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 · 97 changes · Every week

new

New teams start with a free 14-day trial

A new team workspace starts with a free 14-day trial of the Team plan, with no card. Signed in with a verified email address, you name the workspace at zindex.ai/settings/new-team/ and you are its owner. Before, a team had to subscribe before it could try Team. During the trial the workspace has the whole Team plan for up to 10 owners and members; viewers are free and don't count toward the 10. Everyone in the workspace sees how many days are left, and its owners get an email when the trial starts, 3 days before it ends and when it ends. An owner can add 7 days once, from 3 days before the end or any time after it. Nothing is charged unless an owner subscribes, which they can do in Billing at any time. If nobody subscribes, the workspace moves to Free's limits when the trial ends and keeps every diagram, every version and everyone in it. Each person can start one trial.

new

Team trials suggest a next step three days in

Three days into a free trial of Team, the person who started it gets one email with the next step that fits what the team has done: move their agent from their personal workspace into the team, connect an agent, ask a connected agent for a diagram, or invite people. It opens with what the team has done, as counts. A team that already has diagrams and someone else in it or invited to it gets no email, and section 1 of the privacy policy lists when the email was sent.

new

Google Cloud icons draw as Google's product cards

Every gcp:* icon key draws Google's own product icon, about 250 keys with aliases such as gcp:lb and gcp:gcs, where a gcp:* key drew nothing and returned ICON_NOT_FOUND, and Google's 2026 names find theirs too: gcp:agent-platform draws Vertex AI's icon and gcp:agent-runtime Agent Engine's. A node with a gcp:* icon is drawn as Google's product card, whatever its shape: the icon at the left, then the node's label in bold over Google's name for the product, in the same colors on every canvas, and a fill such as Google's gray marks a standby. The new Google Cloud icon keys page in the docs lists every key with Google's name for it, and a product without an icon of its own draws its category's, so name the product in the label.

new

Frames draw as Google Cloud's own groups

containerType takes nine Google Cloud group presets: gcp.cloud, gcp.project, gcp.region, gcp.vpc, gcp.zone, gcp.subnet, gcp.tier, gcp.instanceGroup and gcp.onPremises. Each draws a frame as Google's reference diagrams draw that group, in Google's fill, line and title on every canvas, the blue gcp.cloud boundary drawing the Google Cloud wordmark when it has no title or the title Google Cloud, and a frame's own style overrides its preset, except on a zone. gcp.vpc and gcp.zone are spanning frames written with spans: the VPC is drawn round its subnets as one band across their Regions, and a zone as a strip under each card it spans, green, blue or yellow by its place among its Region's zones, while an instance group names itself a Regional or Zonal MIG from those zones.

new

Architecture: Google Cloud diagrams lay themselves out

Architecture: an architecture diagram with a gcp.cloud, gcp.region or gcp.subnet frame, or a diagram without a family that has one, needs no coordinates and is laid out top to bottom as Google's reference diagrams are, each zone a column through its Region's tiers, the tiers rows in the order their edges run, and the Regions side by side and mirrored, with matching rows level so replication between them runs straight. Global items such as the load balancer and Cloud DNS form a row above the Regions and those the Regions feed, such as storage and BigQuery, a row below, a Region's own services such as Cloud Run form a row above or below its subnets, and what lies outside the boundary, such as users, goes above or below it by its edges. Lines run as Google draws them, with a global load balancer that feeds two Regions standing over the gap between them and its line forking into each Region's first tier, and the layout engine docs say what the arrangement doesn't do, such as reorder Regions or arrange other families.

new

Architecture: Edits to Google Cloud diagrams keep positions

Architecture: an edit to a Google Cloud diagram keeps every position and grows a row, a zone column or a Region for what it adds, as a table inserts a row or a column, and what lies past the growth moves by exactly that much: a card added to a tier stands in its zone's column, a new tier goes in at its place in the subnet's flow, and a new Region goes right of the last. A load balancer over the gap between two Regions moves to stay over it, and the items joined only to it, such as Cloud Armor beside it and the users over it, move with it where their row has room. A whole boundary an edit adds is laid out as on a new diagram, an order constraint an edit adds swaps two nodes that stand the wrong way around in a row, and the save's changed lists every node that moved.

new

Google Cloud diagrams draw Google's connector

In a diagram with Google Cloud groups or gcp:* icons and no AWS ones, every edge is drawn as Google's connector, a 2 px near-black line with rounded bends and 14 px labels, and an edge that sets no endArrow, with its edgeType unset, default, or an er.*, network.* or org.* type, ends in the open arrow, openTriangle, drawn as Google's small chevron. An edge's own stroke, strokeWidth and dash still apply, and where the near-black reads below 3:1 on the canvas, such as a dark one, the line outside Google's filled groups takes the palette's or the theme's edge color. A diagram that mixes AWS and Google Cloud keeps the theme's edges, with the open arrow.

new

Three warnings name Google Cloud frames to change

GCP_GROUP_NESTING names a Google Cloud group where Google doesn't nest it, such as a subnet no Region holds, or a VPC or zone written with children instead of spans, and GCP_GROUP_UNTYPED names the preset to set on a generic frame titled like a Google Cloud group that holds gcp:* icons. GCP_FRAME_HIDES_GROUPS names a generic frame round a Region's subnets, or round a subnet's tiers, that hides them from the Google Cloud layout, and says to list them in the group around it. The Validation rules page says how to change each frame with an edit.

new

Emoji are drawn in their own colors

Every emoji of Unicode 18.0 is drawn in its own colors, whatever the theme or the text's color, where it was an empty box: skin tones, flags, keycaps and ZWJ sequences such as 👩‍💻 are each one emoji, never bold or slanted, while a sign Unicode shows as text by default, such as ⚠ or ❤, stays text unless the emoji selector, U+FE0F, follows it (⚠️). CHARACTER_NOT_DRAWN no longer reports emoji, and its fix no longer suggests writing them as words. An emoji is about a quarter wider than the font size, so a node sized by a label that holds one can be a little wider than before and the diagram's layout can shift slightly; previews and new versions are drawn this way, a version already saved keeps its image, and Text and scripts, at zindex.ai/docs/reference/rendering/#text-and-scripts, lists what is drawn.

new

Chinese, Japanese and Korean text is drawn

Chinese, Japanese and Korean text is drawn as written, with its punctuation and the fullwidth letters and digits typed with it (A, 1), where it was empty boxes with the CHARACTER_NOT_DRAWN warning. It is measured as it is drawn, so a node sized by its label fits it, and the warning's fix names common Chinese, Japanese and Korean characters among the ones to write in. Most rare Chinese characters, those in Unicode's later extensions, are still boxes, as are Arabic, Hebrew, Thai, the Indic scripts and emoji. A diagram already saved shows its text when it is next drawn.

new

scene.lang picks the forms of Chinese characters

scene.lang, a BCP 47 language tag such as ja, ko, zh-Hans, zh-Hant or zh-HK, picks the forms Chinese characters take: Japanese, Korean, Simplified, Traditional or Hong Kong Chinese. Without it, or with a language other than Chinese, Japanese or Korean, the text decides: kana reads as Japanese, Hangul as Korean, and other Chinese characters by the forms they use, with a short label that Traditional Chinese and Japanese share read as Simplified Chinese, so set lang when you know the language; Hong Kong forms need zh-HK. updateScene sets it and null removes it, a Mermaid edit keeps it, and a tag that isn't well formed is refused with LANG_INVALID. The SVG names the diagram's language in its root's lang, for screen readers: the scene's lang, else the language its Chinese, Japanese or Korean text shows.

new

Sequence diagram: Mermaid sequenceDiagram works both ways

Sequence diagram: Mermaid sequenceDiagram text sent as mermaid creates, checks or previews a sequence diagram, and view=mermaid returns any sequence diagram as a sequenceDiagram you can edit and send back, so the VIEW_UNAVAILABLE error is gone. Participants keep their Mermaid IDs, while messages, notes, blocks and activation bars are named in row order, as m1, note1, loop1 and act1. What Zindex leaves out or draws differently comes back in warnings: a participant's links and properties are left out (CONSTRUCT_DROPPED), and a half arrowhead is drawn as the nearest whole one (ARROW_APPROXIMATED).

new

Sequence diagram: Mermaid edits apply only what changed

Sequence diagram: edited sequenceDiagram text sent as mermaid with baseRevision, to diagram_edit or POST /v1/diagrams/:id/revisions, is compared with that version, and only the differences are applied, as operations. Participants are matched by ID and messages and notes as lines, so an unchanged line keeps its ID, a line cut and pasted elsewhere is moved, and text sent back as read returns noop: true. A message whose ends change gets a new ID, and rows added to or cut from a fragment Mermaid has no block for, such as neg or ref, come back as FRAMES_NOT_CHANGED: change those fragments with operations.

new

Sequence diagram: Messages are numbered with badges

Sequence diagram: autonumber in the document's top-level extensions numbers every message top to bottom, as true or with a start and an increment, and each number is a circle badge centered on the start of its arrow. A message's own step draws the same badge and the count goes on from it, so steps draw in every family and the warning STEP_NOT_DRAWN is gone. Adding, deleting or moving a message renumbers the ones after it, notes and hidden messages take no number, and autonumber in any other family draws nothing and gets the warning AUTONUMBER_NOT_DRAWN.

new

Sequence diagram: Lifelines start and end at their messages

Sequence diagram: a lifeline whose first row is a sequence.create message arriving at it has its head on that row, with the arrow ending at the head's side, and a lifeline that a sequence.destroy message arrives at ends at that arrow's X, with no foot. A lifeline used above its create message or below its destroy message is drawn full height on that side, with the warning FAMILY_SEQ_USED_BEFORE_CREATE or FAMILY_SEQ_USED_AFTER_DESTROY.

new

Sequence diagram: Notes sit beside a lifeline

Sequence diagram: a sequence.note with one lifeline's ID as its extensions.anchor and an extensions.side of "left" or "right" sits beside that lifeline instead of over it, and the gap beside the lifeline widens to hold it. A note across several lifelines is always drawn over them, and side there gets the warning FAMILY_SEQ_NOTE_SIDE_IGNORED.

new

Sequence diagram: Participant boxes and highlight bands

Sequence diagram: a frame that holds lifelines or actors is drawn as a participant box around them, with its title centered above their heads, and a frame that holds messages, notes or fragments is drawn as a filled highlight band behind those rows. Neither needs a containerType, since the frame is drawn by what its children hold, and its own layout is ignored. A row placed after or before a row in a band joins the band, while a lifeline placed between two of a box's lifelines stays out of the box's children (FAMILY_SEQ_BOX_SPLIT) unless the operation names the box as its frame.

new

Sequence diagram: Mermaid box and rect carry over

Sequence diagram: a Mermaid box around participants imports as a participant box, with its title and color, and a rect around rows as a highlight band in its color, named box1 and rect1 in the order written; both were left out, with a CONSTRUCT_DROPPED warning. view=mermaid writes them back as box and rect, a hex fill as rgb() or rgba(). In a Mermaid edit a changed title or color is one updateFrame, a new box or rect makes a frame, and a removed one deletes its frame and leaves its participants or rows.

new

Sequence diagram: Six more participant types

Sequence diagram: sequence.boundary, sequence.control, sequence.entity, sequence.database, sequence.collections and sequence.queue join sequence.lifeline and sequence.actor as node types that head a column, and only the head and foot they draw differ. A boundary, a control, an entity and a database draw an icon with the name under it, as an actor does, and collections and a queue draw a box with the name inside, in every theme. Each works wherever a lifeline does, as a message's end, a note's extensions.anchor or the after or before of another participant: a frame that holds any of them is drawn as a participant box, and a sequence.create message arriving at one ends at its icon or box.

new

Edges can end in a cross arrowhead

An edge's startArrow and endArrow take cross: an X on the line just short of the node, the size of the filled arrowhead, in every family and theme, with a dashed line fitted so its last dash ends at the node. A Mermaid flowchart's --x and x--x import as cross ends, with no ARROW_APPROXIMATED note, and view=mermaid writes them back, so a read sent back as an edit changes nothing. Dotted and thick links also keep a circle end (-.-o, ==o), which a read wrote as an arrow.

new

fontFamily mono draws text in a monospace font

A style's fontFamily takes mono, which draws text in a monospace font, for paths, endpoints and types, and measures it in that font, so a node sized by its label fits it; sans, Inter, is the default. In mono an arrow typed in ASCII, such as ->, stays as typed. Any other value draws in the default font, as every value did before, and now returns the warning FONT_FAMILY_UNKNOWN, one for each family, where it was ignored silently. A fontFamily in scene.defaults is never drawn from, so mono there gets the warning too, and a diagram whose styles already set mono draws in it when it is next drawn.

new

updateScene changes document settings in extensions

updateScene takes extensions, which sets four of the document's top-level settings in an edit: autonumber, showLifelineFeet, autoGroup and layoutHints. Each key given replaces its value whole, null removes it, the document's other keys stay, and any other key is refused. A changed layoutHints lays the diagram out again, as a changed layoutStrategy does.

new

CHARACTER_NOT_DRAWN warns of text drawn as boxes

Checking, saving or editing a diagram whose drawn text holds characters the images draw as empty boxes, such as letters outside Latin, Cyrillic and Greek, emoji and some signs, returns one CHARACTER_NOT_DRAWN warning for the whole diagram. Its fix lists up to five of the characters and names up to three of the elements that hold them, and says to write the text in Latin, Cyrillic or Greek letters, with signs and emoji in words, or to keep the text and accept the boxes, since the diagram is saved as written. The warning's section of the rendering reference, at zindex.ai/docs/reference/rendering/#character_not_drawn, says which operation changes each kind of text.

new

Every response carries an X-Request-Id header

Every API response, an error or not, carries an X-Request-Id header: Zindex's own ID for that request, never one you send, to quote to support when something looks wrong. An INTERNAL_ERROR names it in its error, over REST and in an MCP tool's error, so an agent can pass it on. It isn't the requestId you choose for an edit.

new

A docs page says which version link goes where

Diagrams in other tools, at zindex.ai/docs/embedding/, is one home for which version link goes where: @<n> pins one version for an issue, a review or a page an agent publishes, and a link without it follows the latest for a README. Claude artifacts come first, with both ways in: a command that embeds a version's PNG as a data URI with its caption, and the Zindex connector, which the page calls with <ref>@<n> when it opens to pin a version, with the version's page link as each live section's fallback. Linear and Jira join GitHub, a README, Notion, Confluence, Google Docs, Slack, email and a web page, and the REST API page's table of places moves there.

new

The playground has a sequence diagram sample

The playground offers a third Mermaid sample, a checkout sequence diagram with numbered messages, an activation bar, a note beside a lifeline and an alt block. A pasted sequenceDiagram is drawn there as a flowchart or an erDiagram is, and Save diagram creates the diagram from it.

new

An example maps public interfaces for security reviews

The new worked example, Public-facing interfaces diagram from Terraform, gateway routes and ingress (public-interfaces-for-security-review), has an agent draw what a store's Terraform, gateway routes, OpenAPI specs and Kubernetes ingress files open to the internet, for a security review, with no coordinates. Each path from the internet is labeled with how it is authenticated, and a dashed red line marks the one path that skips the WAF and the load balancer and needs no credentials, a public bucket prefix. The review's fix, which closes that prefix, is recorded as revision 2, with the before and after pictures, the compare, the prompt, the operations and the example's own preview card, and as a showcase example its before picture is at /examples/public-interfaces-for-security-review.before.svg.

new

An example draws Google Cloud architecture from Terraform

The new worked example, Google Cloud architecture diagram from Terraform (gcp-architecture-from-terraform), has an agent draw a used-car marketplace's listings platform from its Terraform with no coordinates, laid out as Google's reference diagrams are, and record a pull request that adds a Pub/Sub topic and a Cloud Run function for photo thumbnails as revision 2. It has the before and after pictures reviewers see before terraform apply, the compare that lists the cards the edit moved, the prompt, the operations and the example's own preview card. What agents draw, on the homepage, gains a Google Cloud architecture tab after AWS architecture, with the VPC across two Regions and each VM on its zone's colored strip.

new

The homepage shows a diagram drawn from code

From code, second on the homepage after the hero's replay, shows one request a person makes of their coding agent, to map every public-facing interface across their repositories, with the files the agent read and the diagram it saved with diagram_create as revision 1, the picture Zindex kept for it. It says the diagram shows what the agent found in the code it read, links the worked example, and Present opens the picture full screen. Other things to ask lists three more requests, for a request path, a data model and a deploy pipeline, each linking to its example.

improved

Armenian, Georgian and more signs draw as written

A character the diagram font doesn't have is drawn from a fallback font, where it was an empty box: Armenian and Georgian letters, the rarer Latin, Cyrillic and Greek letters, the phonetic alphabet, and more arrows, math and logic signs (∧, ∀, ∈), symbols and box-drawing characters. Other scripts, among them Chinese, Japanese, Korean, Arabic, Hebrew, Thai and the Indic scripts, the fullwidth letters and digits typed with them (A, 1), emoji and the signs Unicode shows as emoji by default, such as ⭐, and a few rare signs are still boxes with the CHARACTER_NOT_DRAWN warning, whose fix now suggests Latin, Cyrillic, Greek, Armenian or Georgian letters instead, while a sign shown as text by default, such as ⚠ or ♥, is drawn as text, even with the emoji selector after it. Previews and new versions are drawn this way, a version already saved keeps its image, and Text and scripts, at zindex.ai/docs/reference/rendering/#text-and-scripts, lists what is drawn and what is a box.

improved

Text is shaped as browsers shape it

Text in images is shaped as browsers shape it: letters are kerned, and an arrow typed in ASCII, such as ->, <-, <->, => or -->, is drawn as one arrow, in code such as p->next too. Kerned labels are a little narrower, so a node sized by its label can be a few pixels narrower, and a new version's layout can shift by a few pixels, occasionally more. Previews and new versions are drawn this way, and a version already saved keeps its image.

improved

Wrapped text breaks by Unicode's line breaking rules

Wrapped text breaks by the Unicode line breaking rules in every script: at spaces and tabs, after hyphens and slashes, and between Chinese or Japanese characters, never before closing punctuation such as 。 or a small kana such as ッ, while Korean breaks at its spaces. Only "character" breaks inside a word, a number or a date, and only one wider than the line, and a soft hyphen (­) marks a break of your own that ends its line with a hyphen. With "auto", the default, a long hyphenated label, such as a function name under an AWS icon, wraps after its hyphens inside its node, where the node grew as wide as the whole name. When a diagram is next drawn a few labels break at different points, such as a URL or a path in "character" mode.

improved

Edge labels keep clear of nearby nodes

An edge's label keeps 16 px from every node but the two its edge joins, where it kept 6 px, so a label no longer reads as a nearby node's caption. Beside the node its edge leaves or enters it keeps that distance too, except on the face the line actually uses. This applies to architecture, flowchart, workflow and every other family with routed edges; sequence diagrams are drawn as before.

improved

Edge labels find a place between frames

A label with no clear place along its line between two frames is tried beside the line's end, inside the frame the line enters or leaves, before it is left across the frame's border. A label slid across its own line stops while the line still runs under its text, so a line no longer vanishes under a label's first letter.

improved

Edge labels avoid frame titles and crossed frames

An edge label inside a frame keeps 16 px, a line's height, from the frame's title, where it kept 6 px and could read as the frame's second title. A label sits inside a frame that holds an end of its edge, and stays out of a frame its line only passes through wherever the line has a clear place outside, so it no longer reads as a note on what that frame holds; a pool's lanes and a VPC's Availability Zones don't count when the pool or VPC holds an end. This applies in every family, while a pool's or lane's header strip keeps 6 px and a sequence message's label keeps its row.

improved

Architecture: Outside items joined by edges share one side

Architecture: items outside a Google Cloud or AWS Cloud boundary that an edge joins, directly or through other items outside, stand on one side of it together, and a frame outside moves whole, so the line between them stays short, where each took a side by its own lines and a customer's browsers above the boundary could be joined to a CDN vendor below by a line all the way around it. When any of them has more edges into the cloud than out, as users or a customer's environment do, they all stand on the side users take, above a Google Cloud boundary or left of AWS Cloud by default, and otherwise on the other side, as a payment provider the services call does. In a saved diagram each placed item keeps its side and an item an edit adds outside joins its saved partners' side, while a line an edit adds between two saved items on opposite sides goes around the boundary until autoLayout puts them on one side.

improved

Architecture: Storage written and read stands below the Regions

Architecture: in a Google Cloud diagram, a global item with as many edges from the Regions as into them, such as a Cloud Storage bucket that one service in a Region writes and another reads, stands in the row below the Regions, as Google's reference diagrams draw storage, where it stood above and a line could loop around the diagram to reach it. An item joined to the Regions goes above only when more of its edges lead into them than come from them, such as the load balancer or Cloud DNS, and the same rule puts a Region's own items above or below its subnets. In a saved diagram an item already placed keeps its row when an edit adds lines that would give it the other, until autoLayout.

improved

Sequence diagram: Mermaid participant types keep their shape

Sequence diagram: a Mermaid participant declared with a type, as in participant db@{ "type": "database" }, imports as the matching node type, such as sequence.database, for boundary, control, entity, database, collections and queue; each was drawn as a lifeline, with a SHAPE_APPROXIMATED note. view=mermaid writes the type back, and a type changed in a Mermaid edit is one updateNode. A participant imported before this release stays a lifeline and reads back with its type, until updateNode sets its nodeType.

improved

Sequence diagram: Nested activation bars are stacked

Sequence diagram: an activation bar that starts while another is open on its lifeline is drawn 8 px to the right of it, a level for each open bar, so no bar hides another. A message meets the outermost bar's face on the side toward its other end, and a bar that opens on the row another ends on is drawn in line with that one.

improved

SCHEMA_VALIDATION_ERROR messages are in plain words

A SCHEMA_VALIDATION_ERROR, for a scene or an edit that doesn't fit the schema, says in plain words what to send, such as /schemaVersion: '9' is not allowed; use '0.1'. or /elements/0/id: 'has space' can hold only letters, digits, '.', '_', ':' and '-'., where it showed the schema's regular expression or the validator's own phrase. A value of the wrong kind, a missing property and a number or a length out of range read the same way. A layout.mode that doesn't exist is one error that names the two modes and what each needs, where a scene got several.

improved

Mermaid syntax errors name the line you sent

Mermaid text that can't be read answers PARSE_ERROR with a short error that names the place, such as Mermaid couldn't read line 3, column 4. Check the syntax there., with the line counted in the text you sent, where front matter, comments, a directive or a code fence ahead of the diagram shifted it. A mistake Mermaid describes about the diagram itself, such as a participant deactivated twice, is passed on as Mermaid's, after Mermaid says:. IMPORT_TIMEOUT and IMPORT_FAILED say to send shorter text, or a scene.

improved

The 10-warning limit keeps whole-diagram warnings

A response's warnings hold at most 10, and they always include the warnings that cover the whole diagram, CHARACTER_NOT_DRAWN and DRAW_TIMEOUT, when those apply. The first 10 were kept in order, so a response with many per-element warnings could leave them out and read as if nothing else was wrong.

improved

Positions and document depth join the size limits

An element's position or size must be within 100,000 px of zero, and a document, or an edit's ops, may nest at most 32 levels deep. A scene past either is refused with CONTENT_LIMIT_EXCEEDED, as one past the other size limits is, and the error says which value is over and what to change: bring the element within 100,000 px of zero, or make the diagram smaller or split it. Size limits, at zindex.ai/docs/reference/validation-rules/#size-limits, lists every limit with its message.

improved

A PNG over 30 megapixels is refused before drawing

A PNG over 30 megapixels, its width times height, is refused before it is drawn, with DRAW_TIMEOUT and an error that says to bring the elements closer together or split the diagram: a .png image read, a preview as PNG, or diagram_render with preview. The SVG of the same diagram draws, so a save succeeds, and Drawing time, at zindex.ai/docs/limits/#drawing-time, has the details.

improved

Every API request can be up to 4 MiB

A request to any /v1 route can be up to 4 MiB, the limit /mcp already had, and a larger one answers 413 with INPUT_TOO_LARGE. Mermaid text and a feedback snippet keep their 64 KB, and Sizes, at zindex.ai/docs/limits/#sizes, lists each input's size.

improved

Checking a large invalid scene is faster

POST /v1/diagrams/validate and diagram_validate check the size limits before the schema, so a scene over a limit answers with its CONTENT_LIMIT_EXCEEDED errors at once, and the schema errors are worked out for the first ten elements that have them, which is as many as the response listed. A scene of 1,000 elements with hundreds of unknown properties each is checked in a quarter of a second, where it took ten.

improved

Icons at a remote URL return a warning

A node's or a frame's icon given as an http or https URL returns the warning ASSET_URI_POLICY_VIOLATION when the diagram is checked, saved or edited, as an asset's remote URI does. The SVG keeps such an icon as a link most viewers don't load, and PNGs leave it out, so give the icon as a data: URI, which is drawn wherever the image is. Nothing changes in how a diagram is drawn.

improved

The playground shows previews as images

The playground's preview pane shows the drawn SVG as an image, as the diagram pages show saved versions, so an icon given as an http or https address isn't loaded in the preview either, as it isn't in a PNG; give it as a data: URI to see it everywhere. The picture already in the pane stays until the new one is ready, so the pane never collapses between previews, and zooming and panning work as before.

improved

Pages agents publish show one saved version

The Zindex skill, the MCP tools docs page, the MCP server's instructions and diagram_render's description say that a page an agent publishes shows one saved Zindex version, never a diagram the agent draws itself, and that an artifact can't load images from other sites, where they said such pages often can't. Where the agent can run a command it embeds the version's PNG with its caption, and where it can't the page calls diagram_render with <ref>@<n> and preview: true when it opens, with the version's page as its fallback, as Diagrams in other tools describes.

improved

A shorter homepage without Use cases and Revisions

The homepage drops its Use cases section, and its Revisions section with the revision ledger, so it is shorter. From code shows what an agent draws from code and links four worked examples, and the hero's replay shows each revision and what it changed. Every example Use cases linked is still in the examples library.

improved

Storage settings warn when the App reaches every repository

The Storage page in Settings shows a caution when the Zindex GitHub App reaches every repository on its GitHub account rather than the one the workspace connects, on the Location card of a connected workspace and under the repository field while connecting, naming the account and the repository to limit the App to on GitHub. It is advice rather than a hold, so saving and reading go on, and the caution goes once the App is limited. Section 8 of the privacy policy's list of what Zindex records for a connected workspace names whether the App reaches every repository on the account.

improved

Email verification picks up where you left off

After you verify your email for a checkout that asked you to, you come back to it: the Pro checkout on Billing, or a paid team's checkout on New team, opens again. For a free trial, New team comes back with the name you typed and says your email is verified and you can start the trial now, leaving the start to you. Verifying from the banner on New team comes back the same way.

improved

Sign-up from an invitation fills in its address

Signing up or signing in from a team invitation's page starts with the address the invitation was sent to in the email field.

fixed

API and MCP calls no longer fail or stall

API and MCP calls are answered as before, after many failed with 502 or 503 from about 17:09 UTC on 2026-11-11, some after waiting 40 to 60 seconds. Zindex tells MCP clients that its tool list never changes, so Claude Code no longer keeps a connection open to Zindex for each session. Claude Code also connects to Zindex faster, and nothing changes in how you connect.

fixed

MCP confirmations without a yes return a token

Over MCP, diagram_delete and diagram_share with publish, link, report or feedback ask the person to confirm, with a yes or no form in an app that supports elicitation on the 2026-07-28 protocol. Any answer to the form other than a yes returns CONFIRMATION_REQUIRED with a confirm token, the result apps without the form get, so the agent asks the person in chat and, if they agree, calls again with the token, as Confirmation, at zindex.ai/docs/tools/#confirmation, describes. Such an answer returned the error CANCELLED with no way to continue, which stopped these actions in some apps that answer the form without showing it; no tool returns CANCELLED any longer.

fixed

Cyrillic, Greek and Vietnamese text draws as written

Text in Cyrillic, Greek and Vietnamese, and in the other accented Latin letters, arrows and signs the diagram font, Inter, has, is drawn as written, in Russian, Ukrainian, Bulgarian, Serbian, Macedonian, Greek, Turkish and Vietnamese among other languages; those characters were drawn as empty boxes that ran past the node's border. A letter typed with a combining accent is drawn as the one accented letter, and an image already saved keeps the picture it has, while a diagram's next saved version is drawn correctly. Scripts Inter doesn't have (Chinese, Japanese, Korean, Arabic, Hebrew, Thai and the Indic scripts) and emoji are still drawn as empty boxes, in a label that fits inside its node, and Text and scripts, at zindex.ai/docs/reference/rendering/#text-and-scripts, lists what is drawn and what is a box.

fixed

Short Traditional Chinese labels draw in Traditional forms

Without scene.lang, a Chinese character counts for every character list that has it, Simplified, Traditional or Japanese, so a short Traditional Chinese label such as 資料庫 or 通知服務 is drawn in Traditional forms, where it was drawn in Simplified ones. A diagram whose Chinese characters Simplified Chinese, Traditional Chinese and Japanese all share still reads as Simplified Chinese, and Japanese written without kana can read as Traditional Chinese; lang settles both. An image already saved keeps its forms, and a diagram is drawn in the new ones when it is next drawn.

fixed

Nodes are never narrower than their longest word

A node whose label wraps on its own, with textStyle.wrap "auto", the default, is at least as wide as the label's longest word, so a long word beside short ones, or a hyphenated one with no space, no longer runs past the node's border. "auto" never splits a word. With "word" the node keeps the width it is given and a wider word still runs past its edges, which "character" breaks between its characters.

fixed

Crowded frames grow to hold new nodes

In architecture, flowchart, workflow, ER and UI flow diagrams, and those without a family, a node an edit adds to a frame goes inside it, and each frame round it without a layout grows to hold it with its padding and title. Where no place keeps those frames clear of what lies beside them, the node goes past the frame's other members in its column and what lies past the frame's far border moves across the flow by the growth, where the frames could grow over a node or a frame beside them, with FRAME_OVERLAP_DETECTED or NODE_IN_WRONG_FRAME. Only nodes an edit adds move others this way, the save's changed lists what moved, and AWS and Google Cloud diagrams place nodes inside their arrangement by their own rules.

fixed

Only lines drawn alike share an end point

Two lines end at one point of a node only when they are drawn alike there, with the same edgeType, style and arrowhead, or none, at that end, and one drawn differently keeps a point of its own, closer than 12 px to the next where the side has no more room. Before, an arrow could take the point of a line with no arrowheads, so in a Google Cloud diagram a user's HTTPS arrow appeared at the end of Cloud Armor's line into a load balancer, as if Armor sent that traffic. In Google Cloud diagrams a load balancer's line to a card below also leaves its bottom rather than the side Cloud Armor stands close beside, and a line from the global row into a tier's card lands on the card's top, clear of the tier's title.

fixed

Architecture: Lines between outside items go around the boundary

Architecture: a line between two items outside a Google Cloud or AWS Cloud boundary, a gcp.cloud or aws.cloud frame, such as a customer's browsers above it and a CDN vendor below, goes around the boundary however long the way, where a long way around let it run straight through the boundary and what it holds, as if the traffic passed through the cloud account. A line with an end inside the boundary crosses it as before, an edge whose router is straight or curved keeps its own line, and one with no clear way at all still crosses, with EDGE_OBSTACLE_GRAZED. A version already saved keeps its image.

fixed

Architecture: Lines under a tier's title drop straight

Architecture: in a Google Cloud diagram, a tier's line into a card in the next tier drops straight into the card's top, clear of that tier's title, where a title over the card's top sent the line outside the tier and into the card's side. It happened most often with an instance group over a database tier whose primary is in the first zone. A version already saved keeps its image.

fixed

Architecture: Lines between neighboring cards run straight

Architecture: in a Google Cloud diagram, an unlabeled line between two cards that stand next to each other in a tier, in one zone or both in none with no card between them, runs straight between their facing sides, where it dropped below the row, ran across and rose into the second card, a U. A labeled line, a line between cards in two zones, such as failover, and a line past a card between its ends keep the lane under the cards, and a tier whose joined cards are all neighbors joined this way no longer keeps that lane, so it is 25 px shorter. A saved diagram is drawn with the line straight and the tier shorter when it is next drawn, and its nodes keep their places.

fixed

Architecture: Edits place a subnet's nodes above the tier

Architecture: in a Google Cloud diagram, a node an edit adds to a subnet at the same place in the flow as a tier goes in a row above that tier, where it landed in a row under the tier and the line to it ran down through the tier. A cache the app tier feeds, which shares the database tier's place in the flow, stands between the app and database tiers this way. A subnet's nodes at a tier's place in the flow form a row above the tier wherever the subnet lists them, so the edit draws what a new diagram draws.

fixed

Box-drawing trees and tabs line up

Box-drawing characters line up as in a terminal: in a tree such as ├── src/ the spaces beside them take the width of their cells, so each branch sits under its parent's name, and the vertical lines join from line to line. A tab moves to the next tab stop, every eight spaces from the start of its line, where it was drawn as an empty box with the CHARACTER_NOT_DRAWN warning. Other columns lined up with spaces line up only in mono, and a diagram with a tree or a tab changes when it is next drawn.

fixed

Control characters in text no longer break images

A label or other text that holds a control character, such as a NUL or a vertical tab, no longer makes the SVG malformed or the PNG fail to draw. A character no SVG can hold (a control character below U+0020 other than a tab or a line break, U+FFFE or U+FFFF) is left out of the picture, with no warning, and the text is measured and drawn without it. The saved scene keeps the text as written.

fixed

Every ID and name works, __proto__ included

An element ID, a constraint ID, a style name or an asset ID that spells a word such as constructor, toString or __proto__ is checked, drawn, saved and edited like any other, and a missing style or asset is reported whatever its name, so no ID or name is reserved. Before, a preview or a save with an edge named this way failed with an internal error, a node named __proto__ was drawn without its edges, and a style named __proto__ was dropped when the diagram was saved. Diagrams that don't use these names draw exactly as before.

fixed

An edge label's action is drawn in italic

An edge label written as EVENT [guard] / action, as a state machine's transitions are, draws its action in italic, as the docs describe, where it was drawn upright. The action is measured in italic too, so its label's room fits it, and a diagram with such labels changes slightly when it is next drawn.

fixed

Edge labels such as reads / writes are plain text

An edge label written as EVENT [guard] / action is styled as a state machine's transition only in flowchart, workflow, UI flow and sequence diagrams. In every other diagram, one with no diagramFamily included, a label such as reads / writes or HTTPS [mTLS] is drawn as plain text, where 8.42.0 drew the part after the slash small, gray and italic, and a part in brackets small and gray. Such a label can be a little wider, so its diagram changes slightly when it is next drawn.

fixed

Bring your own storage never builds on outside changes

Bring your own storage: a save, or the commit that removes a deleted diagram after its 24 hours, never builds on a change made outside Zindex under .zindex/, such as a commit by someone else or a rewound or force-pushed branch, even one that lands while Zindex is writing. The workspace goes on hold as when the regular check finds such a change: the save is refused with STORAGE_HOLD, the owners are emailed once, and saves wait until an owner restores or disconnects on the Storage page in Settings. Saves of other diagrams, and changes outside .zindex/, are built on as before.

fixed

Moves into your repository finish checking every version

A move of a workspace's diagrams into its own GitHub repository checks every copied version and finishes when the history includes older versions with a frame sized to its contents. Such a move stopped at that check and retried every minute without finishing, and since saves wait while a move runs, the workspace couldn't save until the move was canceled. Nothing was lost and the files written to the repository were correct; a canceled move can be started again from the Storage page in Settings.

fixed

Layouts with labels inside frames aren't penalized

The layout picker, which chooses among candidate arrangements, no longer counts an edge label inside a frame as a label over a node, which penalized layouts that put labels in frames. A label across a frame's border counts instead, so the picker prefers arrangements that keep labels clear of borders.

fixed

Sequence diagram: The first row starts below an actor's name

Sequence diagram: the first message row starts below the tallest participant head at the top, rather than below a lifeline's box, so an actor's name, which hangs under its figure, no longer collides with the first message's label or a first note. Every stored diagram with an actor at the top is drawn up to 30 px taller; a diagram of lifelines only is drawn as before.

fixed

Sequence diagram: An actor's dashed line reaches its foot

Sequence diagram: an actor's dashed line runs down to the figure at its foot, where it stopped 6 px above it, so that line is 6 px longer in stored diagrams too. An actor created by a sequence.create message has its figure on the arrow however many lines its name takes, where the figure and name together were centered on it, and a multi-line name under an actor no longer makes the lifeline boxes beside it taller.

fixed

Sequence diagram: Fragment guards clear open activation bars

Sequence diagram: a fragment's guard, such as [payment ok], starts to the right of any activation bar open in its row, so its background no longer cuts the bar in two. A fragment's guards line up in one column when every row allows it, and the fragment grows to the right to hold a guard moved past a bar; its left edge and its tab stay where they were.

fixed

Sequence diagram: Step badges leaving a bar sit on the bar

Sequence diagram: a numbered message leaving an activation bar has its step badge centered on the bar, covering it from side to side, where it was centered on the bar's face and covered half the bar, leaving a corner of the bar its message closes showing beside it. The arrow still starts at the bar's face.

fixed

Sequence diagram: Sibling fragments share a left edge

Sequence diagram: fragments one under another beside the same lifeline, such as a critical, a break and a par, start at the same X. Each was cleared of the lifeline or bar beside it by its own tab's width, so their left edges stood up to 7 px apart.

fixed

Dashed lines end with a dash under a cross arrowhead

A dashed edge ending in a cross arrowhead has a dash under the center of its X in every line width. The dashes were fitted to the node the line ends at, which at some widths left a gap under the X's center, so the cross stood on nothing.

fixed

Dashed edge types follow their style.dash

The edge types that draw dashed (workflow.messageFlow, sequence.reply, sequence.create, org.dottedLine and uiflow.back) use their 6,3 dash as a default: an edge's style.dash sets another pattern, and [] draws the line solid. They drew 6,3 whatever the style said.

fixed

Sequence diagram: Notes over active lifelines aren't flagged

Sequence diagram: a note drawn over a lifeline whose activation bar is open there no longer gets NODE_OVERLAP_DETECTED, since that is where a note over an active participant belongs. A note over another lifeline, or beside one, that covers a bar is still flagged.

fixed

Scenes that carry computed are accepted

A scene that carries computed, the engine's drawing output that reads leave out, is accepted when it is checked, previewed or saved, over the REST API and MCP, and the engine ignores it, whatever it holds, and works it out again. Before, a computed with fields the JSON Schema didn't list made the scene fail validation. The schema marks computed readOnly and accepts any value for it, and drops the definitions only it used: computedLayoutRegistry, computedEdgePathRegistry, computedEdgePath, diagnostic and bounds.

fixed

Records that carry autoSize: "children" are accepted

A version's record that carries autoSize: "children" on a frame's layout, such as one in a workspace's own GitHub repository for a version saved before early November, is accepted when its scene is checked, previewed or saved as a new diagram, over the REST API and MCP, or pasted into the playground. The API reads such a frame as one without a layout, so the scene goes back as it is, where each frame its author wrote without a layout was refused with 'children' is not allowed. The schema is unchanged, so autoSize still names content or none, and Checking a document, at zindex.ai/docs/reference/schema/#checking-a-document, says how to check such a record locally.

fixed

createFrame keeps each element in one frame

createFrame with frame takes each element its children lists out of that outer frame's children, so every element stays in one frame. A nested frame or fragment can be deleted and created again in one edit, with the same id, its rows in children and frame naming the frame that held it.

fixed

A frame title's TEXT_OVERFLOW fix offers updateFrame

The TEXT_OVERFLOW warning for a frame whose title is wider than the frame offers to shorten the title with updateFrame, beside widening the frame with resizeElement. It said no operation could shorten a saved frame's title, which updateFrame does.

fixed

Switching workspace after a link no longer jumps back

After a link from an email or from GitHub opened a team's Members, Billing or Storage page in Settings, switching workspace from the header menu stays in the workspace you choose, where it went back to the linked one. A link can also open the dashboard's API keys or Diagrams page in the workspace it names, and for a workspace you're not in or have left, the page says so and shows the one you're working in.

fixed

Billing and Storage links act only on their workspace

A link from an email or from GitHub to Billing or Storage in Settings for a workspace you can't open no longer acts on the workspace you're in. Billing adds no trial days and opens no seat review or checkout, Storage skips the GitHub repository step, and each page says why it shows your current workspace.

fixed

Workspace links that couldn't open can be tried again

When a link from an email or from GitHub to Members, Billing or Storage in Settings, or to the dashboard's API keys or Diagrams page, can't open its workspace just then, because Zindex is busy or the connection is lost, the page says so as an error, with Try the link again. That opens the link again with everything it came with, such as the step GitHub sent you back for on Storage, or the extra trial days, the seat review or a checkout on Billing, where the page dropped them and showed a quiet note. Switching workspace afterwards still stays in the workspace you choose.

fixed

Teams on Free's limits take nobody new

Nobody new joins a team workspace whose free trial ended unpaid, or whose subscription ended, by invitation, the invite link or a domain, where an invitation sent earlier still let its person in as a viewer. A pending invitation works again if an owner subscribes before it expires, and its page, like the invite link's, says the team isn't taking new people right now before anyone makes an account.

fixed

Pricing points members to the team's owners

On Pricing, a member or viewer of a team workspace on Free's limits gets See the owners, which opens Members in Settings, with the note that only an owner can subscribe. The Team card offered them Start Team, a button only an owner could use.

fixed

Members shows no charges that may not apply

Members in Settings says nothing about paid seats or charges until it knows whether they apply, so a team whose plan Zindex provides never reads that its members are paid seats, even before the page has loaded its billing details or when they can't be loaded. A change that could add a paid seat still asks you to confirm first.

fixed

The API may also run in Dallas

The privacy policy, updated 2026-11-11, names Dallas beside Ashburn, Virginia, and Chicago as a place in the United States where the API may run, and says that at any time the API runs in one or more of the three. The hosting provider and the country are the same as before.

fixed

This month's usage names the right month

The note under This month's usage on Billing, and under This month on the Usage page, names the month the usage is counted in, the UTC calendar month. In a time zone west of UTC it named the month before, such as October 2026 on November 11.

fixed

The homepage replay keeps the page below still

The caption under the hero's session replay keeps the height of its tallest moment at every width, so From code and the rest of the page below stay still while the replay plays. On phones the page below shifted by up to about 40 px as each moment's caption took more or fewer lines.

breaking

MCP requests need sign-in or an API key

Every request to the MCP server at https://api.zindex.ai/mcp needs sign-in with a Zindex account or an API key, the first a client sends included, and a request with neither answers 401 with a sign-in challenge and code: "AUTH_REQUIRED". An app that supports sign-in asks you to sign in as it connects, so the first save works. Checking a scene with diagram_validate, previewing an unsaved scene and opening a version link with diagram_render no longer work over MCP without an account, while POST /v1/diagrams/validate and POST /v1/diagrams/preview over REST still need no key. An app that connected without sign-in must connect again and sign in, in Claude by reconnecting the connector or removing it and adding it again; apps that already signed in, and anything that sends an API key, need no change. The connect page, at zindex.ai/connect/, and the tools guide, at zindex.ai/docs/tools/#sign-in-with-your-zindex-account, have each app's steps.

breaking

@zindexai/mcp 2.0.0 needs a key for every tool

@zindexai/mcp 2.0.0, the npm package that serves the diagram tools over stdio, needs ZINDEX_API_KEY for every tool, where diagram_validate, unsaved previews and version links worked without one. Without a key it starts and lists the tools, and every tool call answers AUTH_REQUIRED, saying to set a key from zindex.ai/dashboard/api-keys/. Versions before 2.0.0 no longer start without a key, so update the package or remove a version pin; with a key, every version works as before.

breaking

SVG text points to glyph outlines in <defs>

A script that copies an element's data-zindex-element group out of an SVG image must also copy the image's <defs>, since in an image drawn from 8.41.0 the group's text is <use> references to glyph outlines defined once there. An image drawn before 8.41.0, which a version saved earlier keeps, has one path per line of text. Showing an image, or finding an element in it by its id, needs no change, and images are smaller, since each glyph's outline is stored once.

breaking

PARSE_ERROR no longer carries expected

A Mermaid PARSE_ERROR no longer carries expected, the parser's token names, and line and column are absent when Mermaid can't place the mistake, as with one it describes about the diagram itself. Only a caller that reads these details needs a change: treat line and column as optional, and show error, which names the place or the mistake, as it is.