other forms do, the email verification code boxes are one size at every screen width, and autofill keeps the text cursor visible in every form. On the new team workspace page, Pay monthly and Pay yearly keep one width, so "Opening checkout…" no longer shifts the other button, and on a phone each takes the full row.

fixed

The homepage's payments example labels its cache edge again

The payments platform example on the homepage labels the edge from Checkout service to Idempotency cache "check key" again in its second version, on the stretch only that edge draws, past its fork. Every version's picture, and each colored picture of what changed, is drawn with these label fixes.

fixed

The homepage's payments example clears the Workers frame

In the second version of the payments platform example on the homepage, "HTTPS POST" sits clear of the Workers frame's border and off its arrowhead, with Email provider and Data warehouse moved along to make room for it. The second version's list of changes, and its colored picture of what changed since the first, mark those two as changed, since they moved.

fixed

Two examples' descriptions match their diagrams

The PII data-flow example describes the one trust boundary its diagram draws, a dashed frame around the external KYC vendor, where before it described three. The request flow example says its after diagram reads the cart from the Redis cache instead of calling the Cart Service, where before it listed Cart among the handler's collaborators.

breaking

Computed edge paths list junctions for joined edges

In a diagram's computed output, the edge whose last leg carries joined edges into a shared side lists in junctions every point where their lines branch, each drawn as a dot, replacing fanInPoint and isFanInTrunkOwner, and keepFeederArrowhead is removed, so a document whose computed.edgePaths carries any of the three is refused. The info diagnostic EDGE_LABEL_FANIN_BUNDLED, which reported a merge the drawing didn't show, is removed, and edges from different sources that share a label keep their own lines, labels and arrowheads. A document sent without computed, as an authored one is, is unaffected, so only code that reads computed edge paths may need to change.

breaking

Edge SVG groups can hold part of a line

A stretch that edges drawn alike share is drawn once, in one edge's data-zindex-element group, so another edge's group can hold only part of its line, as a colored diff's groups already could. Only code that selects an edge's group to find or restyle its whole line may need to change, and computed paths keep every edge's whole route.