nteract/semiotic

A React data visualization library designed for AI-assisted development.

Simple charts in 5 lines. Network graphs, streaming data, and coordinated dashboards when you need them. Structured schemas and an MCP server so AI coding assistants generate correct chart code on the first try.

What's New in 3.8.2

3.8.2 makes the hosted MCP accessibility contract internally consistent:

• Chart schemas, validation, repair guidance, and audits consistently expose title, description, summary, and accessibleTable.
• Hosted getChartSchema returns structured output for component lists, schemas, and errors, including direct and ChartContainer accessibility guidance.
• The public MCP profile remains a focused five-tool surface with render-proven createChart output.

import { LineChart } from "semiotic/xy"

<LineChart
  data={salesData}
  xAccessor="month"
  yAccessor="revenue"
/>

Why Semiotic

Semiotic is a data visualization library for React that combines broad chart coverage with first-class AI tooling. It handles the chart types that most libraries skip — network graphs, streaming data, statistical distributions, coordinated views — and ships with machine-readable schemas so LLMs can generate correct code without examples.

Built for AI-assisted development

Semiotic ships with everything an AI coding assistant needs to generate correct visualizations without trial and error:

• semiotic/ai — a single import with the schema-backed chart capability catalog (XY, ordinal, network, realtime, geo, value), optimized for LLM code generation. See ai/surface-manifest.json for the generated current inventory. Note: the published entry files are pre-bundled, so importing one chart from semiotic/ai still ships most of the bundle — treat it as a codegen/tooling surface and use family subpaths (semiotic/xy, semiotic/geo, semiotic/value, …) in production code, at roughly half the single-chart cost.
• ai/schema.json — machine-readable prop schemas for every component
• npx semiotic-mcp — an MCP server for tool-based chart rendering in any MCP client
• npx semiotic-ai --doctor — validate component + props JSON from the command line with typo suggestions and anti-pattern detection
• diagnoseConfig(component, props) — programmatic anti-pattern detector with actionable fixes, spanning validation, encoding, accessibility, and misleading-design (deception) checks
• CLAUDE.md — instruction files auto-synced for Claude, Cursor, Copilot, Windsurf, and Cline
• llms.txt — machine-readable documentation following the emerging standard

Every chart includes a built-in error boundary, dev-mode validation warnings with typo suggestions, and accessibility features (canvas aria-label, keyboard-navigable legends, aria-live tooltips, SVG <title>/<desc>) so AI-generated code fails gracefully with actionable diagnostics instead of a blank screen.

Beyond standard charts

Network visualization. Force-directed graphs, Sankey diagrams, chord diagrams, tree layouts, treemaps, circle packing, and orbit diagrams — all as React components with the same prop API as LineChart.

Streaming data. Realtime charts render on canvas at 60fps with a ref-based push API. Built-in decay, pulse, and staleness encoding for monitoring dashboards.

Coordinated views. LinkedCharts provides hover cross-highlighting, brush cross-filtering, coordinate-based linked crosshairs, and selection synchronization across any combination of chart types — zero wiring.

Geographic visualization. Choropleth maps, proportional symbol maps, flow maps with animated particles, and distance cartograms — all canvas-rendered with d3-geo projections, zoom/pan, tile basemaps, and drag-rotate globe spinning.

Statistical summaries. Box plots, violin plots, swarm plots, histograms, LOESS smoothing, forecast with confidence envelopes, and anomaly detection. Marginal distribution graphics on scatterplot axes with a single prop.

First-class annotations. Annotations are data-bound objects, not post-hoc artwork. Labels, callouts, thresholds, enclosures, statistical overlays, and React widgets move with the chart and render through browser, SSR, and export paths. Opt into placement, hierarchy, density, progressive disclosure, audience-aware amount, provenance, and editorial lifecycle when the chart needs to communicate more than its encoding alone.

Start simple, go deep

Every Chart component accepts a frameProps prop to access the underlying Frame API without leaving the simpler interface.

Serialization and interop

Charts serialize to JSON and back: toConfig, fromConfig, toURL, copyConfig, configToJSX. Have Vega-Lite specs? fromVegaLite(spec) translates them to Semiotic configs — works with configToJSX() for full round-trip from notebooks and AI-generated specs.

Need an external pitfall review? The experimental unstable_toDataPitfallsChain() builds a dependency-free chain input for datapitfalls, combining the Semiotic config, JSX, reader grounding, diagnostics, accessibility audit, and optional rendered SVG/image evidence:

import { unstable_toDataPitfallsChain } from "semiotic/experimental"
import { detectPitfalls } from "datapitfalls"

const input = unstable_toDataPitfallsChain("LineChart", props, {
  narrative: "Monthly sales are accelerating.",
  rendered: { svg, evidence },
})

const report = await detectPitfalls(input, { apiKey: process.env.ANTHROPIC_API_KEY })

The return path stays dependency-free too. Use whole-chart findings as ChartContainer notifications, and only turn findings into annotations after your app can anchor them to marks or semantic positions:

import { ChartContainer } from "semiotic"
import { LineChart } from "semiotic/xy"
import {
  unstable_toDataPitfallsAnnotations,
  unstable_toDataPitfallsNotifications,
} from "semiotic/experimental"

const notifications = unstable_toDataPitfallsNotifications(report)
const annotations = unstable_toDataPitfallsAnnotations(report, {
  anchorFor: (finding) =>
    finding.ruleId === "truncated-axis" ? { x: 9, y: 9000 } : null,
})

<ChartContainer notifications={notifications}>
  <LineChart {...props} annotations={annotations} />
</ChartContainer>

When to use something else

Need a standard bar or line chart for a dashboard you'll never need to customize beyond colors and labels? Recharts has a larger ecosystem and more community examples. Need GPU-accelerated rendering for millions of data points? Apache ECharts handles that scale.

Semiotic is for projects that outgrow those libraries — when you need network graphs alongside time series, streaming data alongside static snapshots, or coordinated views across chart types.

Install

npm install semiotic

Requires React 18.1+ or React 19.

Quick Examples

Coordinated Dashboard

Hover one chart, highlight the same data in another — zero wiring:

import { LinkedCharts, Scatterplot, BarChart } from "semiotic"

<LinkedCharts>
  <Scatterplot
    data={data} xAccessor="age" yAccessor="income" colorBy="region"
    linkedHover={{ name: "hl", fields: ["region"] }}
    selection={{ name: "hl" }}
  />
  <BarChart
    data={summary} categoryAccessor="region" valueAccessor="total"
    selection={{ name: "hl" }}
  />
</LinkedCharts>

Streaming Metrics with Decay

Live data fades old points, flashes new ones, flags stale feeds:

import { RealtimeLineChart } from "semiotic"

const chartRef = useRef()
chartRef.current.push({ time: Date.now(), value: cpuLoad })

<RealtimeLineChart
  ref={chartRef}
  timeAccessor="time"
  valueAccessor="value"
  decay={{ type: "exponential", halfLife: 100 }}
  staleness={{ threshold: 5000, showBadge: true }}
/>

Network Graphs

Force-directed graphs and Sankey diagrams — same API as LineChart:

import { ForceDirectedGraph, SankeyDiagram } from "semiotic"

<ForceDirectedGraph
  nodes={people} edges={friendships}
  colorBy="team" nodeSize={8} showLabels
/>

<SankeyDiagram
  edges={budgetFlows}
  sourceAccessor="from" targetAccessor="to" valueAccessor="amount"
/>

Geographic Visualization

Choropleth maps, flow maps, and distance cartograms with canvas rendering, zoom/pan, tile basemaps, and animated particles:

import { ChoroplethMap, FlowMap, DistanceCartogram } from "semiotic/geo"

<ChoroplethMap
  areas={geoJsonFeatures} valueAccessor="gdp"
  colorScheme="viridis" projection="equalEarth" zoomable tooltip
/>

<FlowMap
  nodes={airports} flows={routes} valueAccessor="passengers"
  showParticles particleStyle={{ color: "source", speedMultiplier: 1.5 }}
/>

<DistanceCartogram
  points={cities} center="rome" costAccessor="travelDays"
  showRings costLabel="days" lines={routes}
/>

Streaming System Monitor

Live service topology with threshold alerting and click-to-inspect:

import { StreamNetworkFrame, ChartContainer, DetailsPanel, LinkedCharts } from "semiotic"

const chartRef = useRef()
chartRef.current.push({ source: "API", target: "Orders", value: 15 })

<LinkedCharts>
  <ChartContainer title="System Monitor" status="live"
    detailsPanel={
      <DetailsPanel position="right" trigger="click">
        {(datum) => <div>{datum.id}: {datum.value} req/s</div>}
      </DetailsPanel>
    }>
    <StreamNetworkFrame ref={chartRef} chartType="sankey"
      showParticles particleStyle={{ proportionalSpeed: true }}
      thresholds={{ metric: n => n.value, warning: 100, critical: 250 }}
    />
  </ChartContainer>
</LinkedCharts>

Standard Charts

Line, bar, scatter, area — all the basics, with sensible defaults:

import { LineChart, BarChart } from "semiotic"

<LineChart
  data={salesData}
  xAccessor="month" yAccessor="revenue"
  curve="monotoneX" showPoints
/>

<BarChart
  data={categoryData}
  categoryAccessor="department" valueAccessor="sales"
  orientation="horizontal" colorBy="region"
/>

All Chart Components

Vega-Lite Translation

Paste a Vega-Lite spec, get a Semiotic chart:

import { fromVegaLite } from "semiotic/data"
import { configToJSX, fromConfig } from "semiotic"

const config = fromVegaLite({
  mark: "bar",
  data: { values: [{ a: "A", b: 28 }, { a: "B", b: 55 }] },
  encoding: {
    x: { field: "a", type: "nominal" },
    y: { field: "b", type: "quantitative" },
  },
})

// Render directly
const { componentName, props } = fromConfig(config)
// → componentName: "BarChart", props: { data, categoryAccessor: "a", valueAccessor: "b" }

// Or generate JSX code
configToJSX(config)
// → <BarChart data={[...]} categoryAccessor="a" valueAccessor="b" />

Supports bar, line, area, point, rect, arc, tick marks with encoding translation for color, size, aggregation, and binning.

Conversation Arc Telemetry

Capture and replay the path an AI-assisted chart session took:

import {
  createLocalStorageConversationArcSink,
  enableConversationArc,
  getConversationArcStore,
  loadConversationArc,
  registerConversationArcSink,
} from "semiotic/ai"

const sink = createLocalStorageConversationArcSink({ key: "my-app:arc" })
registerConversationArcSink(sink)
enableConversationArc({ sessionId: "session-abc" })

getConversationArcStore().record({ type: "chart-rendered", component: "LineChart" })
loadConversationArc(sink.load(), { enabled: false })

Bundle Sizes

Semiotic ships 17 stable module entry points. Don't import from "semiotic" unless you need everything — use the sub-path that matches your chart type.

The numbers below are first-party artifact cost: the gzip size of Semiotic's own code for each sub-path. They exclude React and other runtime dependencies, so they are not a prediction of a cold application bundle. Do not add artifact rows to estimate an app: dependency resolution and cross-import deduplication happen in the consumer bundler and are measured separately below.

Cold-consumer named imports

The table above is first-party artifact cost, not an application bundle. The generated table below measures a different thing: a fresh consumer bundles one retained named import from a packed semiotic tarball through the public export path. It includes Semiotic and its resolved runtime dependencies, but externalizes React/React DOM and optional adapter peers that the host application owns. Each row starts cold, so use it to compare one public import choice—not to add together an application's rows. The checked machine-readable baseline is benchmarks/setup/cold-consumer-imports.json; refresh it after a production build with npm run docs:cold-consumer.

Method: fresh npm pack --ignore-scripts tarball → temporary consumer → minified/tree-shaken esbuild ESM bundle → gzip -9. React/React DOM and optional adapter peers are external; Semiotic and its resolved runtime dependencies are included.

// Import from the sub-path, not from "semiotic"
import { LineChart } from "semiotic/xy"
import { BarChart } from "semiotic/ordinal"
import { SankeyDiagram } from "semiotic/network"
import { ChoroplethMap } from "semiotic/geo"

Tree-shaking & multi-subpath imports: Family entries (semiotic/xy, semiotic/network, semiotic/ai, …) are built as one ESM graph with shared chunks. Stream frames, renderers, and other common code ship once and are imported by every entry that needs them — so combining semiotic/ai + semiotic/xy + semiotic/network does not mean paying for three full copies of the runtime. The package is marked "sideEffects": false, so modern bundlers keep only the named exports you retain (e.g. LineChart + suggestCharts). Prefer family subpaths for clarity; import AI helpers from semiotic/ai or the lighter semiotic/ai/core when you do not need the chart catalog.

When to use "semiotic": Fine when you want one import for mixed families. With shared chunks it is no longer “sum every subpath” — the cold-consumer table above is the better guide for a single named import, and multi-family apps should not see the old double-pay tax.

TypeScript

Built with strict: true. Full type definitions ship with the package. Generics for type-safe accessors:

interface Sale { month: number; revenue: number }

<LineChart<Sale>
  data={sales}
  xAccessor="month"    // TS validates this is keyof Sale
  yAccessor="revenue"
/>

Server-Side Rendering

All chart components render SVG automatically in server environments — no special imports or configuration needed:

// Works in Next.js App Router, Remix, Astro — same component, same props
import { LineChart } from "semiotic"

// Server: renders <svg> with path/circle/rect elements
// Client: renders <canvas> with SVG overlay for axes
<LineChart data={data} xAccessor="date" yAccessor="value" />

For standalone SVG/PNG/GIF generation (email, OG images, PDF, Slack), use the server entry point:

import { renderChart, renderToImage, renderToAnimatedGif } from "semiotic/server"

// SVG — sync, no dependencies
const svg = renderChart("LineChart", {
  data, xAccessor: "date", yAccessor: "value",
  theme: "tufte", title: "Revenue Trend",
})

// PNG — async, requires sharp
const png = await renderToImage("BarChart", { data, ... }, { format: "png", scale: 2 })

// Animated GIF — async, requires sharp + gifenc
const gif = await renderToAnimatedGif("line", data, { ... }, { fps: 12 })

MCP Server

mcp-name: io.github.nteract/semiotic

Semiotic ships with an MCP server that lets AI coding assistants render charts, diagnose configuration problems, discover schemas, read packaged AI guidance, and get chart recommendations via tool calls.

Setup

Add to your MCP client config (e.g. claude_desktop_config.json for Claude Desktop):

{
  "mcpServers": {
    "semiotic": {
      "command": "npx",
      "args": ["semiotic-mcp"]
    }
  }
}

No API keys or authentication required. The server runs locally via stdio. HTTP mode is also available for inspectors, web clients, and ChatGPT Apps SDK experiments: npx semiotic-mcp --http --port 3001. It binds to 127.0.0.1 by default; intentionally expose another interface with --host 0.0.0.0 or MCP_HOST=0.0.0.0. Since 3.7.2, HTTP mode is stateless: each request gets a fresh read-only MCP server + transport, so it can autoscale on serverless hosts without sticky sessions.

For ChatGPT developer mode, expose the HTTP endpoint over HTTPS with a tunnel and create a connector that points at https://<your-tunnel>/mcp. The experimental Apps SDK surface is renderInteractiveChart, which returns a text/html;profile=mcp-app widget template plus a hidden SVG payload rendered by Semiotic on the MCP server.

For a hosted deployment, see deploy/cloud-run. The wrapper runs the published semiotic-mcp binary, exposes /mcp plus health endpoints, and supports MCP_ALLOWED_HOSTS for production host-header allowlisting. For ChatGPT Apps domain verification, set OPENAI_APPS_CHALLENGE_TOKEN so HTTP mode serves the raw token from /.well-known/openai-apps-challenge.

Tools

Resources

Prompts

Example: get schema for a component

Tool: getSchema
Args: { "component": "LineChart" }
→ Returns: { "name": "LineChart", "description": "...", "parameters": { "properties": { "data": ..., "xAccessor": ..., ... } } }

Example: suggest a chart for your data

Tool: suggestChart
Args: {
  "data": [
    { "month": "Jan", "revenue": 120, "region": "East" },
    { "month": "Feb", "revenue": 180, "region": "West" }
  ]
}
→ Returns:
  1. BarChart (high confidence) — categorical field (region) with values (revenue)
  2. StackedBarChart (medium confidence) — two categorical fields (month, region)
  3. DonutChart (medium confidence) — 2 categories — proportional composition

Example: render a chart

Tool: renderChart
Args: {
  "component": "BarChart",
  "props": {
    "data": [
      { "category": "Q1", "revenue": 120 },
      { "category": "Q2", "revenue": 180 },
      { "category": "Q3", "revenue": 150 }
    ],
    "categoryAccessor": "category",
    "valueAccessor": "revenue"
  }
}
→ Returns: <svg>...</svg>

Example: render a ChatGPT Apps widget

Tool: renderInteractiveChart
Args: {
  "component": "BarChart",
  "props": {
    "title": "Revenue by Quarter",
    "data": [
      { "quarter": "Q1", "revenue": 120 },
      { "quarter": "Q2", "revenue": 180 }
    ],
    "categoryAccessor": "quarter",
    "valueAccessor": "revenue"
  }
}
→ Returns: structured chart summary for the model + hidden SVG/widget metadata for ChatGPT.

Example: diagnose a broken config

Tool: diagnoseConfig
Args: { "component": "LineChart", "props": { "data": [] } }
→ Returns: ✗ [EMPTY_DATA] data is an empty array — Fix: provide at least one data point

Example: report an issue

Tool: reportIssue
Args: {
  "title": "Bug: BarChart tooltip shows undefined for custom accessor",
  "body": "When using valueAccessor='amount', tooltip displays 'undefined'.\n\ndiagnoseConfig output: ✓ no issues detected.",
  "labels": ["bug"]
}
→ Returns: Open this URL to submit the issue: https://github.com/nteract/semiotic/issues/new?...

CLI alternative

For quick validation without an MCP client:

npx semiotic-ai --list         # list components with import paths and renderability
npx semiotic-ai --list --json  # machine-readable component index
npx semiotic-ai --schema GaugeChart
npx semiotic-ai --suggest '{"data":[{"category":"A","value":10}],"intent":"comparison"}'
npx semiotic-ai --doctor       # validate component + props JSON
npx semiotic-ai --schema       # dump all chart schemas
npx semiotic-ai --compact      # compact schema (fewer tokens)

--doctor uses the full diagnoseConfig checks when dist is available and falls back to schema-only validation in clean source checkouts.

Where to find Semiotic for AI assistants

Semiotic is indexed by AI-coding-agent documentation tools so your assistant (Claude Code, Cursor, Cline, Copilot, etc.) can pull current docs and tools without copy-paste:

• Context7 — context7.com/nteract/semiotic (configured via context7.json)
• DeepWiki — deepwiki.com/nteract/semiotic
• GitMCP — gitmcp.io/nteract/semiotic (exposes the repo as an MCP endpoint directly)
• Official MCP Registry — search "semiotic" at registry.modelcontextprotocol.io
• Smithery — smithery.ai/server/nteract/semiotic

Agent-facing API surface:

• CLAUDE.md, ai/schema.json, ai/surface-manifest.json, ai/behaviorContracts.cjs — bundled in the npm tarball (see package.json#files); agents that install Semiotic locally read these directly. CLAUDE.md is the quick-start cheat sheet (HOC props, push API, theming, usage notes); ai/schema.json is the JSON Schema for every chart's prop surface; ai/surface-manifest.json is the generated inventory; ai/behaviorContracts.cjs carries the agent-visible semantic rules (color precedence, push-mode requirements, ID-accessor contracts).
• semiotic.nteract.io/llms.txt + /llms-full.txt — deployed at the docs site per the llms.txt standard. Agents fetch the navigation map (llms.txt) or the full inlined docs (llms-full.txt) over HTTP; they're not part of the npm package itself.

Documentation

Interactive docs and examples

• Getting Started
• Charts — chart types with live examples
• Frames — full Frame API reference
• Features — axes, tooltips, interaction, responsive behavior, and composition
• Annotations — first-class annotation types, design guidance, provenance, and lifecycle
• Cookbook — advanced patterns and recipes
• Playground — interactive prop exploration

Upgrading

• Migration Guide — upgrading from v1.x or v2.x
• Changelog — full release history

Contributing

See CONTRIBUTING.md. Our community follows the nteract Code of Conduct.

Acknowledgments

Development of this library owes a lot to Susie Lu, Jason Reid, James Womack, Matt Herman, Shelby Sturgis, and Tristan Reid.

The Sankey layout engine is based on sankey-plus by Tom Shanley, which improved on his earlier d3-sankey-circular with better cycle detection, hierarchical arc stacking, and dynamic extent adjustment.

Semiotic icon based on an icon by Andre Schauer.

License

Apache 2.0