swift-artifact

0.16.0

Swift/SwiftUI library for rendering LLM-generated <artifact> blocks in chat interfaces. Supports Markdown, JSON, CSV, Code, SVG, GeoJSON, HTML, React, Mermaid, LaTeX, Vega-Lite, GLTF, USDZ.
1amageek/swift-artifact

What's New

0.16.0

2026-05-22T15:05:14Z

Renders JSON-LD graph presentation styles through the native KnowledgeGraph view. Adds canvas backgrounds, node/group shapes and colors, edge and edge-label styling, scoped layout directives, and a JSON-LD-backed presentation preview. Updates swift-knowledge-graph to 0.3.0 and adds resolver/layout coverage.

swift-artifact

A Swift/SwiftUI library for displaying LLM-generated <artifact> blocks inside chat interfaces. Parses Claude-style artifact tags, models them as values, and renders each one through a pluggable renderer protocol — with first-class support for partial rendering while a model is still streaming.

  • Swift 6.3+, iOS / macOS / iPadOS / visionOS / Mac Catalyst 26+
  • One umbrella module (SwiftArtifact) or five fine-grained libraries — pick your granularity
  • Two-stage rendering model: every renderer declares a refine(_:) step that reduces the in-flight payload to a renderer-valid subset, so the body never sees half-formed input
  • Streaming-aware refiners for JSON, SVG, Mermaid, LaTeX, CSV, Markdown, GeoJSON, Turtle, TriG, N-Quads, RDF/XML, and JSON-LD — partial output is drawn as it arrives, not after the final token. Binary document and image renderers wait for a complete payload before decoding.
  • Environment-driven renderer registry — call .artifactRenderer(_:) once at the top of your view tree and let ArtifactView resolve the right renderer

See SPEC.md for the full specification.

Installation

// Package.swift
dependencies: [
    .package(url: "https://github.com/1amageek/swift-artifact.git", from: "0.14.1"),
]

The simplest option is the umbrella product SwiftArtifact, which re-exports every sub-module:

// Target dependency
.product(name: "SwiftArtifact", package: "swift-artifact"),
import SwiftArtifact   // grants access to every type and renderer below

If you want to pull in only a subset (e.g. you ship a Markdown-only client and don't want WebKit in your binary), depend on individual products instead:

.product(name: "ArtifactCore",            package: "swift-artifact"),
.product(name: "ArtifactRenderer",        package: "swift-artifact"),
.product(name: "ArtifactView",            package: "swift-artifact"),
.product(name: "ArtifactNativeRenderer",  package: "swift-artifact"),
.product(name: "ArtifactWebRenderer",     package: "swift-artifact"),

Modules

Module Depends on Purpose
SwiftArtifact All of the below Umbrella — @_exported re-export of every module
ArtifactCore ArtifactType, AnyArtifact, parsing
ArtifactRenderer Core ArtifactRenderable protocol, RefinedPayload, AnyArtifactRenderer
ArtifactView Core + Renderer ArtifactView, ArtifactCard, ArtifactCanvas, env registry
ArtifactNativeRenderer View Markdown / JSON / CSV / Code / SVG / PDF / raster images / GeoJSON (MapKit) / GLTF (SceneKit) / USDZ (RealityKit) / Turtle / TriG / N-Quads / RDF/XML / JSON-LD
ArtifactWebRenderer View HTML / React / Mermaid / LaTeX (KaTeX) / Vega-Lite via WKWebView

ArtifactNativeRenderer pulls in swift-markdown-ui for block-level Markdown rendering (headings, lists, tables, code blocks). The USDZ renderer uses RealityKit's RealityView with built-in pinch / drag / double-tap gestures.

Minimal example

import SwiftUI
import SwiftArtifact

struct ChatBubble: View {
    let message: String  // raw text containing <artifact> tags

    var body: some View {
        ArtifactCanvas(text: message)
            .artifactRenderer(MarkdownRenderer())
            .artifactRenderer(CodeRenderer())
            .artifactRenderer(JSONRenderer())
            .artifactRenderer(CSVRenderer())
            .artifactRenderer(SVGRenderer())
            .artifactRenderer(PDFRenderer())
            .artifactRenderer(PNGRenderer())
            .artifactRenderer(JPEGRenderer())
            .artifactRenderer(WebPRenderer())
            .artifactRenderer(GIFRenderer())
            .artifactRenderer(TIFFRenderer())
            .artifactRenderer(HEICRenderer())
            .artifactRenderer(BMPRenderer())
            .artifactRenderer(GeoJSONMapKitRenderer())
            .artifactRenderer(USDZModel3DRenderer())
            .artifactRenderer(TurtleRenderer())
            .artifactRenderer(TriGRenderer())
            .artifactRenderer(NQuadsRenderer())
            .artifactRenderer(RDFXMLRenderer())
            .artifactRenderer(JSONLDRenderer())
            .artifactRenderer(MermaidWebViewRenderer())
            .artifactRenderer(LaTeXWebViewRenderer())
            .artifactRenderer(HTMLWebViewRenderer())
            .artifactRenderer(ReactWebViewRenderer())
            .artifactRenderer(VegaLiteWebViewRenderer())
    }
}

Standalone display

ArtifactView(_:) shows an artifact without the card chrome — use it when an artifact already has its own header (sidebar, document outline, etc.).

ArtifactView(artifact)
    .artifactRenderer(MarkdownRenderer())

ArtifactCard(_:) wraps the same thing with a title bar, type badge, streaming indicator, and an optional disclosure button.

ArtifactCard(artifact) {
    Button { share(artifact) } label: { Image(systemName: "square.and.arrow.up") }
    Button { copy(artifact) }  label: { Image(systemName: "doc.on.doc") }
}

The card respects two environment modifiers:

.artifactCardContentInsets(EdgeInsets())    // override inner padding
.artifactCardDisclosure(.hidden)            // hide the expand/collapse button

If the environment override is not set, the card consults the resolved renderer's preferredContentInsets. Bundled renderers that own their own spacing or need edge-to-edge content publish a preference, so the renderer body is the source of truth for padding and sizing. For example, CSV owns table cell padding, Markdown and JSON own their reading/source margins, and raster images own their aspect-ratio layout.

Sizing policy

Renderers split into three layout families:

Family Renderers Sizing behavior
Text / table / source Markdown, JSON, CSV, Code Own their inner spacing and use bounded scrolling for large payloads. CSV measures the visible viewport, fills it for narrow tables, and expands horizontally only when the column count requires scrolling.
Intrinsic media SVG, PNG, JPEG, WebP, GIF, TIFF, HEIC, BMP Size from the rendered content. Raster images make the card body width the source of truth and compute height from the decoded image aspect ratio.
Fill-frame surfaces HTML, React, Mermaid, LaTeX, Vega-Lite, GeoJSON, USDZ, Turtle / TriG / N-Quads / RDF/XML / JSON-LD, PDF Render into a WebView / Map / RealityView / Canvas / document viewport and expand to the frame the caller provides. You are expected to wrap them with .frame(...) at the call site when the default viewport is not appropriate.

The optional artifactContentHeightLimit() modifier scrolls content above a configurable cap (artifactContentMaxHeight, default 360pt). It is used only where a height cap does not corrupt the renderer's own aspect-ratio or viewport contract.

ArtifactCard(artifact)
    .frame(height: 480)         // chat-bubble use: fixed height
    .artifactRenderer(GeoJSONMapKitRenderer())

Earlier versions imposed an internal 240–360pt cap on these renderers, which silently overrode caller-supplied .frame(height:). That cap was removed in 0.6.4 — the library no longer second-guesses your layout.

Partial rendering

Each renderer owns the rules for what counts as a renderable subset of its payload via refine(_:). The view layer never shows a half-parsed structure — it shows either a waiting state (.preRenderable) or whatever the renderer says is safe to draw (.renderable(String)).

Type Strategy while streaming
JSON / GeoJSON longest valid prefix down to the deepest open frame
SVG element-level boundary tracking, last unclosed element dropped
HTML token-level trim — half-typed tag dropped, <script> / <style> blocks withheld until their close tag arrives
Mermaid last incomplete line dropped + brace / quote balance check
LaTeX dangling \command and unbalanced braces trimmed back
CSV drops the last in-flight row
Markdown drops the last in-flight block
Turtle / TriG truncated to the last statement-terminating . outside strings, IRIs, and comments
N-Quads truncated to the last newline (each line is a self-contained quad)
RDF/XML framing pass collects every fully-closed top-level element under <rdf:RDF>
JSON-LD tolerant JSON AST emits triples for every fully-typed property pair, including pre-@context

Renderers without an incremental strategy fall back to the default refiner, which waits for artifact.isComplete.

A type-specific waiting UI is opt-in via preRenderableBody(artifact:progress:) — for example, the React renderer shows highlighted JSX source until the component finishes streaming. If a renderer doesn't override it, the view layer falls back to ArtifactProgressView.

Supported artifact types

The framework keys every artifact on its MIME type. The extension column lists the canonical file suffix(es) for that format — useful when ingesting files from disk or routing on an upload's filename. Extensions are advisory metadata; the renderer registry resolves on ArtifactType (i.e. the MIME) alone.

Tier 1 — Claude-compatible

Format MIME Extensions Renderer
HTML text/html .html, .htm HTMLWebViewRenderer
React application/vnd.ant.react .jsx, .tsx ReactWebViewRenderer
SVG image/svg+xml .svg SVGRenderer
Mermaid application/vnd.ant.mermaid .mmd, .mermaid MermaidWebViewRenderer
Markdown text/markdown .md, .markdown MarkdownRenderer
Code application/vnd.ant.code .diff, .patch for unified diffs; otherwise language carried in attributes CodeRenderer

CodeRenderer uses CodeEditSourceEditor with CodeEditLanguages-backed Tree-sitter highlighting on macOS, and falls back to a readonly monospaced SwiftUI surface on other platforms.

For unified diffs, CodeRenderer switches to CodeEditSourceEditor's code review surface instead of treating the payload as a plain syntax-highlighted file.

Signal Behavior
language, lang, or fileExtension is diff, patch, udiff, or unified-diff Render as a code review diff
title, filename, or fileName ends in .diff or .patch Render as a code review diff
Payload contains a unified diff hunk header such as @@ -1,3 +1,4 @@ Render as a code review diff
Payload has --- / +++ file headers Use the target path to infer the underlying source language
Artifact has contentLanguage, sourceLanguage, or targetLanguage Use that language for hunk syntax highlighting
AnyArtifact(
    id: ArtifactIdentifier("layout-diff"),
    type: .code,
    title: "Layout.diff",
    attributes: ["language": "diff"],
    payload: patch,
    isComplete: true
)

Tier 2 — Common agent output

Format MIME Extensions Renderer
JSON application/json .json JSONRenderer
CSV text/csv .csv CSVRenderer
Vega-Lite application/vnd.vegalite.v5+json .vl.json VegaLiteWebViewRenderer
GeoJSON application/geo+json .geojson GeoJSONMapKitRenderer
LaTeX application/x-latex .tex, .latex LaTeXWebViewRenderer
glTF (JSON) model/gltf+json .gltf GLTFSceneKitRenderer
USDZ model/vnd.usdz+zip .usdz USDZModel3DRenderer

CSVRenderer presents rows as a spreadsheet-style SwiftUI table. It supports sticky headers, type-aware alignment, zebra striping, row/table copy actions, and horizontal scrolling only when the table's minimum column widths exceed the visible viewport.

Documents and raster images

Binary-oriented renderers accept a payload string containing a remote URL, file:// URL, absolute file path, data URL, or base64 data.

Raster image renderers decode only complete payloads. They do not progressively render chunked image data; while artifact.isComplete == false, they stay in the normal pre-renderable waiting state. Once decoded, the image fills the available card body width and sets its height from the source image ratio, so the card content area does not leave unused side space.

Format MIME Extensions Renderer
PDF application/pdf .pdf PDFRenderer
PNG image/png .png PNGRenderer
JPEG image/jpeg .jpg, .jpeg JPEGRenderer
WebP image/webp .webp WebPRenderer
GIF image/gif .gif GIFRenderer
TIFF image/tiff .tif, .tiff TIFFRenderer
HEIC image/heic .heic HEICRenderer
BMP image/bmp .bmp BMPRenderer

Knowledge graph (W3C RDF)

All five RDF renderers share a layered, orthogonal knowledge-graph layout, blank-node-stable identifiers across re-renders, and progressive partial rendering. The layout constraints are specified in Specs/KnowledgeGraphLayout.md: edges use horizontal/vertical routes, avoid non-endpoint nodes, connect to node sides by their normal, and prefer shorter valid paths with fewer corners as the tie-breaker. JSON-LD and RDF/XML use bespoke partial processors so the diagram appears as triples become derivable from the prefix — not when the closing } or </rdf:RDF> arrives.

Format MIME Extensions Renderer
Turtle text/turtle .ttl TurtleRenderer
TriG application/trig .trig TriGRenderer
N-Quads application/n-quads .nq NQuadsRenderer
RDF/XML application/rdf+xml .rdf, .owl RDFXMLRenderer
JSON-LD application/ld+json .jsonld JSONLDRenderer

Tier 3 — User-defined

Register your own MIME type by conforming to Artifactable and adding an ArtifactRenderable implementation.

Writing a custom renderer

struct MyJSONRenderer: ArtifactRenderable, Sendable {
    static let artifactType: ArtifactType = .json

    // Optional: declare what counts as a renderable subset while streaming.
    // Omitting this gives you the default — wait for artifact.isComplete.
    static func refine(_ artifact: AnyArtifact) -> RefinedPayload {
        if artifact.isComplete {
            return .renderable(artifact.payload)
        }
        if let prefix = longestValidPrefix(of: artifact.payload) {
            return .renderable(prefix)
        }
        return .preRenderable(
            PreRenderableProgress(
                receivedCharacters: artifact.payload.count,
                hint: "waiting for first complete value"
            )
        )
    }

    // body receives the refined string, never the raw payload. It can assume
    // the input is well-formed for this renderer's type.
    func body(artifact: AnyArtifact, payload: String) -> some View {
        Text(payload).font(.system(.callout, design: .monospaced))
    }
}

To make the hosting card fill its chrome edge-to-edge (Map / WebView / Code style), override preferredContentInsets:

static let preferredContentInsets: EdgeInsets? = EdgeInsets()

For a type-specific waiting state, add preRenderableBody:

func preRenderableBody(
    artifact: AnyArtifact,
    progress: PreRenderableProgress
) -> some View {
    Text(artifact.payload)      // show the raw stream
        .font(.system(.callout, design: .monospaced))
        .foregroundStyle(.secondary)
}

License

MIT — see LICENSE.

Description

  • Swift Tools 6.3.0
View More Packages from this Author

Dependencies

Last updated: Sun May 31 2026 03:09:59 GMT-0900 (Hawaii-Aleutian Daylight Time)