new
Pro replaces the paid plans
Pro costs $5 a month or $48 a year, bought from the pricing page or the Billing settings page, and replaces the Builder, Growth and Enterprise plans. It has no diagram limit under fair use, 25,000 updates and 50,000 renders a month, 10 API keys, 120 requests a minute and private share links.
new
Team plan opens as a beta
Team is open to every account, labeled Beta and billed from the start, at $8 per paid seat a month or $80 a year with at least 2 seats: owners and members are seats, and viewers and agents are free. Start Team on the pricing page creates a team workspace and opens its Billing page, and each seat adds Pro's monthly updates, renders and API keys to limits the whole workspace shares. Seats follow roles, a seat count changed outside Zindex is put back at once, and Team's features may change during the beta.
new
Team workspaces for people and their agents
Each account has a personal workspace and can create or join team workspaces as an owner, member or viewer, picked with a switcher in the dashboard, and every API key and connected app works in one workspace. Everyone in a team workspace can read all its diagrams, and owners invite people by email, an invite link or an email domain, can turn off public sharing, and see and revoke everyone's keys and apps. A team's Diagrams page points to the diagrams in your personal workspace, which you can move into the team with every version, and the new Workspaces docs page covers roles, joining and seats.
new
Owners choose the role people join with
Owners choose whether people who join through an email domain or the invite link come in as viewers, the default, or as members, and picking Member shows what each one costs: a paid seat at the team's price, or "No charge" when Zindex covers the plan. A join request emails every owner what approving the requester as a member would cost, and on the Members page it has two buttons, Approve as viewer and Approve as member, with the one matching the domain's setting highlighted. When someone joins a paying team as a member without an owner's approval, every owner gets an email with their name, how they joined and the seat's price.
new
Cards leave with a team's payer
When the owner who pays for a team workspace stops being an owner, whether they leave, are removed, become a member or viewer or close their account, its cards are removed and its billing contact moves to the remaining owner who joined first, and the confirmation says so first. Every remaining owner gets an email asking them to add a card before the renewal, or to resume a plan set to cancel and add one, and Billing shows them "No card on file" with Add a card. Until a card is added, a new seat is charged with the renewal instead of at once, and the privacy policy, updated 2026-10-15, covers the change.
new
Read-only API keys and connected apps
An API key can be read-only, chosen with Read only on the dashboard's API keys page or scope: "read" on POST /v1/auth/keys, and the sign-in page for a connected app asks whether it may change diagrams or only read them. Read access finds, reads, renders and compares diagrams and downloads a diagram's history, while anything that changes or shares one, or reports a problem with it, answers 403 READ_ONLY, or the READ_ONLY tool error over MCP. A client that asks for the OAuth scope diagrams:read alone gets read access, and existing keys and connections keep read and write.
new
Each version records who saved it
History over REST and diagram_history gives each version's by, the person's name or else their email, and credential, which is session, api_key or oauth but never which key, and a compare adds authors, each person and client that saved a version in its range. Exports add by and actor to every version, and versions saved before 2026-10-14 name their diagram's owner. Saves made on zindex.ai record zindex.ai as their client instead of the browser's.
new
Privacy policy and terms cover team workspaces
The terms and privacy policy, updated 2026-10-14, say that a team workspace's content belongs to the organization its owners act for, which controls it while Zindex processes it, and that people in a team workspace see your name, email, role and the versions you save there. Closing your account, refused while you are the last owner of a team workspace with other people in it, deletes your personal workspace and leaves the diagrams you made in team workspaces there under a former member. Deleting a team workspace cancels its subscription at once, with no refund for the rest of the period.
new
Permanent links to every version of published diagrams
A published diagram's page is zindex.ai/d/<id>, and zindex.ai/d/<id>@<revision> shows one version with its date and the dates of every version, and serves its SVG and PNG at the same address with .svg or .png. For a published diagram, diagram_render and GET /v1/diagrams/:id/render return these links in url, svg and png with no expiry, other diagrams' links stay signed for 7 days, and GET /v1/public/diagrams/:ref reads a version link without a key. Browsers and other caches may keep an exact version's image for a year and anything that follows the latest version for a minute, and zindex.ai/s/<id> pages redirect to /d/<id>.
new
Every image carries its version's address
Images of versions saved from 2026-10-15 on print their address and date in the bottom-right corner, such as zindex.ai/d/dgm_7Q2kD9a1B2c3@5 · 2026-04-01, on every plan alike, with no separate zindex.ai mark. A teammate holding a screenshot can type the address in, an agent shown the image can open that version, and in an SVG opened in a browser the line is a link. A private diagram's address opens that version for anyone signed in with access, offering to switch workspace if needed. Everyone else sees "diagram not available".
new
Private share links on Pro
POST /v1/diagrams/:id/links, diagram_share with link (after asking the person), or Create share link on the dashboard's diagram page makes a link at zindex.ai/d/<token> that shows the diagram's images and the dates of its versions, never its scene, while the diagram stays private. Revoking a link (DELETE /v1/diagrams/:id/links/:linkId, or unlink for all of a diagram's links) stops it at once, deleting the diagram stops its links until it is restored, and only the visitor's own browser may cache what a link shows. Making a link needs Pro and answers 403 FEATURE_UNAVAILABLE on Free, while links made on Pro keep working after the account leaves Pro, and listing and revoking work on every plan.
new
Share link recipients can report the diagram
A share link's page offers Report this diagram, as a published diagram's page does, and POST /v1/public/diagrams/:ref/report takes either page's ref (a published diagram's ID or a share link's token) with @<revision> for the version seen, keeps a report only for what that page shows, and answers the same 200 whether or not it keeps one. Moderators see what the reported link shows and no more, and if a diagram shared only with links breaks the content policy, Zindex revokes that link only and emails the reason to the person who made it, or to a team workspace's owners if that person has left it. Taking down a published diagram also revokes every share link to it, as the privacy policy and section 5 of the terms, "Published and shared scenes", updated 2026-10-16, say.
new
Diagram links preview in Slack, Teams and iMessage
A published diagram's link or a share link pasted into Slack, Teams or iMessage previews the diagram itself, not just a title. A preview already posted, like a downloaded copy, stays after the link is revoked. On a page that follows the latest version, such as zindex.ai/d/<id> or a share link, the SVG and PNG links follow the latest too. A version's page at @<revision> keeps that version.
new
Ready-made links for docs, issues and READMEs
A new panel under Share on a diagram's dashboard page asks where the diagram is going (GitHub, Google Docs, Notion or Confluence, a web page, Slack or email) and writes the link that place takes, with the one step that puts it there. It links this version, or the latest of a published or link-shared diagram, and offers PNG or SVG only where the place takes both, giving Google Docs a PNG of this version because Docs keeps its own copy. The panel shows the text before you copy it and remembers your choices, its Copy image button puts the PNG itself on the clipboard, and a new section of the REST API docs explains which link suits each place, and its limits.
new
Download every diagram in one file
GET /v1/account/export, or Download in the Export section of the Account settings page, returns every diagram in the account in one JSON Lines file, each diagram written as its own history export, with every version's scene, stored image and hashes. The last line counts the diagrams, so a file cut short shows it, and deleted diagrams are left out. It takes an API key or the website session on every plan; there is no MCP tool for it.
new
Each version carries a chain hash
Saves, edits, renders, view=dsp and view=mermaid reads and exports return chain next to hash, and the history lists each version's chain: a SHA-256 over the version's content hash, the hash of its stored image and the chain hash of the version before it, so if a later export checks out and shows the same chain for a version, no scene or image up to that version has changed. The Concepts page in the docs shows how to check an export without Zindex. Every version saved before chain hashes arrived got its chain hash on 2026-10-13.
new
The playground opens on Mermaid
Paste a Mermaid flowchart or erDiagram at /playground, or try a sample, to see how Zindex draws it, with the import's notes under the picture and a syntax error's line and column. Save diagram creates the diagram from the Mermaid; signed out, the text is kept through sign-up and saved when you come back. Scene JSON is the second tab, where the examples and saved diagrams open.
new
A new homepage and a Connect page
The homepage steps through real versions of Zindex's own architecture and data model diagrams, each with its date and message. zindex.ai/connect has the MCP server's address and the steps for Claude, Claude Code, Cursor, VS Code, ChatGPT and other apps, and Connect is first in the navigation. The agent view has a Human view button at every width, and the login and signup forms work after a switch to the agent view and back.
new
Agents can send feedback to Zindex
diagram_share with feedback, or POST /v1/feedback, sends Zindex staff a bug, a docs problem or a request, with what was expected and what was observed, and optionally a diagram version or a scene or Mermaid snippet, anonymized unless anonymize is false. The person confirms each item before it is sent, an account can send 20 in any 24 hours, and GET /v1/feedback and the dashboard's Your feedback show each item's status and any note from staff.
new
createNode and updateNode take textStyle
Edits set how a node's label wraps and aligns as a new scene does: createNode takes textStyle, with defaults for the fields it leaves out, and updateNode changes only the fields it sends. Both refused it before and pointed at setText.
new
ICON_NOT_FOUND names icons Zindex doesn't have
Validate, create and edit warn with ICON_NOT_FOUND when a node's icon is a family:id key with no built-in icon, such as a misspelled Lucide name or a gcp: or azure: key, where the node drew an empty slot without a warning. The fix suggests the closest keys, as in "did you mean 'lucide:database'?", or lists the icon families there are, and the node keeps its icon's room.
new
TEXT_OVERFLOW flags frames narrower than their title
A frame with its own layout whose title runs past its right border gets TEXT_OVERFLOW, with the width the title needs in requiredWidth and the frame's in availableWidth. Its fix is to widen the frame with resizeElement, shorten the title or drop the frame's layout; pools, lanes and sequence fragments aren't checked.
new
Shared diagrams get their own tab icon
Images and diagram pages at zindex.ai/d/, published or shared by link, and 7-day image links show the Zindex logo on a solid green tile in the browser tab, so they're easy to spot among open tabs and to tell from the app's amber icon.
improved
Layout runs up to ten times faster
A 40-node diagram with 960 edges, and 263 nodes that each link to the next two, lay out in about a tenth of the time they took before, and a hub with 499 leaves and a 520-edge flowchart in well under half. The speed-up changes no drawing: the faster passes choose exactly what the slower ones did.
improved
Frames are laid out with their members
In architecture, workflow, entity-relationship and UI flow diagrams, and diagrams with no family, laying out every node lays out each frame written without a layout together with its members: they sit together, no other node or frame comes inside, and the frame gets room for its padding and title. Before, nodes were placed without regard to frames and each frame was fitted round its members afterwards, so frames could overlap or take in nodes that weren't theirs. Pools, lanes, sequence fragments and frames inside a pool are laid out as before, and org charts and networks still place nodes without regard to frames.
improved
Edits keep frames apart and members inside
On a saved diagram, a new member of a placed frame goes inside it, a new frame's nodes are placed together as one block clear of the other frames, and a new node joined to a frame by an edge goes a rank before or after the frame. When an edit adds or deletes a frame round placed nodes and keeping them where they are would leave frames overlapping, or a frame holding a node that isn't its own, those nodes are placed again together and the edit's changed lists them.
improved
Nodes that grow push their neighbors aside
A saved node drawn bigger than its box, because an edit changed its content, its autoSize: "content" label needs more height or resizeElement enlarged it, pushes the nodes and frames in its way just far enough to keep the gap they had, up to the drawing's spacing, where before it grew in place and could overlap them with NODE_OVERLAP_DETECTED. The edit's changed lists what moved, and the first edit of a diagram saved earlier can move nodes that the engine measures bigger than it did. Frames grow round pushed members, an org chart card brings its reports along, and a pool with its own layout doesn't grow, so a node pushed past its lane gets NODE_OUTSIDE_FRAME.
improved
New order constraints take effect at once
An order constraint added to a saved diagram places its two nodes again in the same edit when they don't already stand in that order, as autoLayout with those two nodes does, and leaves every node where it is when they do. Before, it took effect only when the same edit also sent autoLayout without elements, which lays out every node. The order is judged in the direction the diagram is laid out, top to bottom for an org chart.
improved
Edges route around frames and clear of nodes
An edge to a frame ends at the frame's border, and other edges keep 16 px off the borders of frames they pass, go round a frame they don't need to cross, rarely run under a frame's title, and stay inside a frame when both their ends are in it. Routes also avoid running within 16 px of the side of a node they pass, and an end leg that would turn on a frame's border turns 16 px past it.
improved
Edges keep apart where they meet a node
An edge no longer leaves a node at the spot where another arrives: it takes its own port or another side, and the ends on one side keep 12 px apart. Every arrowhead and crow's foot sits on a straight end leg longer than itself in every theme, so none lands on a bend, and two edges that meet one line from opposite sides are moved apart instead of reading as one line passing through.
improved
Parallel edges cross less and keep their spacing
Edges sharing a corridor are ordered by the side they turn to, so they cross each other less, and where one still crosses another's track the tracks spread up to 24 px apart so the crossing sits clear of the bend. Two edges through a gap too narrow for full spacing share it evenly instead of running 2 px apart, edges drawn differently no longer share one trunk, and a reversed fan of up to sixteen edges untangles fully. A straight edge into a database cylinder's side lands clear of its lid, and a loop's leg stepping off a label keeps clear of other edges and its own node.
improved
Event labels move clear of their flow
In top-to-bottom and bottom-to-top flows, a labeled BPMN event or gateway, or a UI flow entry or exit point, whose edges all use one end of it gets its label at the other end, so its edge no longer runs through its own label. Edges dock on such a shape as it is drawn, where in a box pinned wider than the shape they could start or end inside it.
improved
Rate limits before sign-in suit hosted clients
Without a key, requests count against the address they came from, never X-Forwarded-For, and the published addresses of Claude and ChatGPT, which carry every user of those platforms until they sign in, get 600 a minute instead of 40; each ChatGPT user is also held to 30 plus a burst of 10. Keyless validation and previews also share a ceiling across all callers. A 429 from these limits carries code: "RATE_LIMITED" and retryAfter, in a JSON-RPC error's data on /mcp.
improved
@zindexai/mcp 1.0.1 reports refusals as tool errors
@zindexai/mcp 1.0.1 hands a refused call to the host as a tool error with its code, message and retryAfter in seconds: RATE_LIMITED for a rate limit, BUSY when every drawing worker is busy and DRAW_TIMEOUT for a drawing past its time limit. 1.0.0 did this only for a missing key, so an agent got a transport error it couldn't act on.
improved
Mermaid imports keep titles and larger flowcharts
A title in the front matter becomes the diagram's title on create, and a Mermaid edit that changes it changes the title, as view=mermaid writes it there; text without a title leaves the title as it is. Flowcharts with more than 500 edges import, where they failed with PARSE_ERROR, up to the scene limit of 1,000 elements, edges included, past which create, validate, preview and edits answer VALIDATION_FAILED with CONTENT_LIMIT_EXCEEDED.
improved
Billing shows the plans each workspace can have
A personal workspace's Billing page shows Team's monthly and yearly prices: choosing one asks you to name the new team workspace and goes straight to Stripe Checkout, and a link creates the team without paying. A team's Billing page says Pro is for your personal workspace, with a Switch to Personal button. A Pro or Team plan that Zindex gives you without a subscription shows as free on the Billing and Members pages: no price, no billed seats and no Switch to Free button.
improved
Wide diagrams fit their box, with Actual size
On a diagram's dashboard page and its page at zindex.ai/d/, published or shared by link, a diagram wider than its box is shrunk to fit and a narrower one keeps its size, where a wide one was drawn at its own size and its right side could sit out of view. While the picture is shrunk, Actual size above it, or a click on the picture, shows it at its own size in a box that scrolls, and back. On the zindex.ai/d/ page, the SVG and PNG links sit above the picture after "Open full size:" and open in a new tab.
improved
One control sets who can see a diagram
A diagram's dashboard page has a Share section under the image, with one setting for who can see it: only you (only the team, in a team workspace), anyone with the link, or public. Extra links and revoking sit under Share links, and on a public diagram the section also points to the diagram's public page and its scene JSON. On a private diagram, the Copy 7-day image link button copies a PNG link that stops working after 7 days, where the Copy image link button quietly copied a 7-day SVG link, and Download has its own section, ahead of History.
improved
Docs compare the three sharing choices
A Sharing section in the Concepts docs compares Only you, Anyone with the link and Public side by side. It covers who can see a diagram, what a visitor sees, whether the scene JSON is readable, which plans have each choice and what each change does. On a diagram's dashboard page, Compare the choices beside Who can see it links to it.
improved
Shorter docs, in fourteen pages
The docs are six pages, Quickstart, Concepts, MCP tools, Mermaid scope, REST API and Limits, plus eight reference pages, where there were seventeen. Mermaid scope lists what imports, what is left out with a warning and what is refused, and Limits gives each plan's limits, what counts, the answer at a limit, the rate limits and the drawing time limits. Every old docs URL redirects to the page that replaced it.
fixed
Themes no longer change text size
Themes set colors, strokes, corners and roughness, and text is drawn at the size the layout measured, where each theme set a text size of its own: a sketch card could wrap onto a line it had no room for and report TEXT_OVERFLOW, and ER column names were drawn bigger than the rows their edges dock at. Under blueprint and sketch, node labels are 14 px like the other themes', and ER columns and sequence lifelines and notes are 13 px under every theme. An explicit style.fontSize still applies.
fixed
Edge labels no longer run off the canvas
The canvas grows to hold every edge label with its background, keeping the margin the drawing has on the opposite side, where a multi-line label above the top row, a long label beside a vertical edge or a sequence reply's label could be cut off at the canvas edge.
fixed
PNG images draw start markers the right way
In PNGs, a marker at an edge's start, such as an ER crow's foot at a foreign key leaving a table's right side, was drawn turned around and hidden inside its own node; every PNG, including those of versions saved earlier, draws it facing back along its edge. New SVGs draw start markers without auto-start-reverse, which some SVG viewers don't support.
fixed
ER diagram: Column names no longer overlap types
ER diagram: a table is as wide as its widest row as drawn, with the key badges, the name, a 24 px gap and the type, so a long column name no longer runs into its type. Saved tables whose rows need more width grow on the diagram's next edit and push the nodes beside them.
fixed
ER diagram: forceLabel keeps the column anchor
ER diagram: an edge whose label names a column and sets style.forceLabel draws the label and still anchors at that column, with its cardinality inferred from it. Before, it docked at the table's primary key and read as one-to-one.
fixed
visible: false hides an element and its edges
An element with visible: false isn't drawn, nor is an edge with an end at it, and a hidden frame or group hides only its box and title while its members stay drawn. A hidden element keeps its space in the layout. Before, every element was drawn whatever its visible flag said.
fixed
Edges to frames and groups are valid
Validate and every save no longer return EDGE_ENDPOINT_KIND_INVALID for an edge that starts or ends at a frame or a group, which the docs and the router accepted; a guide, a text element or an image still can't be an edge's end.
fixed
Icons sit above wrapped labels
A node whose label wraps draws its icon above the lines, 4 px from them and on the side the text is aligned to, where the icon sat across the lines, and autoSize: "content" and TEXT_OVERFLOW count its height. Wrapped lines in an object-storage bucket stay inside its tapered sides, updateNode with icon: null removes a node's icon, and a gcp: or azure: key keeps the node's box instead of leaving a floating label.
fixed
Node dashes, label colors and palette accents apply
style.dash dashes a node's outline as it does an edge's, an edge's label takes the edge's textColor instead of its stroke color, and palette.accentPrimary colors the nodes that set no accentColor. Accent-colored text such as ER key badges keeps a shade that reads on any accent, and named colors such as navy count when text picks a contrasting color; palette.accentSecondary has no effect, as the schemas say.
fixed
Frame titles and labels sit on matching backgrounds
A frame's title sits on the frame's own fill, in a color that reads on it, instead of on a tab of the canvas color, and edge labels and nested frame titles inside a filled frame take the color under them instead of showing a pale patch.
fixed
Sequence diagram: Notes draw in their own colors
Sequence diagram: notes draw with their note fill and border, pale yellow on light themes and readable under every theme, where they took the plain node colors, and a note's text sits 5.8 px lower, clear of the folded corner. Two messages with the same label no longer get LABEL_DUPLICATION_DETECTED, since a sequence diagram repeats labels on purpose.
fixed
Sketch theme keeps arrowheads on their edges
In sketch previews, arrowheads and crow's feet sit on the routed line and face along it, where the hand-drawn wobble turned them up to 14 degrees and brought lines within a fraction of a pixel of another edge's glyph. Edges start and end exactly on their routed points, dashed lines draw in one pass with clean gaps, and lifelines are dashed as in the other themes.
fixed
Private and deleted diagrams' images stop being served
A published diagram's image at zindex.ai/d/ stops being served as soon as the diagram is made private or deleted, and browsers keep the latest version's image for a minute at most. Before, a cached copy could go on being served after the change, and browsers could keep the latest image for hours. Copies people already saved stay with them.
fixed
Retries and no-op edits work at the limit
At the monthly update limit, a retried edit replays its first result and an edit that changes nothing returns noop, over REST and MCP, where both were refused with 403 LIMIT_REACHED although neither makes a version.
fixed
Failed payments mark the subscription past due
A Pro or Team payment that fails marks the subscription past due as soon as Stripe reports the failure, while Stripe retries the card. Before, it was marked past due only when Stripe changed the subscription itself, not when the payment failed.
fixed
Root fields inside scene.scene are moved up
diagramFamily, layoutStrategy and palette written inside scene.scene are moved to the document root and stored there, where a create or a check answered 422 although the docs accepted that form. A field written in both places keeps the root's value.
fixed
Unknown paths and server errors answer JSON
A path with no endpoint answers 404 with code: "NOT_FOUND", and a failure inside Zindex answers 500 with code: "INTERNAL_ERROR", as JSON like every other error, where both were plain text. An MCP tool that fails inside Zindex returns INTERNAL_ERROR too, where it returned INTERNAL.
fixed
Refusals keep their fix in one line
The one-line fix of ORDER_CONSTRAINT_CYCLE, PORTS_NOT_ALLOWED_ON_KIND and the workflow rules says what to do, which sat in a second sentence the fix left out, and EDGE_COLUMN_NOT_FOUND lists the table's columns before its advice, so the list isn't cut. A schema error in a restoreRevision op names the field instead of calling restoreRevision an unknown op, and NODE_OVERLAP_DETECTED says when a node is drawn bigger than its layout to fit its content.
fixed
Team workspace deletion links open the right page
The emailed link that confirms deleting a team workspace opens the team's own confirmation page, where it opened the page for deleting an account and deleted nothing.
fixed
Sign-in messages say what happened
The sign-in page says "Your session expired" only when your session has ended, "You're signed out" when you signed out in another tab, and nothing to someone who was never signed in, where it said "Your session expired" every time. A request with a refused session cookie answers 401 with SESSION_EXPIRED, SESSION_REVOKED or SESSION_INVALID.
fixed
Amber accents keep their color on the website
Amber accents on the examples pages, the API keys page and the plan-limit cards, and the email-verification banner's yellow, show in their own colors, where mixing them toward the page's blue-gray turned them teal.
breaking
Drawings have a time limit
Every layout and drawing runs in a worker with a time limit, 5 seconds for checks and previews without a key or sign-in and 20 seconds for everything else, so one large diagram no longer holds up every other request. A save, edit, preview or render past its limit, or past a worker's memory, answers 422 with code: "DRAW_TIMEOUT" and saves nothing, while validate returns a DRAW_TIMEOUT warning in place of the drawing's warnings. When every worker is busy and the queue is full, a call answers 503 with code: "BUSY" and retryAfter, half the refused drawing's limit in seconds.
breaking
Free keeps 25 diagrams, with higher monthly limits
The Free plan keeps 25 diagrams at once, down from 50, and gets 2,500 updates and 5,000 renders a month and 3 API keys, up from 1,000, 2,500 and 1. A Free account with 25 or more diagrams keeps them all and can edit them, but can't create or restore another until it has fewer than 25.
breaking
Free accounts unused for 12 months are deleted
The privacy policy and terms, updated 2026-10-16, say that a free account unused for 12 months is deleted with its diagrams, after emails 30 and 7 days before the deletion date, and the pricing FAQ and the Limits and Concepts docs say so too. Any use keeps the account for another 12 months, whether using the website signed in, a request with one of its API keys or use by an app connected to it, while other people opening its public diagrams, share links or image links don't count. Pro accounts, people in a team workspace that pays and the last owner of a team workspace with other people in it are never deleted for being unused, and no account can get a first warning before 2027-03-23.
breaking
autoLayout lays frames out with their members
In architecture, workflow, entity-relationship and UI flow diagrams, and diagrams with no family, autoLayout without elements, and a new layoutStrategy or diagramFamily, drop the layout of every frame except a pool, lane or fragment and lay each frame out with its members. Before, a frame with its own layout kept its box and its members stayed where they were, as the members of a pool, lane or fragment with its own layout still do. To place one frame's nodes again and leave the other frames as they are, send autoLayout with those nodes in elements; org charts and networks keep every frame's layout.
breaking
Edits that would save new errors are refused
An edit whose version would have an error the version it started from didn't, such as an edge to a missing port, an element in two frames or a relative layout cycle, is refused with 422 VALIDATION_FAILED and each error's code, element and fix, where before it was saved. Errors the diagram already had don't block an edit that leaves them alone, and restoreRevision brings back an earlier version as it was.
breaking
Operations that can't apply are refused
Each operation is checked against what the ones before it left: an edge end, child, style, relative-layout target or constraint element that doesn't exist and isn't created later in the same edit is refused at that operation, as is a constraint ID already in use (CONSTRAINT_ID_DUPLICATE). ungroupElements takes only a group, moveElement refuses an edge and a sequence diagram's nodes and frames (MOVE_NOT_SUPPORTED), and resizeElement refuses kinds it can't size, a sequence diagram's nodes and frames, and a node the engine hasn't placed yet (RESIZE_NOT_SUPPORTED, RESIZE_NOT_PLACED). deleteElement also removes the constraints left without their element and gives nodes placed relative to it an absolute layout where they are drawn, with CONSTRAINT_REMOVED and RELATIVE_LAYOUT_FIXED notes in the edit's warnings.
breaking
Moving a frame moves its members
moveElement on a frame or group moves every placed member with it, members of nested frames included, and a frame without its own layout is fitted round them again; before, such a frame didn't move, and a frame with a layout left its members behind. resizeElement on a frame without a layout gives it one, the box it was last drawn with at the new size, where it changed nothing. NODE_OUTSIDE_FRAME's fix enlarges the frame, since moving it carries the stray member with it.
breaking
ER diagram: Reads return relationships as written
ER diagram: the cardinality the engine infers from a relationship's columns, and the column anchor a label names, are worked out on every save instead of being stored with the edge, so marking a foreign key unique turns its crow's foot into a bar on the next version. view=dsp, the outline, compare, history and exports show the edge as you wrote it, while images and view=mermaid show the drawn values. Diagrams saved earlier keep the values their saves stored, and an end sent through updateEdge without cardinality is inferred again.
breaking
Usage limits cover MCP renders and creates
diagram_render on a saved diagram counts toward rendersPerMonth, as GET /v1/diagrams/:id/render does, where it was refused at the cap without adding to the count. A create, whose first version already counted toward updatesPerMonth, is refused at that limit as an edit is, and restoring a deleted diagram is refused at the diagram limit as a create is. A refusal at a limit gives the limit, the maximum and the count, and PLAN_LIMIT_EXCEEDED no longer carries plan or upgradeUrl.
breaking
An edit holds at most 500 operations
POST /v1/diagrams/:id/revisions refuses more than 500 operations with 400 INVALID_REQUEST, as diagram_edit already did, and an edit whose operations would take the diagram past 1,000 elements, edges included, is refused with CONTENT_LIMIT_EXCEEDED before it runs. A preview or a Mermaid edit past the size limits is refused before anything is drawn.
breaking
The schemas refuse fields the engine ignored
A constraint's required, a grid constraint's rows and a centerIn constraint's axis are schema errors, and the ops schema's envelope drops transactionMode, since Zindex always applies an edit whole; layoutStrategy no longer lists algorithm, nodeWidth or nodeHeight but still accepts them. Both schemas describe ER table columns in extensions.columns, so a column without a name, a flag that isn't true or false, or a key other than "PK" or "FK" is a SCHEMA_VALIDATION_ERROR naming the field, where it was ignored.
breaking
@zindex-ai/mcp is renamed to @zindexai/mcp
The npm package that serves the diagram tools over stdio is published as @zindexai/mcp from 1.0.1, so MCP client configs and npx commands name the new package, as in npx -y @zindexai/mcp, with ZINDEX_API_KEY set as before. @zindex-ai/mcp is deprecated on npm with a message pointing at the new name and stays at 1.0.0: it still installs and runs, with the deprecation warning, but without 1.0.1's refusal errors. Nothing else about the package changes.
breaking
Some earlier versions show a new hash
hash is taken over the scene as reads return it, so the 88 versions saved before chain hashes arrived that hold a frame written without a layout show a different hash than they did. Their scenes and images are unchanged, and a hash recorded earlier for one of them still matches the data-zindex-hash in that version's stored image.
breaking
The content report's operationId is now reportDiagram
In the OpenAPI spec, POST /v1/public/diagrams/:ref/report has the operationId reportDiagram in place of reportPublishedDiagram and names its path parameter ref in place of id, so a client generated from the spec gets a renamed method and argument. The URL, request body and answer are unchanged, so code that calls the endpoint over HTTP needs no change.