new
Remote MCP server with nine diagram tools
Any MCP client can connect to https://api.zindex.ai/mcp over HTTPS with nothing to install, using an API key or a Zindex sign-in, and work through nine tools: diagram_find, diagram_get, diagram_validate, diagram_create, diagram_edit, diagram_render, diagram_history, diagram_delete and diagram_share. Renders come back as SVG and PNG links instead of inline SVG, and deleting, publishing or reporting a diagram asks the person to confirm first. The Connect page in the docs has install links for Cursor and VS Code and the command for Claude Code.
new
MCP server sign-in with your Zindex account
claude.ai, the Claude apps, ChatGPT and other clients that support MCP sign-in connect with only the server URL and a Zindex account, through OAuth 2.1 with PKCE, so there is no API key to copy. diagram_validate and unsaved previews work before signing in, and the first tool call that needs the account opens a Zindex page where you sign in and choose Allow. Every app you allow is listed under Connected apps on the API keys page, where you can disconnect it.
new
@zindex-ai/mcp 1.0 offers the diagram tools over stdio
@zindex-ai/mcp 1.0 passes every call through to https://api.zindex.ai/mcp, so Claude Desktop, Cursor and other hosts that run MCP servers locally over stdio get the same nine diagram_* tools, and the server's own instructions, from npx -y @zindex-ai/mcp. It reads your API key from ZINDEX_API_KEY, and ZINDEX_MCP_URL optionally points it at a different server. Without a key it still starts: diagram_validate and unsaved previews work, and the other tools return AUTH_REQUIRED, naming ZINDEX_API_KEY.
new
Diagrams in the dashboard, with every version
A new Diagrams tab in the dashboard lists your diagrams, most recently changed first, with search. Each diagram's page shows any version's image (step through with Older and Newer, or pick a date), can color what changed since the version before, and lists the history with each version's time, client and message. From the same page you can download the version shown as SVG, PNG, scene JSON or Mermaid, download every version as JSON Lines, copy an image link, publish the diagram or make it private, and delete it.
new
REST API for diagrams at /v1/diagrams
/v1/diagrams does the same work as the MCP tools: Zindex assigns each diagram a dgm_ ID, and an optional key of your own names it, so repeating a create returns the existing diagram instead of a second copy. It covers reads, edits through POST /v1/diagrams/:id/revisions, history with the client and message of every version, comparing two versions or dates, signed SVG and PNG links, a JSON Lines export of the whole history, delete with a 24-hour undo, publishing and problem reports. It replaces /v1/scenes, which is retired.
new
Read, compare and restore versions by date
Reads, renders and history take asOf (a date, meaning the end of that day in UTC, or a timestamp with an offset) as well as a revision number, and comparing two versions accepts numbers or dates, so the diagram as it was last Friday is one call. The new restoreRevision operation makes the next version equal to an earlier one, chosen by number or asOf, without rewriting history. Both work on /v1/diagrams and the MCP server's tools.
new
Each version's image is drawn once and kept
A version's image is drawn when the version is saved and kept: GET /v1/images/{id}@{revision}.svg and .png, the signed links from diagram_render and GET /v1/diagrams/:id/render, and the dashboard's Diagrams pages all serve that image, so later engine updates never change an old version's picture. Versions saved before this got their images on 2026-10-05, drawn by that day's engine, and each version's SVG names the diagram, version and content hash it shows in the data-zindex-id, data-zindex-revision and data-zindex-hash attributes of its root element. To see a version drawn by today's engine, add ?engine=current to an image request sent with your API key; the result isn't kept, signed links ignore it, and any other value is a 400.
new
Version images are kept in Cloudflare R2
Each saved version's image is kept in Cloudflare R2, in its US jurisdiction and encrypted at rest, and the privacy policy's sub-processor list includes it. A deleted diagram's images last through its 24-hour undo and are removed within a day of the diagram itself. Deleting your account removes all of its images within a day.
new
History export includes each version's image
Each version's line in the history export, from GET /v1/diagrams/:id/export or the dashboard's JSON Lines download, carries image: the stored SVG (svg) with its hash, the engine build that drew it, when (renderedAt), and whether it was drawn after the version was saved (backfilled). One download holds the whole history, pictures included.
new
Create diagrams from Mermaid text
POST /v1/diagrams, /v1/diagrams/validate and /v1/diagrams/preview, and the diagram_create, diagram_validate and diagram_render tools, accept an erDiagram or flowchart as mermaid in place of a scene. The conversion is deterministic: element IDs come from the Mermaid IDs, the same text always gives the same diagram, and anything approximated or left out is listed in warnings. Flowchart subgraphs and styles are left out with a warning, other Mermaid diagram types are refused with UNSUPPORTED_DIAGRAM_TYPE, and a syntax error returns PARSE_ERROR with its line and column.
new
Read and edit diagrams as Mermaid
view=mermaid on GET /v1/diagrams/:id, or view: "mermaid" on diagram_get, returns an entity-relationship diagram as an erDiagram and other families as a flowchart; sequence diagrams have no Mermaid view yet. Sending the whole diagram back as edited Mermaid to diagram_edit or POST /v1/diagrams/:id/revisions applies only the differences: unchanged elements keep their IDs, nodes and edges left out are deleted, and unchanged text makes no new version. Mermaid edits work on architecture and entity-relationship diagrams and never change frames, groups or text, which stay editable through operations.
new
Nodes keep their positions after the first save
The engine records where it places each node in the node's layout, which view=dsp reads return, and every later edit starts from those positions instead of laying the diagram out afresh: new nodes go next to their neighbors, clear of the rest, and a node whose content changes is resized where it stands without pushing others aside. Edges, labels and frames are redrawn on every save, and sequence diagrams are laid out from their messages each time. A diagram saved before 2026-10-07 records its positions on its next edit that changes it, laid out once by today's engine, and that edit's changed list, history and comparisons show only what the edit itself changed.
new
autoLayout lays out all or chosen nodes again
autoLayout without elements lays out every node again, as on a new diagram, and with elements places only the listed nodes again, next to their neighbors; that is the fix NODE_OVERLAP_DETECTED names. Changing or removing layoutStrategy or diagramFamily with updateScene lays out every node too, and a layout of every node leaves the members of a frame with its own layout where they are, since the frame keeps its size. A new order constraint on a saved diagram takes effect when the same edit sends autoLayout without elements, and an autoLayout that lands every node where it was makes no new version.
new
Unpositioned nodes are placed around pinned ones
A node without a position is placed next to the nodes it connects to, inside its own frame, lane or band, while pinned nodes, those whose layout gives a position, stay where they are; before, a diagram that mixed the two drew all the unpositioned nodes on top of each other in the top-left corner. A new rank gets the spacing the drawing already uses between its ranks, and a node with layout.x and layout.y set to 0 counts as not placed, so the engine places it and sizes it to its content. Nodes the engine places keep those positions in later versions, and sequence diagrams place their lifelines in the order they are written.
new
NODE_OUTSIDE_FRAME warns of nodes outside their frame
Validate, create and edit return NODE_OUTSIDE_FRAME for a node outside a frame that keeps its size: a frame with its own layout, or a lane of a pool with one. A node lands there when it is given coordinates outside, placed by a new diagram's layout, which ignores such a frame's box, or placed again where the frame has no room. Its fix names the frame: move the node inside with updateNode, or move or enlarge the frame (for a lane, its pool) with moveElement or resizeElement.
new
Warnings flag overlapping frames and misplaced nodes
Validate, create and edit return FRAME_OVERLAP_DETECTED when two frames overlap and neither holds the other, and NODE_IN_WRONG_FRAME when a node is drawn inside a frame that doesn't list it. A layout of every node, as on a new diagram, places nodes without regard to frames, so a frame sized around its members can grow over another frame or around a node that isn't its own. For FRAME_OVERLAP_DETECTED, send autoLayout with one frame's nodes in elements, or move a frame that has its own layout with moveElement; for NODE_IN_WRONG_FRAME, send autoLayout with the stray node once its own frame is clear, or move the node with updateNode.
new
autoSize: "content" fits wrapped nodes to their text
A node whose label wraps (textStyle.wrap: "word" or "character") and whose layout sets autoSize: "content" keeps its width and grows down from its top edge to the height the label needs, the same height TEXT_OVERFLOW reports; other nodes don't move to make room, so one that grows into a neighbor gets NODE_OVERLAP_DETECTED. The height you give is the minimum and stays in the saved layout, so the node re-fits its label on every edit that changes it; a diagram that already uses autoSize: "content" draws those nodes taller, where their text needs it, from its next version. It has no effect on labels that don't wrap, which the engine already fits, or on nodes that place their own text, such as UI flow screens and modals, ER tables with columns, org chart people and BPMN events.
new
Every diagram write is safe to retry
An edit that changes nothing makes no new version and returns noop: true, and an edit sent with a requestId returns its first result with replayed: true when retried, for the life of the diagram. Repeating a create with the same key and the same content returns the existing diagram instead of a second one. Replays and no-op edits don't count toward your monthly diagram updates.
new
Compact outline reads and a diagram brief
diagram_get and GET /v1/diagrams/:id return a compact text outline by default (one line per element and edge, about 200 to 600 tokens for a typical diagram), with view=dsp for the full scene. knownRevision returns only what changed since a version the agent already has, and elements narrows the outline to named elements and their neighbors. The new scene.brief field (up to 1,000 characters) keeps an agent's notes on what the diagram shows and the conventions to keep, versioned with the diagram.
new
Report a problem with one exact version
A version that draws wrong can be sent to Zindex support with a one-line description, from the diagram's dashboard page, the playground, diagram_share or POST /v1/diagrams/:id/reports. Reports are anonymized by default, with labels and IDs replaced by random characters of the same width; you preview exactly what support will see before sending, and can view or withdraw a sent report later. Reporting doesn't change who can see the diagram, so public diagrams can be reported too.
new
Published diagrams are readable without a key
GET /v1/public/diagrams/:id returns the scene and kept image of a published diagram's current version without an API key; it takes the ID only, answers a private, deleted or unknown ID with the same 404, and is never cached, so making a diagram private takes effect at once. Anyone, signed in or not, can report a published diagram that breaks the terms of service with POST /v1/public/diagrams/:id/report. The public page at zindex.ai/s/<id> shows that kept image, watermark included, so an engine update never changes a published page.
new
Agent skill for the Zindex diagram loop
https://zindex.ai/skill/SKILL.md, in the Agent Skills format, teaches an agent to find, read, change and render diagrams, keep context in the brief, and track a diagram's ID in a .zindex.json file between sessions, with a longer authoring guide alongside. The MCP server's instructions link to it.
new
Preview themes, restorable diagrams and publish dates
POST /v1/diagrams/preview takes a theme (clean, dark, blueprint or sketch); the theme parameter applies to previews only, not to saved versions' images. GET /v1/diagrams?deleted=true lists the diagrams that can still be restored, each with deletedAt and expiresAt. Publishing a diagram returns madePublicAt, and a view=dsp read includes it while the diagram is public.
improved
Every plan keeps a diagram's full history
Versions are no longer removed after a plan's retention window: every plan, Free included, keeps every version of every diagram for as long as the account is active.
improved
Fewer edge crossings, and pinned layouts hold still
Diagrams the engine lays out cross fewer edges and bend less: edges that run backward loop around the outside, edges that dock on the same side of a node nest instead of crossing, and labels prefer spots clear of other edges' lines. Edges are routed from the final node positions the same way whether the engine placed the nodes or the diagram pins them, so hand-placed diagrams cross fewer edges too, and a layout pinned where the engine placed it draws the same picture again. Architecture diagrams with extensions.autoGroup draw their bands when nodes are pinned too, unless a type's nodes are scattered among other types.
improved
Growing drawings keep balanced margins
When nodes are placed past the canvas's right or bottom edge, or left of or above the drawing, the canvas grows and the margin on that side matches the one on the opposite side, up to a fifth of the content's size, instead of a fixed 40 px (60 px at the bottom). A diagram that grows edit by edit stays balanced on its canvas.
improved
Validate, create and edit report drawing problems
Validate, create and edit, on /v1/diagrams and through the diagram_validate, diagram_create and diagram_edit tools, return what drawing the diagram notices in warnings: TEXT_OVERFLOW, NODE_OVERLAP_DETECTED, NODE_OUTSIDE_FRAME, FRAME_OVERLAP_DETECTED, NODE_IN_WRONG_FRAME and LABEL_DUPLICATION_DETECTED, each with a one-line fix and, where there is one, the element. Edits also check the version they make against the diagram's rules, such as a decision with one way out or an ER edge to a missing column, which only a fresh validate reported before. Each rule warning names its element, so a warning that applies to several elements is listed for each one.
improved
Render links count toward the monthly render cap
Each successful GET /v1/diagrams/:id/render counts toward rendersPerMonth and shows in the dashboard's render usage. diagram_render on a saved diagram is refused once the cap is reached but doesn't add to the count, and previews and opening an image link never count.
improved
Docs and examples use only /v1/diagrams
The workflow examples, revisions, rendering and authentication pages use only /v1/diagrams, and the nine workflow walkthroughs are curl calls shown with the responses the API returns. The seven examples in the examples library teach the diagram_* tools and /v1/diagrams, and each example's .diff.json is what GET /v1/diagrams/:id/compare returns for the change it shows.
fixed
Website sessions refuse changes from other sites
A request that uses the zindex.ai sign-in cookie to change anything must come from zindex.ai itself; one sent from another site is refused with 403 CROSS_SITE_REQUEST. Reads, and requests made with an API key, are unaffected.
fixed
Edits accept every value a new scene does
Operations accept the same values as a scene you create, including objectStorage, the workflow message, timer, error and terminate events, workflow.gatewayEvent, the uiflow.* node types, the openTriangle and openCircle arrowheads, order constraints, the accentColor and forceLabel style properties, and fragment frames, all of which were refused with 422 on edits. createFrame no longer requires layout: a frame without one is sized around its children, as it is in a new scene.
fixed
Edits keep every field they send
createNode and updateNode keep metadata and extensions, so ER columns added through operations stay in the diagram, createEdge and updateEdge keep edgeType and metadata, and createFrame keeps containerType and laneDirection. setStyle with styleOverride merges the override onto the element's current style instead of erasing it.
fixed
Validation errors name what to use instead
A validation error's one-line fix, from validate, create or edit, keeps its hint: an unknown property is named with what to use instead, such as extensions.columns for ER columns or layoutStrategy on updateScene for autoLayout's old strategy, or with the properties allowed there. A value outside an allowed list is named along with the allowed values, nearest first. Before, a hint sat in a second sentence that the fix left out, and an unknown property on a scene element wasn't named.
fixed
createEdge gives edges their usual arrowheads
An edge added with createEdge and no arrowheads of its own gets the ones the same edge gets in a new scene, so a default edge points at its target instead of drawing as an undirected line. Edges added earlier keep the arrowheads they were saved with, which updateEdge can change through endpoints. The scene schema and the element types page give each edge type's defaults, where they said a default edge had none.
fixed
ER diagram: Edits draw relationships as a new diagram does
ER diagram: a version saved by an edit infers each relationship's missing cardinality, and anchors an edge whose label names a column, such as "user_id", at that column, as creating the diagram does. Before, an edge added with createEdge at a column and without cardinality drew arrowheads in the saved image where a preview drew crow's feet, with no FAMILY_ER_MISSING_CARDINALITY warning. The first edit of a diagram built by edits also fixes the edges it doesn't name, without moving any node, and images of earlier versions don't change.
fixed
Edits can re-create edges a delete removed
Deleting an element in an edit also deletes the edges that start or end at it, and their IDs are free again for the rest of the edit, where creating one of them again failed with OP_ID_ALREADY_EXISTS. One edit can delete a frame, create it again with new children and reconnect its edges.
fixed
updateScene removes fields set to null
updateScene accepts null for palette, layoutStrategy and diagramFamily and removes the field, as its description says, where the schema refused it with VALIDATION_FAILED.
fixed
resizeElement keeps the side its anchor names
resizeElement's anchor ("top", "right", "bottom" or "left") is the side that stays put: the element grows or shrinks away from it and stays centered on that side, so with "bottom" it grows upward. The option was accepted and ignored, and every resize kept the top-left corner, as one without an anchor still does.
fixed
Edges to Mermaid subgraphs no longer fail imports
A Mermaid flowchart edge to or from a subgraph is left out with a CONSTRUCT_DROPPED warning instead of failing the whole import; point it at a node inside the subgraph to keep it. view=mermaid writes an edge to a frame or group as an edge to the subgraph that shows it, and a Mermaid edit reads it back as that edge, changed or deleted like any other; before, such an edit failed. In an edit, an edge to a subgraph that shows none of the diagram's frames or groups is left out with the same warning.
fixed
Edges end on the side they connect to
An edge enters the node it ends on through the side it docks at, instead of running under the node to reach its far side, as some ER relationships did. Relationships between ER tables stacked one above the other run as a bracket along whichever side is closer, and diagrams cross fewer edges overall.
fixed
Each version's image matches its own content
An edit that changed only an ER table's columns, an edge's column anchor, ports or style, or a frame's title could be drawn with the previous version's layout, so an edge moved to another column stayed on its old row in that version's image. Each version is laid out from its own content, and images already kept for earlier versions don't change.
fixed
textStyle.wrap: "character" breaks long words
In node labels and text elements, "character" wraps at spaces like "word" and, when a word is too long for the line, breaks it between two characters, which suits long identifiers, paths and URLs. It was drawn like "none", and every layout widened a node that used it to fit its label on one line instead of keeping the node's width.
fixed
Wrapped labels stay inside their padding
Wrapped lines no longer run into the padding or up to the border, left- and right-aligned lines keep their padding, and a \n in a wrapped label is measured as the line break it is drawn as. TEXT_OVERFLOW on a wrapped node reports the height the node needs, padding included, where its fix came up short, and catches top- or bottom-aligned lines that cross the node's edge. An edit that changes a wrapped node's content no longer widens the node to fit its label on one line.
fixed
Multi-line edge labels sit on their backgrounds
An edge label with a \n is drawn by its edge, on its own background, where its text was drawn at the top of the canvas and the background sat empty beside the edge. Images kept for earlier versions keep the misplaced text, and ?engine=current on an image request sent with your API key draws one with the fix.
fixed
Nested frames no longer disappear behind their container
A frame is drawn on top of every frame that contains it, whatever order the scene lists them in and whatever their zIndex. A frame listed before its container was painted over by the container's fill and disappeared, border and title included.
fixed
Long frame titles stay inside the frame
A frame without its own layout, which the engine sizes around its members, is at least as wide as its title, where a long title over narrow members ran past its right border. The extra width goes on the right, so the members and the title keep their places. A frame with its own layout keeps the width you give it.
fixed
Frame borders no longer show through long titles
The background behind a lane's or frame's title spans the title as it is drawn, where it stopped short of the last letters and a border or edge running under the end of a long title showed through them.
fixed
Edge labels keep clear of frame borders
In a diagram with frames, edge labels keep their distance from each frame's border as the frame is finally drawn, and the leg of an edge that loops back steps clear of another edge's label beside a frame. Before, both were checked against where each frame stood before the drawing was centered on its canvas, so a label could sit on a border and a leg could stay beside another edge's label. A later edit no longer moves those labels and legs.
fixed
ER diagram: Relationship lines follow crow's-foot notation
ER diagram: the crow's foot at a many, zero-or-many or one-or-many end spreads its prongs on the table and meets on the line; it was drawn backwards and read as an arrowhead pointing into the table. er.identifying relationships draw a solid line instead of a dashed one, as Information Engineering notation draws them, and to draw a non-identifying relationship dashed, give an er.relationship edge a style.dash. Images of versions saved earlier keep their look, and the next version draws the fix.
fixed
ER diagram: Rows meet their edges in every theme
ER diagram: under the clean and dark themes, and on saved diagrams with a dark canvas background, which use the dark theme, rows sat a little lower than the edges docked at them, more with each row, so a long table ran past its box and its last edges met the wrong row. Rows keep the same height in every theme, and the text keeps the theme's size.
fixed
ER diagram: Self-relationships end without an arrowhead
ER diagram: a relationship from a table to itself, such as a parent_id key, ends with no marker. Its crow's-foot marks are left off because the loop is too tight for them, and the default arrowhead drawn in their place could merge with another relationship's mark on the same key column. Self-loops in other families keep their arrowheads.
fixed
ER diagram: Join tables relate many-to-one to their parents
ER diagram: a column counts as unique only when it has unique: true or is its table's only pk column, so a join table keyed by its foreign keys, such as post_tags with post_id and tag_id, draws many-to-one to each parent where it drew one-to-one. A table whose whole primary key is a foreign key to its parent, such as user_profiles.user_id, still draws one-to-one, and diagrams saved earlier keep the cardinality their saves stored.
fixed
ER diagram: PK+FK columns show both badges
ER diagram: a column that is both pk and fk shows a PK and an FK badge, where the PK badge hid the foreign key. Every row of its table leaves room for both badges, so names stay aligned, and the table grows wide enough that no name runs into its type.
fixed
Org chart: Reports hang from their manager's bus
Org chart: a manager's direct reports share one trunk and bus that leaves the manager's card on the side facing them and enters each report's card head-on, whichever way the chart runs. An edge no longer leaves a card where another arrives and runs along it, which made a report read as reporting to its manager's manager.
fixed
Sketch theme fills cover their shapes
In sketch theme previews, a rectangle's fill painted only thin slivers along its hand-drawn sides, so an ER table's header all but lost its color and its dark title sat on the dark body. Fills cover the whole shape, and ER headers and org chart accent stripes sit under the hand-drawn outline instead of over it.
fixed
Diagrams read with view=dsp save again
A frame written without a layout came back from view=dsp reads, the history export and public reads with a layout the scene schema refuses, so saving the document again failed with 422. Reads return such a frame as it was written, so a diagram read back saves again unchanged.
fixed
Edit warnings describe the version the edit made
An edit's warnings come from the version it saves, so a drawing warning the edit fixed, such as NODE_OVERLAP_DETECTED, no longer comes back on that version and every later one. NODE_OVERLAP_DETECTED names a fix an edit can make: send autoLayout with one of the nodes in elements to place it again, or move one of them.
fixed
ER diagram: Key-anchored ends skip the cardinality warning
ER diagram: FAMILY_ER_MISSING_CARDINALITY fires only for an edge end whose cardinality the engine can't infer from a pk or fk column, and names that end. Creating a diagram whose relationships anchor at key columns no longer warns about cardinality the engine infers and draws.
fixed
Malformed input answers 4xx instead of 500
Five malformed inputs to /v1/diagrams and /v1/images answer a 4xx with a code instead of a 500: an edit whose ops isn't an array, an edit with malformed ops on an older baseRevision, restoreRevision: null, an image path with a broken percent escape, and a preview of a scene too malformed to draw, which answers 422 VALIDATION_FAILED with the scene's errors. A request without an API key or sign-in to an endpoint that needs one answers 401 with code: "AUTH_REQUIRED". POST /v1/diagrams/validate accepts a bare scene document, as documented, instead of mistaking the document's own scene field for a wrapper.
fixed
History works on diagrams missing old versions
diagram_history, GET /v1/diagrams/:id/revisions and the history export work on diagrams whose older versions an earlier retention limit removed, where they failed with 404 at the first missing version and the export could stop partway through a download. They list the versions that are kept, with a note naming the ones that are gone.
fixed
Playground, Recently deleted and public page fixes
The playground, Recently deleted and the public page use the new API; diagrams with a dot in their key open in the playground, and a late preview there no longer paints over a newer one. Recently deleted says when a restored diagram came back without its key, because another diagram took the key while it was deleted. The public read leaves out the owner's account ID, and on the public page, wide diagrams scroll instead of losing their left edge and the report form resets each time it opens.
fixed
Sign-out, agent view and phone layout fixes
Sign out shows its progress and, if the server doesn't confirm, says it failed instead of leaving the page as if it had worked. The Human / Agent view banner sits directly under the navigation bar instead of below an empty band, the agent view fits phone screens on the login and signup pages, and a ?view=agent link always opens the agent view. On phones, the pricing and changelog pages and the open menu keep the side margins the rest of the site uses.
fixed
Watermarks show the date of the version drawn
A rendered diagram's watermark shows the date the drawn version was made, rather than the date the diagram was first created.
fixed
JSON Schema $id values point at zindex.ai
scene.schema.json and ops.schema.json declare their $id as the https://zindex.ai/schemas/ URLs they are served from, replacing a domain that never resolved, so validators that resolve references by $id find them.
fixed
Reference docs match what the API returns
The validation rules page lists every code the API returns, with its severity, message and what to change, in place of rule numbers that no response carried. The OpenAPI spec's layouts, error codes and parameters match what the API takes and sends, including relative layout, which it showed in a shape validate refused. The operations page sends edits to diagram_edit and POST /v1/diagrams/:id/revisions instead of a retired endpoint.
fixed
Markdown for agents and real 404 pages
Requests for the homepage, pricing, signup and login pages with Accept: text/markdown return the markdown written for agents again, instead of the HTML page. Unknown URLs on zindex.ai return a real 404 page instead of the homepage with a 200.
fixed
llms.txt serves its content again
zindex.ai/llms.txt, the index of the docs for agents, was served empty for part of 2026-10-06 and has its content back.
fixed
Footer and inline code fit phone screens
On phones, the footer links wrap instead of running off both edges of the page, and long inline code in the docs, examples and legal pages wraps instead of widening the page. Example pages also line up with the navigation bar on wide screens.
breaking
/v1/scenes is retired; /v1/diagrams replaces it
Every /v1/scenes path answers 410 with code: "MOVED" and the method and path of the request that replaces it, with or without an API key, and the scene ID in the old path is the diagram's ID in the new one. @zindex-ai/mcp 0.x called these paths, so its tools get the same error until you move to 1.0. The OpenAPI spec drops the /v1/scenes operations, and their reference page redirects to the Diagrams API page.
breaking
@zindex-ai/mcp 1.0 replaces the dsp_* tools
Version 1.0 removes the dsp_* tools of 0.x rather than aliasing them, so agents and prompts that call them by name move to the diagram_* tools. Diagrams saved with 0.x keep their IDs and work with the new tools, and the pre-1.0 ZINDEX_API_URL setting still works. Every 0.x version is deprecated on npm with a message pointing at @zindex-ai/mcp@latest.
breaking
Problem reports replace sharing a diagram with support
A diagram can no longer be shared with support: send a problem report with diagram_share or POST /v1/diagrams/:id/reports instead. A diagram shared with support earlier reads as private, and publishing it ends that share instead of failing with 409.
breaking
Edits no longer lay out the whole diagram
After a diagram's first save, an edit keeps its nodes where they are and places only new ones; before, every save placed afresh each node without a position of its own, so one edit could rearrange the whole diagram. To lay a saved diagram out again, send autoLayout without elements.
breaking
The normalizeScene operation is removed
The normalizeScene operation returns OP_REMOVED: Zindex normalizes every diagram and places its new nodes when it is saved, so drop the op, and send autoLayout to have the engine place nodes again. The playground's Normalize button is gone.
breaking
autoLayout refuses strategy, scope and writeBack
autoLayout's strategy, scope and writeBack never changed what it did, and an edit that sends one is refused with VALIDATION_FAILED; to change how nodes are placed, send updateScene with a new layoutStrategy. elements, when given, lists at least one node ID, naming a frame or edge in it fails with OP_TARGET_KIND_MISMATCH, and the AUTO_LAYOUT_SELECTION_EMPTY code is retired.
breaking
Nodes at the origin are no longer pinned
A node with layout.x and layout.y set to 0 counts as not placed: the engine places it next to its neighbors and sizes it to its content, where before a node pinned there stayed in the corner at its own size. To keep a node where you pin it, give it any other position.
breaking
Example pages drop the CI/CD recipes
The /examples pages no longer include CI/CD recipe panels, the per-example .github-actions.yml URLs return 404, and /examples/index.json drops the ciWorkflowUrl field. Each example keeps its scene, rendered SVG and revision diff.