Mviz

mviz v1.7.0

mviz

Generate clean, data-focused charts and dashboards from compact JSON specs or markdown. Maximizes data-ink ratio with minimal chartjunk, gridlines, and decorative elements. Uses a 16-column grid layout system.

Setup

No installation required. Use npx -y -q mviz which auto-downloads from npm. The -q flag reduces npm output while still showing lint errors.

For faster repeated use, install globally: npm install -g mviz

What This Skill Does

mviz is two things:

1. A spec language — markdown with fenced code blocks describing components (charts, KPIs, tables, notes, etc.). One markdown file (or a single JSON spec) describes a whole visualization.
2. A set of renderers that turn that spec into HTML for different display contexts. Today there are two:
   • visualize:show_widget (inline) — when you're in claude.ai with the visualization tool available, render the mviz spec as a widget_code HTML fragment using claude.ai's design system (Chart.js, var(--color-*) tokens). The widget appears inline in the conversation.
   • mviz CLI (standalone) — mviz dashboard.md -o report.html produces a complete HTML document with mviz's own design system (mdsinabox theme, ECharts). Suitable for downloads, iframe tiles, print/PDF.

The mviz markdown spec is canonical; the renderer is chosen based on where the visualization will appear. Same spec, two renderings. A third-party could build a different renderer (Chart.js-based CLI, Slack-themed renderer, etc.) without changing the spec.

Picking a renderer

Always compose the mviz markdown spec first, even when rendering inline. The spec is the canonical artifact — emit it in a fenced code block in your response so the user can copy / reuse / pipe-to-CLI later. Then render it with whichever renderer fits the request.

Before rendering, decide which renderer fits — and state your inference so the user can correct you. Some signals:

Inline (visualize:show_widget):

• "show me X" / "what was X" / "how did X change" / "render"
• A small visualization (typically 1–4 components) that flows with the surrounding prose
• The user is exploring data, not preparing something to share
• visualize:show_widget is available in this conversation

Standalone (mviz CLI):

• "make a report" / "send me the file" / "export to PDF" / "save"
• Multi-section composed view (5+ components, dividers, page breaks, full-width tables)
• The user wants something to save, share, print, or embed in an external host (iframe tile, dashboard cell)
• visualize:show_widget is unavailable (no inline rendering surface)

Default to inline when ambiguous. State the inference explicitly:

"I'll render this inline since you asked to see it in chat. If you'd rather a downloadable file, I can run the mviz CLI on the same spec — just say the word."

If the user later asks for the file version after an inline render, the spec is already in the conversation — pipe it through mviz ... -o out.html without re-composing.

Visual Style (mdsinabox theme — for the CLI renderer)

• Font: Helvetica Neue, Arial (clean sans-serif)
• Signature: Orange accent line at top of dashboards
• Palette: Blue primary, orange secondary, semantic colors (green=positive, amber=warning, red=error)
• Background: Paper (#f8f8f8 light) / Dark (#231f20 dark)
• Principles: High data-ink ratio, no chartjunk, minimal gridlines, data speaks for itself

Renderer: mviz CLI (standalone HTML)

The CLI renderer produces a complete HTML document styled in the mdsinabox theme above. Output is suitable for direct browser viewing, PDF export, iframe embedding, or sharing as a file.

Single Chart (JSON)

echo '<json_spec>' | npx -y -q mviz -o chart.html

Dashboard from Markdown

npx -y -q mviz dashboard.md -o dashboard.html

Dashboard from Folder

npx -y -q mviz my-dashboard/ -o dashboard.html

Iframe-embedded output (--embed)

When the rendered HTML will live inside another product's iframe (mdw-turbo Prism tiles, dashboard cells, etc.), pass --embed to strip page chrome (red accent bar, title row, theme toggle) and apply a battery of post-processes that lean the body: CSS pruning, transparent backgrounds so the host's card chrome shows through, ECharts script dropped when no charts are present, marked dropped when no marked.parse(...) caller, JS minify, etc.

npx -y -q mviz dashboard.md --embed -o tile.html

Embed mode emits a complete HTML document (DOCTYPE + html + head + body intact). It's not the right tool for visualize:show_widget — for that, use the other renderer (see below).

16-Column Grid System

Components are sized using size=[cols,rows] syntax:

```big_value size=[4,2]
{"value": 1250000, "label": "Revenue", "format": "currency0m"}
```
```bar size=[8,6]
{"title": "Sales", "x": "month", "y": "sales", "file": "data/sales.json"}
```

• 16 columns total width (both portrait and landscape)
• Row height: ~32px per row unit (approximate - charts have padding)
• Page capacity: Portrait [16c × 30r], Landscape [16c × 22r]
• Components on same line share the row
• Empty line = new row

Height Guidelines:

For charts with many categories (10+ bars, 10+ rows in dumbbell), increase row units to prevent compression.

Side-by-Side Layout

Critical: To place components side-by-side, their code blocks must have NO blank lines between them:

```bar size=[8,5]
{"title": "Chart A", ...}
```
```line size=[8,5]
{"title": "Chart B", ...}
```

This renders Chart A and Chart B on the same row. Adding a blank line between them would put them on separate rows.

Headings and Section Breaks

Heading Guidelines:

• Set the page title with either the frontmatter title: or a leading # H1, not both.
• Use ## H2 for section titles within a page (most common).
• Use ### H3 for lightweight subheadings that don't interrupt flow.

Section vs Page Breaks:

• Use --- to separate logical sections visually. Content flows naturally to the next page when needed.
• Use === only when you explicitly want to force a new page (e.g., separating chapters or major report sections for PDF output).
• Never use === by default. Only add page breaks when the user specifically requests them.

Default Sizes

Recommended Size Pairings

Example: Dense KPI Row

```big_value size=[3,2]
{"value": 1250000, "label": "Revenue", "format": "currency0m"}
```
```big_value size=[3,2]
{"value": 8450, "label": "Orders", "format": "num0k"}
```
```big_value size=[3,2]
{"value": 2400000000, "label": "Queries", "format": "num0b"}
```
```delta size=[3,2]
{"value": 0.15, "label": "MoM", "format": "pct0"}
```
```delta size=[4,2]
{"value": 0.08, "label": "vs Target", "format": "pct0"}
```

This creates a row with 5 KPIs (3+3+3+3+4 = 16 columns).

Example: Two Charts Side by Side

```bar size=[8,6] file=data/region-sales.json
```
```line size=[8,6] file=data/monthly-trend.json
```

Supported Types

Charts: bar, line, area, pie, scatter, bubble, boxplot, histogram, waterfall, xmr, sankey, funnel, heatmap, calendar, sparkline, combo, dumbbell, mermaid

UI Components: big_value, delta, alert, note, text, textarea, empty_space, table

Table Formatting

Tables support column-level and cell-level formatting:

Column options: bold, italic, type ("sparkline" or "heatmap")

{
  "type": "table",
  "columns": [
    {"id": "product", "title": "Product", "bold": true},
    {"id": "category", "title": "Category", "italic": true},
    {"id": "sales", "title": "Sales", "fmt": "currency"},
    {"id": "margin", "title": "Margin", "type": "heatmap", "fmt": "pct"},
    {"id": "trend", "title": "Trend", "type": "sparkline", "sparkType": "line"}
  ],
  "data": [
    {"product": "Widget", "category": "Electronics", "sales": 125000, "margin": 0.85, "trend": [85, 92, 88, 95, 102, 125]}
  ]
}

Cell-level overrides: Use {"value": "text", "bold": true} to override column defaults.

Heatmap: Applies color gradient from low to high values. Text auto-switches to white on dark backgrounds.

Sparkline types: line, bar, area, pct_bar (progress bar), dumbbell (before/after comparison)

Sort & filter:

Example: {"type": "table", "filter": true, "columns": [...], "data": [...]}

Note Types

Notes support three severity levels via noteType:

Notes also support an optional label for bold prefix text:

{"type": "note", "label": "Pro Tip:", "content": "Use keyboard shortcuts for faster navigation.", "noteType": "tip"}

Specialized Chart Examples

big_value - Hero metrics with large display:

{"type": "big_value", "value": 1250000, "label": "Revenue", "format": "currency0m"}

• Optional comparison object: {"value": 10300, "format": "currency", "label": "vs last month"} shows change with arrow

dumbbell - Before/after comparisons with directional coloring:

{
  "type": "dumbbell",
  "title": "ELO Changes",
  "category": "team",
  "start": "before",
  "end": "after",
  "startLabel": "Week 1",
  "endLabel": "Week 2",
  "higherIsBetter": true,
  "data": [
    {"team": "Chiefs", "before": 1650, "after": 1720},
    {"team": "Bills", "before": 1600, "after": 1550}
  ]
}

• Green = improvement, Red = decline, Grey = no change
• higherIsBetter: false for rankings (lower = better)
• Labels auto-abbreviate large numbers (7450 → "7k")

delta - Change metrics with directional coloring:

{"type": "delta", "value": 0.15, "label": "MoM Growth", "format": "pct0"}

• Positive values show green with ▲, negative show red with ▼
• Optional comparison object: {"value": 0.05, "label": "vs Target"}

area - Filled line chart for cumulative/volume data:

{
  "type": "area",
  "title": "Daily Active Users",
  "x": "date",
  "y": "users",
  "data": [{"date": "Mon", "users": 1200}, {"date": "Tue", "users": 1450}]
}

combo - Bar + line with dual Y-axis:

{
  "type": "combo",
  "title": "Revenue vs Growth Rate",
  "x": "quarter",
  "y": ["revenue", "growth_rate"],
  "data": [
    {"quarter": "Q1", "revenue": 1000000, "growth_rate": 0.15},
    {"quarter": "Q2", "revenue": 1200000, "growth_rate": 0.20}
  ]
}

• First y-field renders as bars, second as line
• Dual Y-axes with independent scales

heatmap - 2D matrix visualization:

{
  "type": "heatmap",
  "title": "Activity by Hour",
  "xCategories": ["Mon", "Tue", "Wed", "Thu", "Fri"],
  "yCategories": ["9am", "12pm", "3pm", "6pm"],
  "format": "num0",
  "data": [[0, 0, 85], [1, 0, 90], [2, 0, 72]]
}

• format option applies to cell labels (e.g., num0k, currency0k, pct)

funnel - Conversion or elimination flows:

{
  "type": "funnel",
  "title": "Sales Pipeline",
  "format": "num0",
  "data": [
    {"stage": "Leads", "value": 1000},
    {"stage": "Qualified", "value": 600},
    {"stage": "Proposal", "value": 300},
    {"stage": "Closed", "value": 100}
  ]
}

• format option applies to labels/tooltips (e.g., currency_auto, pct, num0)

waterfall - Cumulative change visualization:

{
  "type": "waterfall",
  "title": "Revenue Bridge",
  "x": "item",
  "y": "value",
  "data": [
    {"item": "Start", "value": 1000, "isTotal": true},
    {"item": "Growth", "value": 200},
    {"item": "Churn", "value": -50},
    {"item": "End", "value": 1150, "isTotal": true}
  ]
}

bubble - Scatter with size dimension. Supports series for color grouping and showLabels for persistent labels:

{
  "type": "bubble",
  "title": "Market Analysis",
  "x": "growth",
  "y": "profit",
  "size": "revenue",
  "series": "region",
  "label": "company",
  "data": [
    {"growth": 5, "profit": 20, "revenue": 100, "region": "US", "company": "Acme"},
    {"growth": 10, "profit": 15, "revenue": 200, "region": "EU", "company": "Beta"}
  ]
}

sankey - Flow diagrams showing relationships:

{
  "type": "sankey",
  "title": "Traffic Sources",
  "data": [
    {"source": "Organic", "target": "Landing", "value": 500},
    {"source": "Paid", "target": "Landing", "value": 300},
    {"source": "Landing", "target": "Signup", "value": 400}
  ]
}

mermaid - Diagrams from Mermaid syntax (flowcharts, sequence, state, class, ER). Use array for multi-line code:

{
  "type": "mermaid",
  "title": "User Flow",
  "code": [
    "graph TD",
    "  A[Start] --> B{Decision}",
    "  B -->|Yes| C[Action]",
    "  B -->|No| D[End]"
  ]
}

mermaid (ASCII) - ASCII/Unicode text-based diagrams (set ascii: true):

{
  "type": "mermaid",
  "title": "Process Flow",
  "code": ["graph LR", "  A[Input] --> B[Process] --> C[Output]"],
  "ascii": true
}

Mermaid lint rules (errors that will fail validation):

• No <br/> tags in labels (render as literal text, not line breaks)
• No quoted labels like A["text"] in flowcharts (quotes appear in output)

Number Format Options

Percentage formats: pct, pct0, pct1 always multiply by 100 for scalar components (big_value, delta) — pass 0.15 for 15%. Charts and tables are series-aware: they inspect the column/series and skip the multiply when any value exceeds 1, so [15, 22, 18] renders as 15.0%, 22.0%, 18.0% (not 1500%+) while [0.15, 0.22, 0.18] still renders as 15.0%, 22.0%, 18.0%. Pass [0.8, 1.2] for a growth-rate series and it'll render as 80%, 120%. If big_value / delta get a pct value > 1, the linter warns "did you mean value/100?".

Smart formatting (auto/currency_auto) is recommended. The format option applies to both axis labels and data labels on bar charts. It automatically picks the right suffix (k, m, b) based on magnitude and always shows 4 significant digits. Negative values are wrapped in parentheses: (1.000m).

When no format is specified, smart formatting is used by default.

Auto-Detected Axis Formatting

Chart axes automatically detect the appropriate format based on field names:

Override with an explicit format field in the chart spec.

Columnar Data Format

The chart generator auto-detects columnar query results. Instead of manually converting columns/rows to data, pass the result directly:

{
  "type": "bar",
  "title": "Sales by Region",
  "x": "region",
  "y": "sales",
  "columns": ["region", "sales"],
  "rows": [["North", 45000], ["South", 32000], ["East", 28000]]
}

This is automatically converted internally. No manual JSON reconstruction needed.

Axis Bounds (yMin/yMax)

For line, area, bar, and combo charts, control y-axis range with yMin and yMax:

{
  "type": "line",
  "title": "Elo Rating Trend",
  "x": "date",
  "y": "elo",
  "yMin": 1400,
  "data": [{"date": "Oct", "elo": 1511}, {"date": "Jan", "elo": 1636}]
}

Use yMin when:

• Data doesn't start at 0 (ratings, stock prices, temperatures)
• You want to emphasize relative changes over absolute values

Use yMax when:

• Labels are being cut off at the top of the chart
• You need headroom above the highest data point

Validation & Lint Rules

The CLI validates specs automatically using built-in lint rules. Use --lint flag for validation-only mode:

npx -y -q mviz --lint dashboard.md  # Validate without generating HTML

Lint Rules

Errors exit with code 1. Warnings log to stderr but don't fail.

Common Fixes

Time series error: Sort your data by date before passing to the chart.

Sankey wrong keys: Use source, target, value in your data:

{"source": "A", "target": "B", "value": 100}

big_value string: Pass numeric value with format option:

{"type": "big_value", "value": 0.625, "format": "pct0", "label": "Rate"}

Troubleshooting

Warning Messages

The generator outputs helpful warnings to stderr when issues are detected:

Warnings include context like content previews, similar type suggestions, and section/row info.

Labels Cut Off at Chart Edges

If data labels on bar, line, or area charts are being cut off at the top:

1. Find the maximum value in your data
2. Set yMax to ~10-15% higher than that value

Example: If max value is 200, set "yMax": 220

{
  "type": "bar",
  "title": "Sales",
  "x": "month",
  "y": "sales",
  "yMax": 250,
  "data": [{"month": "Jan", "sales": 180}, {"month": "Feb", "sales": 220}]
}

This provides headroom for the label text above the bars.

Data Generation Best Practice

Use SQL to generate data files instead of manually authoring JSON. This reduces errors and ensures data accuracy:

-- Generate chart data file
COPY (
  SELECT month, SUM(sales) as sales, SUM(revenue) as revenue
  FROM orders
  GROUP BY month
  ORDER BY month
) TO 'data/monthly-sales.json' (FORMAT JSON, ARRAY true);

Then reference the generated file:

```bar file=data/monthly-sales.json
{"title": "Monthly Sales", "x": "month", "y": "sales"}
```

This approach:

• Ensures data accuracy (no manual transcription errors)
• Keeps data in sync with source systems
• Reduces token usage (SQL is more compact than JSON arrays)
• Makes updates easy (re-run query to refresh)

File References (JSON and CSV)

Reference external data files to save tokens and enable data/visualization separation:

JSON Files

```bar size=[8,6] file=data/sales.json
```

CSV Files (DuckDB Workflow)

CSV files work great with DuckDB for data exploration:

# Export query results to CSV
duckdb -csv -c "SELECT quarter, revenue FROM sales" > data/quarterly.csv

```bar file=data/quarterly.csv
{"title": "Quarterly Revenue", "x": "quarter", "y": "revenue"}
```

• CSV provides data, inline JSON provides chart options (title, x, y, format)
• Auto-detection: If no inline options, first column = x, second column = y
• Type conversion: Numeric strings auto-convert to int/float

Benefits of File References

Dashboard Markdown Format

---
theme: light
title: My Dashboard
---

# Page Title

## Section Name

```big_value size=[4,2]
{"value": 125000, "label": "Revenue", "format": "currency0k"}
```
```bar size=[12,6] file=data/sales.json
```

Rules:

• # Title sets the page title (first occurrence only)
• ## Section creates a new section with divider (border, spacing)
• ### Header creates a soft header within the current section (no divider)
• --- creates a section break (untitled, visual divider only)
• === creates a page break (forces new page when printing to PDF)
• size=[cols,rows] controls layout (16-column grid)
• size=auto auto-calculates size from data
• file=path references external JSON
• Empty lines = new rows

Theme Toggle

Dashboards include a theme toggle button (top right) that switches between light and dark modes. All charts dynamically update when the theme changes.

Set the default theme in frontmatter:

---
title: My Dashboard
theme: dark
orientation: landscape
print: true
---

Page capacity: Portrait fits 30 row units, landscape fits 22 row units (Letter paper, 0.5" margins).

The theme toggle affects all charts globally - individual chart theme settings are ignored in favor of the global toggle.

Renderer: visualize:show_widget (inline)

The show_widget renderer lives inside Claude. Claude reads the mviz spec, translates it to a widget_code HTML fragment using claude.ai's design system (Chart.js + var(--color-*) tokens), and passes it to visualize:show_widget. The widget appears inline in the conversation.

Workflow

1. Compose the mviz markdown spec and emit it visibly in your response, in a fenced code block. The spec is the canonical artifact — even when the user only wants an inline render, write the spec down so it's available for later (copy / reuse / pipe-to-CLI). Then render below it.

   ```mviz
   ---
   title: Q3 sales
   ---
   ```big_value
   {"value": 1250000, "label": "Q3 Revenue", "format": "currency_auto"}
   

   ...

2. State your renderer inference, one short sentence. Example: "Rendering inline since you asked to see it in chat — say the word if you want the file too."

3. Translate the spec to widget_code following the mapping tables below. Use claude.ai's design tokens (var(--color-*)), Chart.js for charts, metric cards for KPIs.

4. Call visualize:show_widget with the widget_code.

5. Write the surrounding explanation as normal response text, outside the tool call. The widget is the visual; your prose carries the meaning.

When this renderer is the right pick

Trigger inline rendering when all of these hold:

• The user is in claude.ai web/desktop (not Claude Code, not the API).
• visualize:show_widget is in your available tools.
• The visualization is composed of mviz component types that have chat-native equivalents (see mapping below).
• The visualization is small enough to emit naturally as widget_code — model emit time dominates latency, so 6–8 components is comfortable, 15+ starts to feel slow.

If any of those fails — mviz-only chart types like sankey/heatmap, very large dashboards, no show_widget — render through the mviz CLI instead and offer the file as a follow-up.

What the show_widget renderer is NOT

• Not the mviz CLI. Don't invoke npx mviz to produce HTML for show_widget. The CLI's mdsinabox-styled output doesn't match the chat host, and the subprocess + larger emit costs latency. The CLI is a peer renderer used for standalone files.
• Not a paste-mviz-output-into-widget_code path. mviz CLI emits a complete document with its own design system; that wouldn't render correctly inside the widget sandbox even if show_widget accepted documents (it doesn't — it requires fragments).

The translation: mviz spec → claude.ai-native HTML

mviz markdown describes structure. Read the spec, then write a widget_code HTML fragment using claude.ai's design system.

Component type mapping

Layout mapping

mviz uses a 16-col grid with explicit size=[cols, rows] directives. Inside a 680px-wide show_widget container, that translates to:

.mviz-grid {
  display: grid;
  grid-template-columns: repeat(16, 1fr);
  gap: 8px;
}
.mviz-grid > .item-12 { grid-column: span 12; }
.mviz-grid > .item-8  { grid-column: span 8; }
.mviz-grid > .item-4  { grid-column: span 4; }

For simple 2–3 component widgets you can also use claude.ai's recommended responsive pattern:

display: grid;
grid-template-columns: repeat(auto-fit, minmax(160px, 1fr));
gap: 12px;

This auto-fits to the host's chosen width, which is the right choice when the exact column count is not load-bearing.

Format hint mapping

mviz format strings → JS formatters used inside Chart.js label/tooltip callbacks and metric-card text:

The currency_auto / auto helper, mirroring mviz's smartFormatNumber. ~4 significant digits, magnitude tier with suffix, negatives in parentheses — match these exact thresholds so the inline render and the standalone report show identical numbers:

function fmtAuto(v, sym = '') {
  const neg = v < 0, abs = Math.abs(v);
  let s;
  if (abs >= 1e9) {
    const x = abs / 1e9;
    s = sym + (x >= 100 ? x.toFixed(1) : x >= 10 ? x.toFixed(2) : x.toFixed(3)) + 'b';
  } else if (abs >= 1e6) {
    const x = abs / 1e6;
    s = sym + (x >= 100 ? x.toFixed(1) : x >= 10 ? x.toFixed(2) : x.toFixed(3)) + 'm';
  } else if (abs >= 1e4) {
    const x = abs / 1e3;
    s = sym + (x >= 100 ? x.toFixed(1) : x.toFixed(2)) + 'k';
  } else if (abs >= 1e3) {
    s = sym + abs.toLocaleString('en-US', { maximumFractionDigits: 0 });
  } else {
    s = sym + (abs === Math.floor(abs) ? abs : abs.toFixed(2));
  }
  return neg ? '(' + s + ')' : s;
}
// currency_auto: fmtAuto(v, '$')      → 1_250_000 → "$1.250m", 12_500 → "$12.50k", 1_250 → "$1,250"
// auto:          fmtAuto(v)            → 1_250_000 →  "1.250m", 12_500 →  "12.50k", 1_250 →  "1,250"

Always Math.round / .toFixed(n) displayed numbers — 0.1 + 0.2 is 0.30000000000000004 and that artifact will leak into the widget if you don't.

Color mapping

Chart.js renders to <canvas> and cannot resolve CSS variables. Use hardcoded hex from claude.ai's color ramps (the c-{ramp} stops listed in the visualize:read_me design module). A safe default palette for mviz series:

• Primary (mviz blue #0777b3) → c-blue 600 #185FA5
• Secondary (mviz orange #bd4e35) → c-coral 600 #993C1D
• Positive (mviz green #2d7a00) → c-green 600 #3B6D11
• Warning (mviz amber #e18727) → c-amber 600 #854F0B
• Error (mviz red #bc1200) → c-red 600 #A32D2D
• Tertiary / neutral → c-gray 600 #5F5E5A

For HTML elements (cards, table borders, text), use var(--color-text-primary) / var(--color-background-secondary) / var(--color-border-tertiary) and friends — they auto-adapt to dark mode.

Chart.js setup (canonical patterns from visualize:read_me)

<script src="https://cdnjs.cloudflare.com/ajax/libs/Chart.js/4.4.1/chart.umd.js"></script>

Wrap every canvas in a sized div:

<div style="position: relative; width: 100%; height: 240px;">
  <canvas id="c1" role="img" aria-label="Sales by region — NA $540k, EU $380k, APAC $220k, LATAM $110k">
    Sales by region (bar chart). NA 540, EU 380, APAC 220, LATAM 110.
  </canvas>
</div>

Constraints to honor:

• responsive: true, maintainAspectRatio: false — height goes on the wrapper div only, never on the canvas itself.
• Horizontal bar charts: wrapper height ≥ (bars * 40) + 80 px.
• ≤12 categories where every label must be visible: scales.x.ticks: { autoSkip: false, maxRotation: 45 }.
• Bubble/scatter: pad scales.{x,y}.{min,max} ~10 % beyond data range so radii don't clip.
• Negative currency: -$5M not $-5M — sign before symbol. Use a formatter (v) => (v < 0 ? '-' : '') + '$' + Math.abs(v) + 'M'.

Default Chart.js legends use round dots and no values — disable them and emit your own legend as inline HTML (small color squares, tight spacing, include the value/percentage when categorical):

plugins: { legend: { display: false } }

<div style="display: flex; flex-wrap: wrap; gap: 16px; margin-bottom: 8px;
            font-size: 12px; color: var(--color-text-secondary);">
  <span style="display: flex; align-items: center; gap: 4px;">
    <span style="width: 10px; height: 10px; border-radius: 2px; background: #185FA5;"></span>NA $540k
  </span>
  <!-- repeat -->
</div>

Metric card (for big_value and delta)

<div style="background: var(--color-background-secondary); border-radius: var(--border-radius-md);
            padding: 1rem;">
  <div style="font-size: 13px; color: var(--color-text-secondary);">Q3 Revenue</div>
  <div style="font-size: 24px; font-weight: 500; color: var(--color-text-primary);
              line-height: 1; margin-top: 4px;">$1.250m</div>
</div>

Use in grids of 2–4 with gap: 12px. For delta, add a small directional indicator with color: var(--color-success) or var(--color-danger).

Worked example

Spec (mviz markdown):

---
title: Q3 sales
---

```big_value size=[4,2]
{"value": 1250000, "label": "Q3 Revenue", "format": "currency_auto"}
```
```big_value size=[4,2]
{"value": 0.184, "label": "Growth", "format": "pct1"}
```

```bar size=[16,5]
{"title": "Sales by region", "x": "region", "y": "sales", "format": "currency0k",
 "data": [{"region":"NA","sales":540000},{"region":"EU","sales":380000},
          {"region":"APAC","sales":220000},{"region":"LATAM","sales":110000}]}
```

Inline render (the widget_code you pass to visualize:show_widget):

<h2 class="sr-only">Q3 sales overview: $1.250m revenue, 18.4% growth, regional breakdown.</h2>
<div style="display: grid; grid-template-columns: repeat(16, 1fr); gap: 12px;">
  <div style="grid-column: span 4; background: var(--color-background-secondary);
              border-radius: var(--border-radius-md); padding: 1rem;">
    <div style="font-size: 13px; color: var(--color-text-secondary);">Q3 Revenue</div>
    <div style="font-size: 24px; font-weight: 500; margin-top: 4px;">$1.250m</div>
  </div>
  <div style="grid-column: span 4; background: var(--color-background-secondary);
              border-radius: var(--border-radius-md); padding: 1rem;">
    <div style="font-size: 13px; color: var(--color-text-secondary);">Growth</div>
    <div style="font-size: 24px; font-weight: 500; margin-top: 4px;
                color: var(--color-success);">+18.4%</div>
  </div>
  <div style="grid-column: span 16;">
    <div style="font-size: 14px; font-weight: 500; margin-bottom: 8px;">Sales by region</div>
    <div style="position: relative; width: 100%; height: 240px;">
      <canvas id="mviz-bar-1" role="img"
              aria-label="Sales by region — NA $540k, EU $380k, APAC $220k, LATAM $110k">
        NA 540, EU 380, APAC 220, LATAM 110.
      </canvas>
    </div>
  </div>
</div>
<script src="https://cdnjs.cloudflare.com/ajax/libs/Chart.js/4.4.1/chart.umd.js"></script>
<script>
  new Chart(document.getElementById('mviz-bar-1'), {
    type: 'bar',
    data: {
      labels: ['NA', 'EU', 'APAC', 'LATAM'],
      datasets: [{ label: 'Sales', data: [540000, 380000, 220000, 110000], backgroundColor: '#185FA5' }],
    },
    options: {
      responsive: true, maintainAspectRatio: false,
      plugins: { legend: { display: false } },
      scales: {
        y: { ticks: { callback: (v) => '$' + Math.round(v/1000) + 'k' } },
        x: { ticks: { autoSkip: false, maxRotation: 45 } },
      },
    },
  });
</script>

Constraints + things you must not do

• Do not invoke the mviz CLI for show_widget output. This path renders without it. The CLI is only for standalone HTML reports.
• Do not paste mviz's own HTML into show_widget. mviz emits its own styled fragments — they will not match the chat host and the bytes are wasted.
• No <!DOCTYPE>, <html>, <head>, or <body> in the widget — show_widget rejects documents.
• No position: fixed. It collapses the iframe's auto-sized viewport.
• Only load from the allowlist: cdnjs.cloudflare.com, esm.sh, cdn.jsdelivr.net, unpkg.com, fonts.googleapis.com, fonts.gstatic.com. Chart.js from cdnjs.cloudflare.com is the canonical source.
• Tables with more than ~3 columns or any prose-y content: render as markdown in your response text instead of inside the widget. The visualize:read_me design guide is explicit about this.

Falling back to the CLI renderer

When the inline renderer isn't a fit, the spec is still useful — pipe it through the CLI:

• Chart type with no Chart.js equivalent (sankey, heatmap, calendar, etc.) — render via the CLI, offer the file. Mention that the mviz design system kicks in for the standalone output.
• visualize:show_widget is not available — render via the CLI and present as an artifact.
• User is in Claude Code — render via the CLI, write the file path, they'll open it in a browser. No inline widget surface in the TUI.
• Visualization exceeds the inline size budget (15+ components, dense tables with sparklines, multi-section reports) — render via the CLI. Offer inline as a follow-up if they want a smaller summary view.
• Single tiny visual (one sparkline, one 3-cell metric) — consider hand-rolling SVG directly via show_widget without going through the mviz spec at all. mviz earns its keep when composing multiple components; for a one-off icon it's overkill.

Custom Themes

Load custom brand colors and fonts from a YAML file:

npx -y -q mviz --theme my_theme.yaml dashboard.md -o dashboard.html

Example theme file:

name: 

…(truncated)