Built-in Widget Types
What each widget VizChat ships with is for — charts, maths, tables, text, documents, web content, infographics and slides — and where the full reference for each one lives.
Everything you place on a canvas is a widget. VizChat ships with a set of built-in widgets covering the common cases, and anything they don't cover can be built as a custom widget.
Each widget defines its own data fields and its own operations. You rarely touch those directly — you describe what you want in the chat and the agent fills them in.
This page says what each built-in widget is for. Every section ends with a link to that widget's Marketplace page, which is the full reference: fields, operations, and a live demo you can drive.
Charts and data
Bar Chart
Use it to compare values across discrete categories — regions, products, time buckets — where ranking and relative size are the point. The data is a list of labelled items you can edit on the canvas, drawn vertically or horizontally with a shared colour scheme and optional value labels. Every bar is its own anchor, so a storyboard cue can point at one category while you narrate, and the agent can add, correct or re-colour a bar between steps.
Line Chart
Use it when the question is how a metric moves — over time, across an experiment, along any ordered axis. Each item here is a whole series of data points rather than a single value, so several series share one pair of axes, with curve smoothing and per-point styling. It pairs naturally with a bar chart when the bars are buckets and the line is the running total. Series and points are addressable, so cues can walk an audience along the curve one segment at a time.
Pie Chart
Use it for composition — how a handful of parts make up a whole. It draws pie or donut form, with colour schemes, an explode effect for the slice under discussion, and label and legend strategies. Keep it to roughly five to seven slices; past that a bar chart reads better. Exploding one slice from a cue is the usual way to spotlight a single share during a walkthrough.
Math Coordinate
An interactive Cartesian plane for maths explanations. It draws twelve kinds of object — points, lines, circles, polygons, function curves, implicit curves, loci, inequality regions, angles, axis intervals, measurements and conic sections — and lines come in four flavours: segment, ray, infinite line and vector (with an arrow head and optional component markers). Objects can depend on each other: midpoints, intersections, points driven along a curve by a parameter slider, marked angles and live measurements (distance, slope, area, angle) all recompute and follow when you move what they were built from, and an object whose dependencies are gone is dimmed and dashed rather than quietly disappearing. Conic sections are first-class too: an ellipse, hyperbola or parabola carries only its own parameters, while its foci, directrices, asymptotes and vertices are separate objects built on top of it — so a cue can reveal “the right-hand focus” on its own, and a chord through that focus can be built on top of it in turn. Calculus is built the same way: the derivative f′ is a curve of its own whose expression is worked out from the function it came from — edit the function, or drag a parameter it mentions, and f′ follows — and tangent lines, extrema, monotonic intervals marked along the x axis, the common tangents of two curves and a live count of how many times two curves meet are each ordinary objects built on a function. Where a construction has several answers, each answer is its own object, so a cue can reveal “the second extremum” or “the decreasing stretch” on its own. Curves that will not fit y = f(x) are written as they stand: give it F(x, y) — x*y - 4, a rotated ellipse in full, sin(x) + sin(y) — and it draws every branch of F(x, y) = 0, labelled with the equation itself, with a parameter slider free to appear in the expression so a whole family of curves is one object you can drag through. A locus is the path a derived point sweeps out as its driver runs through its range — the midpoint of a chord as one end travels round a circle, the vertex of a parabola as a coefficient slides — sampled from the same constructions and redrawn whenever anything it is built on changes, while dragging the driver itself costs nothing extra. A function or implicit curve that names a parameter can also leave a family of ghost copies behind itself for a range of that parameter’s values, so “what the graph does as a grows” is one picture rather than a slider to drag. A parameter can also play by itself — looping, bouncing or running once across its range — and a point can leave a fading trail behind it, so a cue can set a construction in motion and let readers watch the extremum travel. Switching the plane to complex mode relabels the axes Re / Im and writes points as a + bi, so the same canvas covers complex-number questions. Each axis carries its own tick unit, so a trigonometry plot can tick the x axis in multiples of π — labelled π/2 and 3π/2 — while y stays on plain numbers, and an axis can equally be read in fractions like 1/2 and 2/3; the grid follows whatever the ticks do. Everything is reachable through operations, so narration can build a worked solution one step at a time.
Number Line
One horizontal axis and the values or ranges that sit on it. Points mark single values as solid (included) or hollow (excluded); intervals are ranges whose two ends each carry their own open/closed setting, with an arrow wherever an end runs to infinity. Overlapping ranges are stacked on separate levels automatically, so a union like x < 0 or x > 2 reads as two strokes rather than one smear. The visible span is a setting rather than a pan-and-zoom gesture, and the ticks can be read in multiples of π — labelled π/2 and 3π/2 — for trigonometry. Every point and interval is individually addressable, so a cue can point at "this half-line" while you narrate, and the solution set can grow, shrink, or change its ticks one step at a time.
Math 3D
Solid geometry, drawn the way a textbook draws it. Seven shapes cover the syllabus — cuboids, prisms and pyramids on any regular base, plus the solids of revolution: cylinders, cones, frustums and spheres. Every edge is drawn twice, once for the part you can see and once, dashed, for the part a solid is in front of, so the picture stays readable from any angle without anything being marked by hand; rotating solids have no edges to speak of, so their outline is worked out afresh from wherever you are looking from. Coordinates are the ones you already draw on paper — z upwards, x and y across the horizontal plane — so a height goes in the third number rather than the second. Named vertices, construction lines and cross-sections sit on top of the solids and are never hidden by them, because a cut you cannot see is a cut that is not doing its job — and each construction line is solid or dashed exactly as you say, which is how a height dropped inside a solid reads correctly. The view orbits the origin and can be set as a step in a walkthrough, but readers can also turn it themselves and scroll to zoom, even mid-playback, since turning a picture does not change it; wherever they leave it is remembered. Parallel projection is the default so that parallel edges stay parallel, with perspective available when a solid should look photographed. Every solid, vertex, line and cut is individually addressable, and what a cue lights up is the box the thing occupies on screen, so the annotation frames the pyramid rather than pointing at its centre.
Sequence
Turns a rule into a picture: give it the nth-term formula (2*n - 1, 3^n) or a recurrence with its starting values (a1 + a2 from 1, 1 is Fibonacci) and it draws the terms one by one, as bare dots or on stems down to zero. Turn on the running total and the partial sums appear as a second series on the same axes, which is what you want when the question is about growth, monotonicity or convergence. Terms that cannot be worked out are skipped rather than blanking the plot, and a runaway rule still shows the part that reads. Every drawn term is its own anchor and marked terms get a ring and a note, so a storyboard cue can point straight at "the fifth term" and the annotation follows the widget when it moves.
Distribution
Use it when the question is a probability: a bell curve with the stretch you are asking about shaded in and its chance printed on it, a binomial experiment drawn as the bars it really is, or a plain table of outcomes when no formula sits behind them. Shading takes a lower and an upper end and either may run to infinity, so "at least 600" is one stretch with one end left open — and the probability printed on it comes from the real infinite tail, even though the drawing stops four standard deviations out. The mean line and the one, two and three standard deviation marks work identically for all three, so the three-sigma argument reads the same whichever distribution is on screen. Every shaded stretch, called-out value and listed outcome is its own anchor, and so is the curve itself, so a cue can take in the whole distribution before zooming in on one tail.
Stat Chart
One batch of measurements, three ways to look at it. A histogram lays the classes along a continuous axis, each drawn from where it starts to where it ends — so neighbouring classes touch, unequal class widths look unequal, and a gap between classes stays a gap rather than being filled in with a zero bar. What the bar height means travels with the data rather than being a display setting, and the vertical axis is labelled to match: a plain count, a share of the whole, or a share divided by class width, the form where each bar's area is the share and the areas add up to one. A box plot draws one box per group from its five-number summary, with outliers as separate rings and every group on one shared scale, which is what makes the comparison honest; a group whose five numbers are out of order is marked as inconsistent instead of taking the rest of the chart down with it. A stem and leaf plot writes the readings out digit by digit, leading digits down the left of a rule and last digits to the right, so nothing is rounded away and the middle value can be counted off. Every class, box and row is its own anchor for a cue to point at.
Venn
Use it when things have to be sorted by which of two or three descriptions they satisfy, and the interesting cases are the ones that satisfy more than one — or none. Two circles overlap by half a diameter, three sit as a triangle, and every member carries the list of sets it belongs to rather than a position, so it lands in the piece that list describes and moves there by itself when the list changes. Shading is how you point at a piece: name some sets and say whether you mean everything inside all of them, or only the part that no other circle covers. One region paints one piece, so a union is two of them and the complement of an overlap is three; a highlight takes its colour from the sets it names. A fourth set needs stretched ovals rather than circles, so only the first three are drawn and a note says how many were left out. Circles, members and shaded regions are each addressable, which is what lets a walkthrough drop the numbers in one at a time and shade the answer at the end.
Table
Use it when the numbers matter more than a picture: frequency tables, value tables, any tidy grid. Every cell holds text, and anything wrapped in $…$ is typeset as a formula — headers go through the same path, so $f(x) = x^2$ works as a column title (write \$ for a literal dollar sign). Row and column totals sum the numeric columns only, so a column of group names stays out of the arithmetic without being marked as a label, and a 2×2 numeric area also gets a chi-square readout under the grid. Rows and column headers are addressable, so a cue can point at "this row" while you narrate, and the table can grow a row, get corrected, then reveal its margins step by step.
Probability Tree
Use it when a question runs in stages — draw a ball then another, test then re-test — and the chance at the second stage depends on how the first one came out. You give a flat list of points, each saying which point it comes from, and the branching picture is drawn from that; the number on a point is the chance of taking that branch given you got to where it starts. At the end of every path the odds of that whole outcome are written out, the branch chances along the way multiplied together — and if any branch is still unstated the result stays a question mark rather than being quietly filled in with a number. Branches leaving a point that do not come to 1 get a warning mark instead of a rejection, because a half-built tree is the normal state of one being explained, and a point that refers to a parent which is not there is drawn dashed and off on its own rather than vanishing. Every point is its own anchor, so a cue can light up one branch while you narrate, and the tree can grow a stage, get a chance corrected, then drop the branch the question rules out.
Text and code
Rich Text
Use it for the prose around the data — slide narratives, summaries, captions, static tables and formulas. It renders Markdown with KaTeX maths and per-level typography settings; there is no editing surface on the canvas, the text comes from the widget's content and only operations change it. Sections marked with <!-- section:ID --> comments can be hidden and revealed one at a time, which is how a cue brings a document in paragraph by paragraph.
Code Viewer
Use it to walk an audience through code. Syntax highlighting (Shiki) covers the usual languages, with line numbers, wrapping, light/dark themes and highlighted line ranges. Moving the highlighted range from a cue is the point: one block of code carries a whole explanation as the storyboard advances.
Documents and web content
PDF Viewer
Use it to keep a paper, report or contract on the canvas and cite it precisely. It renders a workspace PDF (referenced by workspace ref, so renames are safe) or an external URL, page by page, and the agent can turn to a page and frame one region — a paragraph, a table, a figure — in a single step. That framing is what makes a citation checkable: "it says so here" arrives with the box already drawn around it. Cues move page and highlight together, so a document walkthrough runs alongside the narration.
Web Viewer
Use it to quote the web without leaving the canvas. The live site is never embedded — the canvas blocks external frames outright, whatever the target site allows — so the backend extracts the page's article text and the widget renders that sanitized copy, with the extracted title, the favicon and a link back to the original. A phrase can be highlighted and scrolled to, and the agent reads the same extracted text you see, which is what makes "summarize these pages" a reliable request. Lay a dozen reference pages side by side for a literature review, then cue through them.
Web Link
Use it as a bookmark card: the title, description and favicon of one external URL, with no page body fetched and nothing embedded. It suits resource lists and reference rows where the link itself is the content. When you want the page's text on the canvas instead, reach for Web Viewer.
HTML Embed
Use it for HTML files that live in your workspace — a generated report, an exported notebook, an authored page. The document is referenced by its workspace ref, so moving or renaming it is safe, and it renders sandboxed: scripts inside it never run. The agent can scroll to and highlight an element, so a cue can steer the audience to one section of a long report. External live pages can't be shown here — Web Viewer covers those.
Graphics
Infographic
Use it when the layout carries as much of the message as the data — a hero section, a step diagram, a comparison, a stat callout. The content is a declarative DSL whose first line names a template from the built-in library; the renderer handles typography, palette and responsive layout, and theme settings (style, palette override, fine-grained theme config) restyle it without touching the content. The agent writes and extends that DSL through operations, so you ask for the change in chat rather than looking for a text field. It works well as the visual anchor of a storyboard step, paired with cue captions and a persona.
Image
Use it to put the figure a question belongs to — an apparatus drawing, a geometry figure, a photo, a screenshot — next to the work about it. The picture comes from a workspace file (by ref, so renames are safe) or from a URL, with alt text and an optional caption. What makes it more than a picture frame is regions: named boxes given in percentages of the original picture, so each one stays on the same spot however the widget is resized. Every region is its own anchor, so a cue can light up one part of a busy drawing while you talk about it, then move on to the next.
Slides
Cover Slide
Use it to open a storyboard, mark a chapter break, or close on a thank-you page. Three text layers — eyebrow, title, subtitle — sit over a solid or gradient background, with alignment settings. A reveal counter says how far the three-layer sequence has advanced, so a cue can bring the cover in one line at a time, each step fading and sliding in.
Custom widgets
When none of the built-in widgets fit — a timer, a kanban board, a calculator, a bespoke interactive component — the agent can build one for you and publish it to your private library. Published widgets are reusable across documents and can be shared through the Marketplace.
Built-in and user-published widgets are searched from the same pool, so the agent reuses an existing widget before building a new one.
Next step
- Marketplace — every widget's detail page, plus everything users have published
- Custom widgets — how to get one that isn't here
- What a widget is — the underlying model
Canvas Overview
The canvas is where you compose a VizChat document — a storyboard for step-by-step narration, or a dashboard for a static layout — with the AI reading and writing it directly.
Canvas Keyboard Shortcuts
Keyboard reference for the VizChat canvas — selection, copy/paste, delete, zoom, and storyboard playback navigation.