Choosing Opus in chat now runs Opus 5.5, still with the full 1M-token context. It used to run Opus 4.8.
The Components stage lists only built-in and shared Components — the ones it exists to audit — instead of every card-scoped copy in the workspace.
v0.4.0
Added
Diagrams pan and zoom in fullscreen — an Ilograph card opened fullscreen is an infinite canvas: drag to pan, scroll or pinch to zoom, and Fit to re-frame. Each perspective remembers where you were looking, and one you have not opened yet is centred for you. The view is yours alone — moving it does not move anyone else's.
Pull a resource aside to untangle a diagram while you read it. The arrangement stays in your tab and is remembered per perspective.
Click a resource to see what it is part of — everything upstream and downstream of it stays lit and the rest dims, so following a path is one click instead of tracing a line by eye.
Changed
A perspective now draws only the resources it relates. A model carries more than any single perspective wires up, and the surplus used to collect in a column of disconnected boxes down the left-hand side.
Diagrams lay out with far fewer crossed lines. On a busy perspective this took ten crossings to none.
Fixed
Hovering a component no longer stops you zooming the board. Ctrl+scroll and trackpad pinch now zoom wherever the pointer is; before, you had to find a gap between cards first.
Diagram labels read over the lines and boxes they cross instead of vanishing behind them.
Removed
Ten built-in components — AgentChat, CanvasArch, CedarPalette, CedarShell, CedarpadLogoTest, CedarpadLogos, CedarpadMerch, Dashboard, KpiCard and LandingV5. Any card already placed from one of these will no longer load.
v0.3.0
Added
Fullscreen browses the whole board — a collapsible rail lists every component on the page and ‹ › step between them, wrapping at both ends. The canvas underneath follows along, so Escape lands you on whatever you read last, and deleting the component you are viewing advances to the next instead of dropping you out. A page holding a single component looks exactly as it did.
Open any component fullscreen from the side panel — each row of the "On this board" list gains a ⛶, revealed on hover and on keyboard focus.
Edit a component over MCP without resending it — read_component takes offset/limit and returns just that window with line numbers, and the new edit_component replaces the lines you just read. A one-line change no longer costs the whole module, and an impossible range is refused before anything is written.
v0.2.4
Fixed
An outline is now shared by everyone on the canvas. Edits to an Outliner card stayed in the tab that made them — each person was quietly editing a private copy of the same card, and publishing an edit could wipe out rows someone else had added since. Bullets, notes, indents and completions now reach everyone. Which bullet you have zoomed into stays yours.
v0.2.3
Changed
The default workspace now runs on hardware we own, reached over a private tunnel instead of a Cloudflare Container. The container's filesystem did not survive sleep, so every cold start rebuilt its dependency bundle from scratch — and a page load that raced that rebuild got a canvas referring to files that briefly did not exist. A long-lived guest keeps its filesystem, so that way of failing to load is gone.
Fixed
A canvas that had gone to sleep now wakes on the first try. Reopening an idle canvas could fail for minutes with canvas error: ProcessExitedBeforeReadyError instead of showing the board. Waking was tied to the request that asked for it, so a browser giving up mid-load canceled the wake — and each retry canceled the one before it. A canvas now wakes once, and finishes waking even when the request that triggered it is gone.
A workspace that only needed waking no longer answers "Bad gateway". When its stack was not yet listening, the error page went straight to the browser and stayed there, with no retry — for exactly the case a single boot would have fixed.
v0.2.2
Fixed
Adding a component now puts it on the page you are looking at. On a canvas whose pages had been reordered, cards added from any page landed on the first one instead.
Removed
FlintGallery — the chart showcase. Every chart type it demonstrated is documented in the chart-authoring reference, and ChartCard renders them all.
v0.2.1
Added
Filter the Components page by tier — chips for built-in, shared, canvas and no-source, each with a count, so you can audit one tier without scrolling past the others. A chip only appears for a tier that has components in it.
State machines in components — xstate is available to anything built on the canvas, so a component with real states and transitions can model them directly instead of juggling booleans.
Removed
BarChart — superseded by ChartCard, which renders Bar, Grouped Bar, Stacked Bar and Histogram among 35 chart types, with axes, scales and legends derived from your data. Any card still placed as a BarChart will show as having no source; place a ChartCard instead.
Fixed
Chart slides no longer render their bar labels in near-invisible text.
v0.2.0
Added
Outliner — a Workflowy-style outliner you can place on any canvas. Enter splits a bullet, Tab and Shift+Tab indent and outdent, Shift+Enter adds a description, ⌘/Ctrl+Enter marks an item done, and clicking a bullet zooms into it. Edits persist across reloads, and an agent can write a whole outline in a single call.
Fixed
An outline that references a missing item no longer crashes the card — the stray row is skipped.
An outline containing a loop no longer hangs the tab.
v0.1.0
Added
Notebook search — type in the Notebook and get ranked results with highlighted excerpts instead of a filter over what is already on screen.
Agents write the Notebook — an agent can record an Observation and where it came from, create a Problem, tie Observations to it, link Problems, and record a Solution. Works from the in-canvas agent and from an outside agent over MCP.
A canvas reads the Notebook live — place a component against a Notebook query (open problems, a problem's observations, recent observations) and it shows what is there now, not a nightly snapshot. Members only.
Solutions and Problem links — record how a Problem was solved, and link one Problem to another.
The chart gallery — browse every chart type on the pad with live data and the spec that produced it, so you can see what an agent can build.
An agent-terminal chat card — a chat rendered in the grammar of Claude Code, Codex or Grok, with three skins.
Changed
The Notebook is a reading surface. The capture box and modal are gone — agents write it, you read it.
Chat reads as a terminal. Turns, tool calls, plans and approvals share one column, and a tool call opens to show the full command it ran.
The interface is a true neutral. Surfaces lost their warm tint; the cedar ember is now the only colour in the chrome.
Fixed
Charts fill the card they are placed in, at any zoom. Before, one drew at a fixed size and left the rest of the card empty.
Security
Notebook graph reads are confined to your workspace. A problem-context or neighbourhood read could previously return a record from another workspace, and a traversal could route through one.
v0.0.1
Added
The shared canvas — a multiplayer surface where you and AI agents work on the same artifacts. Multi-page, with the picker reorderable by drag; a reload returns you to the page and camera you left.
Live components — agents write real React components onto the canvas instead of describing them. Each canvas has its own; promote one and every canvas sees it. A gallery to place them, a Components stage for the workspace-wide inventory, and preview on demand.
The Notebook — record an Observation and where it came from, and synthesize Observations into a Problem. API-first, so agents can write to it while working with you.
Ilograph diagrams — architecture views with perspectives, drill-down, and a sequence walkthrough.
Passwordless sign-in — a magic link by email; no password to set or forget.
The Inbox — conversations with agents, with an unread badge pushed to you rather than polled for.
Agents reach cedarpad over MCP — an external agent can read and write your canvases with the same reach the in-app agent has.
Huddles — talk over voice on a canvas. Join from the side panel, see who is in, and get a live transcript as you go; guests on a share link can join too.
The dev agent — point an agent at a connected repo and it works in a branch and opens a pull request. Run it from Jobs, or trigger it from an outside agent over MCP.
Marts — components read the workspace warehouse by name. A component names a mart, never a URL, and a shared canvas exposes only the marts it actually uses.
Shareable canvases — invite a guest by link, see their presence, revoke it.
Workspace settings — members and invitations, private canvases, connected GitHub repositories, and web push notifications.
This changelog, and the running version in Settings.
Fixed
Components no longer fail to render after the container restarts — one that imports another now materializes together with it.