Note

This entire codebase — Rust parsers, TypeScript renderers, tests, and tooling — is implemented by AI coding agents, primarily Claude and Codex, through iterative prompting. No human-written application code exists in this repository.

Why this project exists — a note from the author

OOXML's behavior is defined by a written specification (ECMA-376 / ISO-29500), and there is a clear answer to compare against: Word, Excel, and PowerPoint themselves. In principle, anyone with enough patience could have built a faithful viewer — the spec says what to implement, and the Office applications show whether you got it right.

In practice, it didn't happen. For more than a decade, no free, open-source library reached a rendering quality good enough for real use. There are a few commercial libraries with decent fidelity (and editing support), but their pricing makes them hard to adopt casually. I think the reason is simply cost: the specification is huge, and reading and implementing it faithfully takes far more effort than volunteers can afford.

Generative AI changed that. A viewer is an unusually good fit for AI-driven iterative development ("vibe coding"): there is a spec to read and a correct output to aim for, so the work comes down to interpreting the specification and refining the rendering until it matches. Limiting the scope to viewing also avoids the most serious risk an Office library can carry — corrupting a user's files.

So I'm building this library with AI coding agents, spec-first, and keeping it free to use. For some documents it already reproduces the desktop Office applications more faithfully than commercial libraries — and sometimes even the official Microsoft 365 web apps.

Office Open XML Viewer

Live demo

A browser-based viewer for Office Open XML documents that renders to an HTML Canvas element. The parsers are written in Rust and compiled to WebAssembly; the renderers use the Canvas 2D API. Each format also exposes a headless engine (DocxDocument / XlsxWorkbook / PptxPresentation) that renders into any caller-supplied canvas, so you can compose your own UI — scroll views, thumbnail grids, master-detail panes — instead of being locked into the built-in viewer. See the live framework examples for runnable React, Vue, Svelte, and Solid projects.

Project scope: read-only viewing

This project is intentionally a read-only viewer. Editing interfaces, mutation APIs, editable or lossless document models, saving / round-tripping, and partial editing of DOCX, XLSX, or PPTX files are out of scope. Read-only interactions such as selection, copy, search, shape IDs, hit testing, and annotation or external-tool integrations remain in scope, and editing-focused forks are welcome. See #496 for the rationale.

DOCX	XLSX	PPTX

npm install @silurus/ooxml
# or
pnpm add @silurus/ooxml

Bundler note: the Rust parsers ship as real .wasm asset files next to the JavaScript, referenced with the standard new URL('…', import.meta.url) form and fetched (streaming-compiled) at load time. Verified to work with zero config: webpack 5, Next.js (Turbopack, dev and build), Vite 8 (dev and build), Vite 7 production builds, and a plain <script type="module"> with no bundler at all. Two setups need a hand:

• Vite 7 dev server: the dependency optimizer rewrites the asset reference into its own cache path and the load fails (fixed in Vite 8). Add optimizeDeps: { exclude: ['@silurus/ooxml'] } to your vite.config — production builds are unaffected.
• esbuild / Angular CLI (whose application builder is esbuild-based): new URL asset references are not processed (esbuild#795). Copy the .wasm into your served output and point the viewer at it with the wasmUrl load option. For Angular CLI, copy the required *_parser_bg.wasm asset from node_modules/@silurus/ooxml/dist into the served output and pass its public URL to the viewer.

wasmUrl also serves the parser WASM from a CDN or any path you control:

new DocxViewer(canvas, { wasmUrl: 'https://cdn.example.com/docx_parser_bg.wasm' });

Bundle size note: the package is ESM-only (.mjs). npm's Unpacked Size includes every entry and optional asset, so it is not the size of a format-specific application graph. See the current production measurements on the stable Bundle size page. Import only the format and optional renderer entries your application needs. In main mode, MathJax, ChartEx, 3-D, Region Map and TIFF implementations remain outside the graph unless imported; the separately loaded worker asset is self-contained.

---

Quick Start

import { DocxViewer } from '@silurus/ooxml/docx';
import { XlsxSheetViewer, XlsxViewer, XlsxWorkbook } from '@silurus/ooxml/xlsx';
import { PptxViewer } from '@silurus/ooxml/pptx';

// DOCX — caller provides the <canvas>
const docxCanvas = document.getElementById('docx-canvas') as HTMLCanvasElement;
const docx = new DocxViewer(docxCanvas);
await docx.load('/document.docx');
docx.nextPage();

// XLSX — viewer manages its own <canvas> + tab bar
const container = document.getElementById('xlsx-container') as HTMLElement;
const xlsx = new XlsxViewer(container);
await xlsx.load('/workbook.xlsx');
xlsx.setSelection('B2:D5'); // A1 strings describe geometry; the normalized upper-left is ActiveCell

// Excel keeps Selection and ActiveCell separate. Use structured state when the
// active cell or Shift-extension anchor is not the area's upper-left cell.
xlsx.setSelection({
  areas: [{ kind: 'cells', top: 2, left: 2, bottom: 5, right: 4 }],
  activeAreaIndex: 0,
  activeCell: { row: 3, col: 3 },
  extensionAnchor: { row: 2, col: 2 },
});

// Read-only, serializable context for an AI/MCP request. Populated cells are
// bounded and detached; formulas and Viewer-formatted display text are retained.
const context = xlsx.getSelectionContext({
  maxCells: 1_000,
  maxTextCharacters: 1_048_576,
});

// XLSX active-sheet surface only — caller provides the <canvas>
const sheetCanvas = document.getElementById('xlsx-canvas') as HTMLCanvasElement;
const sheet = new XlsxSheetViewer(sheetCanvas);
await sheet.load('/workbook.xlsx');
await sheet.goToSheet(1);

// Share one parsed workbook across independently scrollable sheet canvases,
// including canvases created in same-origin popup windows.
const workbook = await XlsxWorkbook.load('/workbook.xlsx');
const firstCanvas = document.getElementById('xlsx-first-canvas') as HTMLCanvasElement;
const secondCanvas = document.getElementById('xlsx-second-canvas') as HTMLCanvasElement;
const firstSheet = XlsxSheetViewer.fromWorkbook(firstCanvas, workbook);
const secondSheet = XlsxSheetViewer.fromWorkbook(secondCanvas, workbook);
await Promise.all([
  firstSheet.goToSheet(0),
  secondSheet.goToSheet(1),
]);

// Each viewer borrows the workbook; the caller closes it after the viewers.
firstSheet.destroy();
secondSheet.destroy();
workbook.destroy();

// PPTX — caller provides the <canvas>
const pptxCanvas = document.getElementById('pptx-canvas') as HTMLCanvasElement;
const pptx = new PptxViewer(pptxCanvas);
await pptx.load('/deck.pptx');
pptx.nextSlide();

Rendering equations

OMML equations (m:oMath / m:oMathPara) in .docx, .pptx and .xlsx are rendered with MathJax + STIX Two Math. That engine is ~3 MB, so it is opt-in: import the math engine from the separate @silurus/ooxml/math entry and pass it to the viewer. Pass it and equations render; omit it and the engine asset is not fetched or evaluated (equations are simply skipped; the on-demand render-worker asset retains a small loader). When you do pass it, the ~3 MB engine ships as a standalone asset file next to the bundle rather than an inline data URL, and is fetched on demand — only the first time a document actually contains an equation, so equation-free documents never pay for it. It is fully self-contained: served from your own origin, no cross-origin requests.

import { DocxViewer } from '@silurus/ooxml/docx';
import { math } from '@silurus/ooxml/math';

const canvas = document.getElementById('docx-canvas') as HTMLCanvasElement;
const docx = new DocxViewer(canvas, { math }); // ← equations now render
await docx.load('/paper-with-equations.docx');

The same math engine works for every viewer (DocxViewer, PptxViewer, XlsxViewer) and every headless engine (DocxDocument, PptxPresentation, XlsxWorkbook). You inject it once where you create the object — the viewer constructor or the .load() options — and every render reuses it; it is never a per-render argument. (Excel stores "Insert > Equation" as OMML inside the shared DrawingML <xdr:txBody> grammar, so XlsxViewer renders equations embedded in shapes / text boxes the same way.)

Optional rendering modules

Classic DrawingML 2-D chart families are included in every format entry. Microsoft ChartEx, model-space 3-D charts and offline country-level Region Maps are separate entries. TIFF decoding is also a separate entry shared by DOCX, XLSX and PPTX. Inject the modules once in the same load options object as math; omitting one keeps its main-mode implementation out of the ordinary format graph. The built-in modules work in both main and worker modes. Without chartEx, ChartEx families show the standard unsupported-chart placeholder. Without threeD, 3-D chart groups fall back to their canonical 2-D family. Without regionMap, Region Maps show the standard unsupported-chart placeholder. Without tiff, recognized TIFF parts report a render error instead of relying on the browser decoder or being silently omitted. The code-size boundary applies to the default main-mode application graph. The separately loaded render-worker asset stays self-contained for broad bundler compatibility and therefore contains its built-in optional renderer implementations.

import { XlsxViewer } from '@silurus/ooxml/xlsx';
import { threeD } from '@silurus/ooxml/three-d';
import { regionMap } from '@silurus/ooxml/region-map';
import { chartEx } from '@silurus/ooxml/chart-ex';
import { tiff } from '@silurus/ooxml/tiff';

const container = document.getElementById('xlsx-container') as HTMLElement;
const workbookViewer = new XlsxViewer(container, {
  threeD,
  regionMap,
  chartEx,
  tiff,
  mode: 'worker',
});
await workbookViewer.load('/workbook-with-advanced-charts.xlsx');

The Region Map renderer is deterministic and network-free. It uses a pinned Natural Earth country dataset, supports authored world projections and two/three-stop value ramps, and fails closed for cached identities or sub-country/view-specific layouts that the bounded offline model cannot yet represent safely. The specification/Office evidence boundary for automatic chart behavior is documented in Chart compatibility evidence and scope.

The bounded TIFF 6.0 codec decodes the first IFD/page of stripped, top-left images in the uncompressed bilevel, 8-bit grayscale, RGB, RGBA and process-CMYK classes, plus 1-bit CCITT Group 4. Unsupported or malformed first-page classes fail with a diagnostic render error. It is not a general-purpose TIFF library, but as a small by-product the same tiff.render() method can provide a simple preview of a supported standalone TIFF file. Its optional TiffRenderOptions argument can request a bounded display-sized bitmap; the Office viewers supply those targets automatically for layout-backed image paints. Try Yours and the VS Code extension enable every first-party optional module; library applications can choose only the entries they need. See Production decisions for the full module list and Bundle size for current measurements.

Off-main-thread rendering

By default the headless engines parse in a worker but render on the main thread. Pass mode: 'worker' to .load() to normally parse and render inside a Web Worker — the main thread presents the returned ImageBitmap via a bitmaprenderer context, keeping it free for scrolling and input. It requires Worker + OffscreenCanvas.

import { PptxPresentation } from '@silurus/ooxml/pptx';

// Render entirely inside a Web Worker — the main thread only paints bitmaps.
const pres = await PptxPresentation.load('/deck.pptx', { mode: 'worker' });
const canvas = document.getElementById('pptx-canvas') as HTMLCanvasElement;
const bitmap = await pres.renderSlideToBitmap(0, { width: 960, dpr: window.devicePixelRatio });
const ctx = canvas.getContext('bitmaprenderer') as ImageBitmapRenderingContext;
ctx.transferFromImageBitmap(bitmap); // consumes the bitmap

The *ToBitmap method exists on all three engines — PptxPresentation.renderSlideToBitmap(slideIndex, opts), DocxDocument.renderPageToBitmap(pageIndex, opts), and XlsxWorkbook.renderViewportToBitmap(sheetIndex, viewport, opts) (the xlsx variant requires opts.width and opts.height, since a worker has no DOM element to measure). They work in both modes — in main mode they render to an internal OffscreenCanvas — so you can write mode-agnostic code.

Notes:

• The returned ImageBitmap is owned by the caller: transferFromImageBitmap consumes it, or call bitmap.close() when done.
• The canvas-target methods (renderSlide(canvas), renderPage(canvas), renderViewport(canvas)) are unavailable in worker mode — use the *ToBitmap variants instead.
• The built-in math, ChartEx, 3-D chart, and Region Map renderers work in both modes through the same math, chartEx, threeD, and regionMap options. Custom renderer objects are main-realm code and therefore use the feature's documented fallback in mode: 'worker'.
• A DOCX document that requires browser-only OpenType vertical-glyph selection automatically uses effective main mode for correct shaping. Read document.mode after loading when your integration needs to observe this fallback.
• Trade-off: worker mode keeps the main thread responsive, but each frame is transferred back as an ImageBitmap, so a single render can be marginally slower than mode: 'main'. Choose it for non-blocking UI, not raw speed.

Continuous scroll viewers

DocxScrollViewer and PptxScrollViewer render the whole document as one vertically-scrolling, PDF-reader-style surface instead of a single page/slide at a time. Unlike DocxViewer / PptxViewer (which take a <canvas>), the scroll viewers take a container <div> — they own the scroll host, virtualize the page/slide list (only the visible window plus a small overscan is mounted), and recycle canvases as you scroll.

import { DocxScrollViewer } from '@silurus/ooxml/docx';

const container = document.getElementById('docx-scroll') as HTMLElement;
const viewer = new DocxScrollViewer(container);
await viewer.load('/document.docx');
// viewer.scrollToPage(3);
// viewer.pageCount, viewer.topVisiblePage

import { PptxScrollViewer } from '@silurus/ooxml/pptx';

const container = document.getElementById('pptx-scroll') as HTMLElement;
const viewer = new PptxScrollViewer(container);
await viewer.load('/deck.pptx');
// viewer.scrollToSlide(2);
// viewer.slideCount, viewer.topVisibleSlide

The container must have a bounded height (e.g. height: 100vh or a flex child) so the viewer can size its scroll host to it. Base zoom fits the widest available DOCX page, or the PPTX slide width, to the container and re-fits on resize. A progressively loaded DOCX re-fits if a wider page appears; a 0-width container defers layout until it has width. Call destroy() to tear down (a self-loaded engine is destroyed with it; a borrowed one is not — see below).

Pass refitOnResize: false when the viewport must not determine the document's physical display size. An explicit pre-load setScale(1) then keeps the same authored font size at roughly the same on-screen size for portrait pages, landscape pages, and slides; users can still zoom or call fitWidth() / fitPage() explicitly.

Desk appearance. The viewer paints each page/slide on its own white canvas with a soft drop shadow, over a transparent "desk". Style the desk and the sheet gaps without any wrapper CSS:

const viewer = new DocxScrollViewer(container, {
  background: '#f3f4f6',            // the desk behind / between pages
  gap: 24,                          // vertical gap between pages
  paddingTop: 32,                   // desk padding above the first page
  pageShadow: '0 0 0 1px #c8ccd0',  // crisp 1px "border" look (box-shadow never shifts layout)
  // pageShadow: false,             // flat pages, no shadow
});

paddingBottom, paddingLeft and paddingRight each default to gap, so the sheet sits inside a uniform desk margin; pass 0 for a flush edge.

Zoom. Ctrl/⌘ + mouse-wheel (and trackpad pinch) zooms the surface; bare-wheel still scrolls natively. Zoom is flicker-free — a rapid gesture shows a CSS preview and settles into a crisp re-render when it pauses. Bounds are the absolute scale factors zoomMin / zoomMax (default 0.1 / 4), and setScale(scale) sets it programmatically. When fitting needs a scale below zoomMin, that fitted scale becomes the effective minimum so users can zoom in and still return to the original fit. Pass enableZoom: false to disable.

Text selection and find. Pass enableTextSelection: true to overlay a transparent, selectable text layer per page/slide for native copy. It works in both mode: 'main' and mode: 'worker'; worker rendering returns the retained text-run geometry beside each bitmap. findText(), findNext(), findPrev(), and clearFind() search the complete document, including virtualized pages or slides outside the mounted window. Set findHighlightColors: { match, active } on any viewer to override the two overlay backgrounds with CSS colors; use an alpha color when the canvas text should remain visible through the highlight.

Selection context for AI/MCP. Every Viewer exposes one read-only query, getSelectionContext(), for handing the user's current focus to an external assistant. The result is a detached, JSON-serializable snapshot discriminated by format and kind; it never exposes a mutable document model or sends data over the network. Text, run locators, and populated XLSX cells have hard resource limits. Check truncated / truncationReasons before building a prompt.

const docx = new DocxViewer(docxCanvas, {
  enableTextSelection: true,
  enableElementSelection: true,
  onSelectionContextChange(context) {
    // kind === 'text' or 'element': selected text or a clicked drawing
    updateAskAiButton(context);
  },
});

const pptx = new PptxViewer(pptxCanvas, {
  enableTextSelection: true,
  enableElementSelection: true,
  onSelectionContextChange(context) {
    // Text selection wins while it exists; otherwise a slide-element click
    // yields kind === 'element' with compact bounds, provenance and content.
    updateAskAiButton(context);
  },
});

const spreadsheet = new XlsxViewer(container, {
  enableElementSelection: true,
  onSelectionContextChange(context) {
    updateAskAiButton(context);
  },
  onContextMenu: async ({ originalEvent, getContext }) => {
    // Browser-menu control must happen synchronously, before the first await.
    originalEvent.preventDefault();
    const { clientX, clientY } = originalEvent;
    const context = await getContext();
    openContextMenu({ clientX, clientY, context });
  },
});
const spreadsheetContext = spreadsheet.getSelectionContext({ maxCells: 1_000 });
// kind === 'range': selection state, values and formulas.
// kind === 'element': a clicked chart, picture or shape.

All three formats expose the same onSelectionContextChange(context) handoff; callers may instead query on demand. enableElementSelection is an independent, explicit opt-in because it enables object selection and draws a non-editable outline around the focused object; adding the callback alone never enables object hit-testing. XLSX separately retains onSelectionStateChange for canonical UI state such as ActiveCell and multiple areas. Its context callback is frame-coalesced so drag selection does not build one snapshot per pointer event. onContextMenu is also common to every Viewer. It receives the real browser event synchronously so the host can call preventDefault(), plus a clearly asynchronous getContext() lookup for the right-click target. The lookup starts on the first call and is memoized. Omitting the callback installs no listener and leaves the native browser menu unchanged. DocxDocument.getElementContextAt() and PptxPresentation.getElementContextAt() provide the identical compact element query for custom/headless page or slide surfaces in both modes. PPTX element provenance is limited to master | layout | slide; editor tree indexes, archive paths, save/round-trip handles, and mutation APIs are deliberately absent. When switching on kind, retain a default branch so a future read-only focus kind can be added without changing the transport envelope. See the selection-context guide for the complete contract, resource bounds, PPTX hit-testing semantics, and extension policy.

Hyperlinks. For DOCX/PPTX the link hit regions live on the text-selection overlay, so hyperlink interaction requires enableTextSelection: true; when that overlay is enabled, links are interactive by default. XLSX hit-tests cells directly, so links are interactive out of the box. An external link opens in a new tab (scheme-sanitized to http / https / mailto / tel, noopener), and an internal target navigates within the document (docx bookmark, pptx slide jump, xlsx defined name or cell reference). XLSX references may switch sheets and then scroll the destination cell into view; a range navigates to its first cell. Pass onHyperlinkClick(target) to take over the click yourself. Pass enableHyperlinks: false to disable hyperlink interactivity entirely — no hit-testing, no pointer cursor over links, no default navigation, and onHyperlinkClick is never called; links still render as authored but are inert. This applies to every viewer that supports hyperlinks (DocxViewer, DocxScrollViewer, PptxViewer, PptxScrollViewer, XlsxViewer, XlsxSheetViewer).

Choose one loading mode. For the normal one-view case, construct a Viewer and call viewer.load(source): the Viewer owns the parsed engine, may replace it on a later load, and destroys it during teardown. For master–detail, multi-pane, or multi-window UIs, create the Viewer from an already-loaded headless engine so every view shares one parse. The two modes are mutually exclusive: on a Viewer returned by a from*() factory, load() is unsupported, the engine's own mode wins, and destroy() leaves the borrowed engine intact — the caller owns its lifecycle.

Use case	Acquisition	Engine owner
One Viewer	new Viewer(target) then viewer.load(source)	Viewer
Shared parse	Engine.load(source) then Viewer.fromDocument / fromPresentation / fromWorkbook(...)	Caller

import { DocxDocument, DocxScrollViewer, DocxViewer } from '@silurus/ooxml/docx';

const document = await DocxDocument.load('/document.docx'); // parse once
const scroll = DocxScrollViewer.fromDocument(container, document);
const page = DocxViewer.fromDocument(canvas, document);
await page.goToPage(0);
// ...also drive a thumbnail grid or more panes from the same document.
page.destroy();
scroll.destroy();  // the borrowed document is NOT destroyed — you own it
document.destroy(); // release it yourself when every pane is gone

PptxViewer and PptxScrollViewer use fromPresentation(...); XlsxViewer and XlsxSheetViewer use fromWorkbook(...).

Progressive DOCX layout

For a large DOCX file, set progressiveLayout: true to resolve load() when the opening pages are paintable while the same paginator continues in the background. It is available on DocxViewer, DocxScrollViewer, and DocxDocument.load() in both render modes. Pair it with mode: 'worker' when the remaining layout and paint work should also stay off the UI thread.

import { DocxScrollViewer } from '@silurus/ooxml/docx';

const container = document.querySelector('#document') as HTMLElement;
const pager = document.querySelector('#pager') as HTMLElement;

const viewer = new DocxScrollViewer(container, {
  progressiveLayout: true,
  mode: 'worker',
  onVisiblePageChange(pageIndex, availablePages, layoutComplete) {
    pager.textContent =
      `Page ${pageIndex + 1} of ${availablePages}${layoutComplete ? '' : '…'}`;
    pager.setAttribute('aria-busy', String(!layoutComplete));
  },
});

await viewer.load('/document.docx');

// Print, export, and final-page-count UI need the converged layout.
await viewer.waitUntilLayoutComplete();
console.log('Final page count:', viewer.pageCount);

While layoutComplete is false, pageCount and the callback's total argument mean pages available so far, not the final total. The page callbacks fire again when that count grows, even if the visible page does not change. Await waitUntilLayoutComplete() before printing, exporting, or snapshotting a final count. In-document NUMPAGES fields are repainted with their authoritative value after pagination converges. See the progressive layout guide and DOCX API reference.

Progressive PPTX layout

PPTX exposes the same progressive lifecycle on PptxViewer, PptxScrollViewer, and PptxPresentation.load():

import { PptxScrollViewer } from '@silurus/ooxml/pptx';

const viewer = new PptxScrollViewer(container, {
  progressiveLayout: true,
  mode: 'worker',
  onVisibleSlideChange(slideIndex, slideCount, layoutComplete) {
    pager.textContent = `Slide ${slideIndex + 1} of ${slideCount}`;
    pager.setAttribute('aria-busy', String(!layoutComplete));
  },
});

await viewer.load('/presentation.pptx');
await viewer.waitUntilLayoutComplete();

Unlike DOCX pagination, a PPTX bootstrap already provides the final slide list and uniform dimensions. slideCount and the ScrollViewer's scroll extent are therefore stable from first paint; availableSlideCount grows as the paintable opening prefix is prepared. Scrolling ahead shows a loading state without changing the scrollbar length. See the PPTX progressive layout guide and PPTX API reference.

For presentations, enableMediaPlayback: true makes embedded audio and video interactive inside the real viewport plus mediaOverscan slides. Other mounted slides remain static and selectable, avoiding offscreen media blobs and animation loops.

Both viewers also expose relayout() (force a re-fit when the container resizes in a way a ResizeObserver cannot see — e.g. a late web-font load), onVisiblePageChange (fires when the top-most visible page, provisional DOCX page count, or completion state changes), onVisibleSlideChange (fires when the top-most visible slide or PPTX completion state changes; its slide count is final from first paint), and onError (async per-page render failures are routed here instead of crashing the scroll loop). The parse/render knobs from the headless engines (mode, useGoogleFonts, resourceLimits, the deprecated maxZipEntryBytes alias, math, dpr) are accepted too.

Markdown export

Every headless engine can project its document to GitHub-flavoured markdown for LLM ingestion, full-text search, or diffing — headings, lists, tables, and (for docx) footnotes / comments are preserved; layout, fonts, and positioning are dropped. The projection is compiled into the parser WASM you already ship, so it adds zero bundle weight. toMarkdown() works in both mode: 'main' and mode: 'worker' (it runs off the archive opened at load()):

import { DocxDocument } from '@silurus/ooxml/docx';

const doc = await DocxDocument.load('/document.docx');
const md = await doc.toMarkdown();

PptxPresentation.toMarkdown() (title slides → # headings, body → nested bullets, notes / comments collated) and XlsxWorkbook.toMarkdown() (each sheet → a ## SheetName pipe table) are the twins.

The repository also contains a low-level adapter and CLI for workspace tooling. They are internal implementation utilities, not separately published packages; installed applications should use the format model's toMarkdown() method.

---

Architecture diagram

flowchart TB
    subgraph build["🦀  Build-time  (Rust → WebAssembly)"]
        direction LR
        docx_rs["packages/docx/parser/src/lib.rs"]
        xlsx_rs["packages/xlsx/parser/src/lib.rs"]
        pptx_rs["packages/pptx/parser/src/lib.rs"]
        docx_rs -- wasm-pack --> docx_wasm["docx_parser.wasm"]
        xlsx_rs -- wasm-pack --> xlsx_wasm["xlsx_parser.wasm"]
        pptx_rs -- wasm-pack --> pptx_wasm["pptx_parser.wasm"]
    end

    subgraph browser["🌐  Runtime  (Browser)"]
        VIEWER["DOCX / XLSX / PPTX Viewer"] --> ENGINE["format headless engine"]
        ENGINE -->|mode: main| PARSE["parser worker\n〈WASM parse〉"]
        PARSE --> MODEL["validated document model"]
        MODEL --> MAIN["layout + Canvas paint\n〈main thread〉"]
        ENGINE -->|mode: worker| RENDER_WORKER["render-worker.ts\n〈WASM parse + layout + paint〉"]
        RENDER_WORKER --> BITMAP["ImageBitmap"]
        MAIN --> CANVAS["<canvas>"]
        BITMAP --> CANVAS
        CORE["@silurus/ooxml-core\nshared layout and paint primitives"]
        MAIN -. uses .-> CORE
        RENDER_WORKER -. uses .-> CORE
    end

    docx_wasm --> PARSE
    xlsx_wasm --> PARSE
    pptx_wasm --> PARSE
    docx_wasm --> RENDER_WORKER
    xlsx_wasm --> RENDER_WORKER
    pptx_wasm --> RENDER_WORKER

Loading

All three formats use the same public mode boundary. In the default main mode, WASM parsing runs in a Worker and the validated model is laid out and painted on the main thread. In worker mode, parsing, layout and Canvas paint run in a render Worker, which returns an ImageBitmap for presentation. Built-in optional renderers use the same injection options in both modes. @silurus/ooxml-core owns the layout, paint and type primitives shared by DOCX, XLSX and PPTX.

Key files

File	Role
packages/docx/parser/src/lib.rs	Rust WASM parser — DOCX ZIP → Document JSON
packages/xlsx/parser/src/lib.rs	Rust WASM parser — XLSX ZIP → Workbook JSON
packages/pptx/parser/src/lib.rs	Rust WASM parser — PPTX ZIP → Presentation JSON
packages/docx/src/renderer.ts	DOCX Canvas 2D rendering and text layout
packages/xlsx/src/renderer.ts	XLSX Canvas 2D viewport rendering
packages/pptx/src/renderer.ts	PPTX Canvas 2D rendering
packages/*/src/worker.ts	Main-mode parser Worker entry
packages/*/src/render-worker.ts	Worker-mode parser, layout and Canvas paint entry
packages/*/src/viewer.ts	Public Viewer API — canvas lifecycle, navigation
packages/core/src/index.ts	Cross-format primitives — chart renderer, shape helpers, autoResize, shared types

---

Framework Examples

Runnable TypeScript projects are available for React, Vue, Svelte, and Solid. Each guide embeds the complete StackBlitz project and supports selecting a local DOCX, XLSX, or PPTX file without uploading it.

---

Feature Support

Word (.docx)

Category	Feature	Status
Document	Page rendering	✅
	Page size and margins	✅
	Headers / footers (default / first / even)	✅
	Section breaks (continuous / nextPage / oddPage / evenPage)	✅
	Page borders (w:pgBorders, §17.6.10 — standard line styles, offsetFrom / display / zOrder; art borders not yet supported)	✅
	Line numbering (w:lnNumType, §17.6.8)	✅
	Section vertical alignment (w:vAlign, §17.6.22)	✅
Text	Paragraphs	✅
	Bold, italic, underline, strikethrough	✅
	Font family, size, color	✅
	Hyperlinks	✅
	Superscript / subscript (w:vertAlign)	✅
	Ruby annotations / furigana (w:ruby)	✅
Formatting	Paragraph alignment (left / center / right / justify / distribute — CJK both/distribute spread by inter-character pitch, §17.18.44)	✅
	Line spacing (auto / atLeast / exact)	✅
	Document grid (w:docGrid, §17.6.5 — line pitch + East Asian character grid / 字詰め)	✅
	Margin collapsing between paragraphs	✅
	Indents and tab stops	✅
	Multi-column section layout (w:cols, §17.6.4 — newspaper-flow columns; full-width floats span all columns)	✅
	Lists (bullet and numbered, multi-level %N markers §17.9.11)	✅
	Paragraph styles (Heading 1–9, Normal, custom)	✅
	Table style w:pPr cascade (§17.7.6)	✅
	Table style borders / shading / banding (tblStylePr, cnfStyle, §17.4.7)	✅
	Table of contents (TOC field) — dot leaders, right-aligned page numbers	✅
	keepNext / keepLines / widowControl	✅
	Right-to-left text — UAX#9 bidi, w:bidi / w:rtl, complex-script formatting (w:szCs / w:bCs / rFonts@cs, §17.3.2.26), RTL lists and indents	✅
	Japanese kinsoku line breaking (w:kinsoku, §17.15.1.58 — 行頭/行末禁則)	✅
	Vertical writing (縦書き — UAX#50 vertical glyph forms, 縦中横 tate-chu-yoko runs, 、。 upper-right positioning; §17.3.2 vertical text)	✅
Elements	Tables (with borders, fills, merges, banding, alignment)	✅
	Table auto-layout by preferred widths (w:tblLayout autofit, §17.4.52; min content width)	✅
	Table indent (w:tblInd, §17.4.50)	✅
	Right-to-left table column order (w:bidiVisual, §17.4.1)	✅
	Charts (embedded DrawingML c:chart — bar / line / area / pie / doughnut / radar / scatter, via the shared core chart renderer; data labels honour dLblPos, §21.2.2.48)	✅
	ChartEx (waterfall / histogram / Pareto / funnel / box & whisker / treemap / sunburst)	✅ opt-in
	Math equations (OMML m:oMath / m:oMathPara, rendered via MathJax — opt-in @silurus/ooxml/math)	✅
	Images (inline and anchored, with text wrap and adaptive display-sized decoding for oversized rasters)	✅
	TIFF images (opt-in @silurus/ooxml/tiff; bounded bilevel, grayscale, RGB(A), process-CMYK and CCITT Group 4)	✅
	SVG images (asvg:svgBlip MS-2016 extension — vector drawn from the embedded .svg, raster fallback)	✅
	Text boxes / drawing shapes (inline and anchored wps:wsp / wps:txbx, including solid, gradient, and image fills; a:prstGeom — 186 preset geometries via the shared engine; connector arrow heads headEnd / tailEnd (§20.1.8.3) and prstDash dash patterns (§20.1.8.48)). Text-box paragraphs run through the same line-layout engine as body text, so kinsoku 行頭/行末禁則 (§17.15.1.58–60), UAX#9 bidi (w:bidi, §17.3.1.6), justification (§17.18.44) and tab stops (§17.3.1.37) all apply inside a box	✅
	WMF and EMF metafile images (legacy vector, incl. inside text boxes) — rasterized via a built-in player: window→viewport mapping (MS-EMF map modes, world transform), pens/brushes, poly/rect/ellipse, text-out, path clipping, and embedded DIB blits	✅
	Legacy VML content — positioned shapes, text boxes, image previews, and authored text wrapping	✅
	OLE embedded objects (w:object — the baked VML v:imagedata preview is drawn; the embedded app is not run)	✅
Advanced	Footnotes — reference markers + bottom-of-page bodies with separator rule, numbered (w:footnoteReference / w:footnoteRef, §17.11)	✅
	Endnotes — reference markers + bodies at document end (w:endnoteReference, §17.11)	✅
	Page-number formats (w:pgNumType restart / format §17.6.12; PAGE \* switches — decimal / roman / letter / hex / ordinal-dash / hebrew2 / koreanLegal, §17.18.59)	✅
	Field date/time pictures (TIME / DATE field \@ format, §17.16.5.72 / .16)	✅
	w:snapToGrid opt-out of the document grid (§17.3.1.32)	✅
	Track changes (§17.13.5 w:ins / w:del / w:moveFrom / w:moveTo) — the default render is the FINAL state (deletions and moved-away text hidden); the opt-in markup view (showTrackedChanges) draws author-coloured underline / strikethrough plus margin change bars, and body-story revision records are available as data	✅
	Comments (§17.13.4) — opt-in margin balloons (comments: true): commented ranges tinted, threaded replies via commentsExtended.xml, resolved threads hidden, click-to-select stacking; also available as data (doc.comments, doc.commentAnchorRanges())	✅
	Markdown export (DocxDocument.toMarkdown() — headings, lists, tables, footnotes / comments)	✅
	Mail merge fields	❌ Not planned
Interaction	Text selection, including table-cell text (transparent overlay, native copy)	✅
	Bounded read-only text/element context (getSelectionContext(), page/source locators, element selection, AI/MCP callback)	✅
	In-document find (findText / findNext / findPrev / clearFind — full-text search, all hits highlighted, each match tagged with its page)	✅
	Runtime zoom (getScale / setScale / fitWidth / fitPage)	✅
	Clickable hyperlinks (overlay hit-test, onHyperlinkClick; internal bookmark / anchor navigation)	✅
	Continuous scroll viewer (DocxScrollViewer — virtualized page list, desk background / shadow, Ctrl/⌘+wheel zoom, borrowed-engine factory)	✅
Loading	Password-protected files ([MS-OFFCRYPTO] Agile Encryption — load(bytes, { password }), decrypted client-side via WebCrypto; legacy Standard / Extensible encryption → typed unsupported-encryption)	✅

---

Excel (.xlsx)

Category	Feature	Status
Workbook	Multiple sheets, sheet names	✅
	Sheet tab colors (<sheetPr><tabColor> — theme / tint / indexed / rgb)	✅
Cells	Text, number, boolean, error values	✅
	Formula results (from cached <v>)	✅
	Dates (ECMA-376 date format codes)	✅
	Rich text (per-run formatting)	✅
	East-Asian furigana (<rPh> §18.4.6 + <phoneticPr> §18.4.3 — drawn when a cell opts in via ph="1"; row-level <row ph> inheritance)	✅
Formatting	Bold, italic, underline (single / double / singleAccounting / doubleAccounting), strikethrough	✅
	Superscript / subscript (vertAlign)	✅
	Font family, size, color	✅
	Cell background color (solid + gradient)	✅
	Pattern fills (gray125 / gray0625 / lightGray / mediumGray / darkGray and the 12 light* / dark* directional hatches)	✅
	Borders (thin, medium, thick, hair, double, dashed, dotted, dashDotDot, …)	✅
	Diagonal borders (diagonalUp / diagonalDown, single + double)	✅
	Horizontal / vertical alignment	✅
	Text wrapping	✅
	Japanese kinsoku line breaking in wrapped cells (行頭/行末禁則, shared core engine)	✅
	Number formats (0.00, %, #,##0, custom date/time)	✅
Structure	Merged cells	✅
	Right-to-left sheets (sheetView rightToLeft, §18.3.1.87 — mirrored grid, headers, selection, scroll)	✅
	Frozen panes	✅
	Row / column sizing (custom widths and heights)	✅
	Hidden rows / columns	✅
	Row / column outline grouping (outlineLevel / collapsed §18.3.1.73 / .13, <outlinePr> — gutter brackets, +/− collapse, numbered level buttons; view-only)	✅
Elements	Images (<xdr:twoCellAnchor>, with adaptive display-sized decoding for oversized rasters)	✅
	TIFF images (opt-in @silurus/ooxml/tiff; bounded bilevel, grayscale, RGB(A), process-CMYK and CCITT Group 4)	✅
	OLE embedded objects (<oleObjects> — the legacy VML v:imagedata preview keyed by oleObject@shapeId is drawn; an image-typed objectPr target is preferred when present, and the embedded app is not run)	✅
	SVG images (asvg:svgBlip MS-2016 extension — vector drawn from the embedded .svg, raster fallback)	✅
	Drawing shapes / text boxes (xdr:sp, xdr:txBody — 186 preset geometries via the shared engine, with avLst adjust handles)	✅
	Math equations in shapes (OMML m:oMath / m:oMathPara in xdr:txBody, incl. a14:m / mc:AlternateContent; rendered via MathJax — opt-in @silurus/ooxml/math)	✅
	Charts (bar, line, area, pie, doughnut, radar, scatter / bubble)	✅
	ChartEx (waterfall / histogram / Pareto / funnel / box & whisker / treemap / sunburst)	✅ opt-in
	Chart markers (circle / square / diamond / triangle / x / plus / star / dot / dash, per-point <c:dPt> overrides; markers-only scatter series draw a marker legend key)	✅
	Chart data labels (<c:dLbl> per-point with CELLRANGE / VALUE / SERIESNAME / CATEGORYNAME field references, position l/r/t/b/ctr/outEnd)	✅
	Chart error bars (<c:errBars> X/Y direction, cust / fixedVal / stdErr / stdDev / percentage, dashed/styled lines)	✅
	Chart manual layout (<c:title><c:layout> and <c:plotArea><c:layout>)	✅
	Sparklines (x14:sparklineGroup — line / column / win-loss, with markers and high/low/first/last/negative highlights)	✅
Advanced	Conditional formatting (cellIs, colorScale, dataBar, iconSet, top10, aboveAverage)	✅
	Slicers (static, Office 2010 extension)	✅
	Pivot tables (saved worksheet output renders unchanged; read-only metadata is exposed. Refresh, recalculation, filtering, restructuring, and interactivity are unsupported)	⚠️ Partial
	Cell comments / notes (classic xl/commentsN.xml + Office-365 threaded comments — red triangle indicator + author / text via the worksheet model; pointer or keyboard users can open the popup, with a polite screen-reader status)	✅
	Data validation (rules via the worksheet model; list-type dropdown arrow on the selected cell whose click opens a panel showing the allowed values — read-only)	✅
	Markdown export (XlsxWorkbook.toMarkdown() — each sheet as a ## SheetName pipe table)	✅
Interaction	Cell selection (single / range / row / column / all / multiple areas; setSelection('B2:D5') or canonical structured state)	✅
	Excel-style row / column header highlight on selection	✅
	Shift+click to extend, Ctrl/⌘+drag to add another area, Ctrl+C to copy as TSV	✅
	Text selection inside cells (transparent overlay)	✅
	onSelectionStateChange, bounded range/element getSelectionContext() / copySelection(), chart/picture/shape selection, getCellAt(x, y)	✅
	Zoom slider (Excel-style, right of the tab bar, 10–400% with 100% centered; showZoomSlider option)	✅
	Ctrl/⌘ + mouse-wheel and trackpad-pinch zoom (in addition to the slider)	✅
	Runtime fit / zoom API (fitWidth / fitPage / getScale / setScale, in addition to the slider)	✅
	In-document find (findText / findNext / findPrev / clearFind — matches tagged with sheet + cell)	✅
	Clickable hyperlinks (onHyperlinkClick; internal defined-name / sheet-and-cell navigation, ranges use the first cell)	✅
	Drag-to-resize columns / rows by dragging header borders (resizable option, default on) — view-only: changes the on-screen view only and never modifies the loaded file	✅
	Customizable cell-selection color (selectionColor option, setSelectionColor())	✅
Loading	Password-protected files ([MS-OFFCRYPTO] Agile Encryption — load(bytes, { password }), decrypted client-side via WebCrypto; legacy Standard / Extensible encryption → typed unsupported-encryption)	✅

---

PowerPoint (.pptx)

Category	Feature	Status
Slides	Slide rendering	✅
	Slide layout / master inheritance	✅
	Slide size (custom dimensions)	✅
	Slide background (solid, gradient, image)	✅
	Slide numbers	✅
	Speaker notes (plain text via getNotes())	✅
	Markdown export (PptxPresentation.toMarkdown() — title slides → headings, body → nested bullets, notes / comments collated)	✅
	Animations / transitions	❌ Not planned
Element types	Shapes (sp)	✅
	Pictures (pic, with adaptive display-sized decoding for oversized rasters)	✅
	TIFF images (opt-in @silurus/ooxml/tiff; bounded bilevel, grayscale, RGB(A), process-CMYK and CCITT Group 4)	✅
	SVG images (asvg:svgBlip MS-2016 extension — vector drawn from the embedded .svg, PNG fallback)	✅
	Groups (grpSp) with nested transforms	✅
	Connectors (cxnSp)	✅
	Tables (tbl in graphicFrame)	✅
	Charts (bar, line, area, radar)	✅
	Charts (pie, doughnut)	✅
	Charts (scatter — scatterStyle marker / line / smooth variants)	✅
	Charts (bubble — bubbleSize per-point area scaling)	✅
	Charts (ordered classic combo groups — observed bar/line/area, scatter/bubble, and stock/line combinations; unsupported mixes fail closed)	✅
	ChartEx (waterfall / histogram / Pareto / funnel / box & whisker / treemap / sunburst)	✅ opt-in
	Charts (stock — high / low / close candlesticks)	✅
	SmartArt (renders the PowerPoint-saved drawing layout dsp:drawing, or a staged fallback to a text list when no drawing part is present; no native diagram layout engine)	✅
	OLE embedded objects (p:oleObj — the baked preview p:pic is drawn; the embedded app is not run)	✅
	Video / audio (poster + interactive playback)	✅
	Ink / handwriting (p:contentPart, raster fallback)	✅
Shape geometry	186 preset shapes (prstGeom — incl. 3D presets cube / can / bevel / frame)	✅
	Custom geometry (custGeom) on shapes and pictures (clipping)	✅
	Rotation and flip (flipH / flipV)	✅
Fills	Solid fill (solidFill)	✅
	Linear / radial gradient (gradFill)	✅
	No fill (noFill)	✅
	Pattern fill (pattFill) — 30 preset bitmaps incl. pct5–pct90 / horz / vert / cross / diag / grid / brick / check / trellis	✅
	Image fill on shapes (blipFill in sp)	✅
Strokes	Solid line color and width	✅
	Dash / dot styles	✅
	Arrow heads (headEnd / tailEnd)	✅
	Compound / double lines (`<a:ln cmpd="dbl	thinThick
	Picture border (a:ln on p:pic) — stroked along the clip silhouette	✅
Shape effects	Drop shadow (outerShdw)	✅
	Glow (glow — radius + colour)	✅
	Inner shadow (innerShdw)	✅
	Soft edge (softEdge)	✅
	Reflection (reflection)	✅
	3D camera / perspective projection (scene3d camera + rot) on pictures and shapes — projected shape text is drawn but not selectable	✅
	3D contour edge (sp3d contourW / contourClr) — flat approximation	⚠️
	Bevel shading (sp3d bevelT / bevelB) — distance-field lip lit by lightRig, matte/plastic materials	✅
	3D extrusion (sp3d extrusionH / extrusionClr) — swept side-wall approximation (visible only under a tilted camera)	⚠️
Text — characters	Bold, italic, strikethrough (incl. dblStrike)	✅
	Underline styles (sng / dbl / dotted / dash / dashLong / dotDash / dotDotDash / wavy / wavyDbl and *Heavy variants)	✅
	Per-run underline colour (uFill / uFillTx)	✅
	Font family, size, color	✅
	East Asian font (rPr > a:ea — separate typeface for CJK glyphs)	✅
	Symbol font runs (a:sym — e.g. Wingdings / Webdings glyphs)	✅
	Caps transform (all / small)	✅
	Letter spacing (spc)	✅
	Superscript / subscript	✅
	Hyperlinks (hlinkClick — theme hlink colour + auto underline)	✅
	Text shadow (rPr > effectLst > outerShdw)	✅
	Text outline (rPr > a:ln)	✅
	WordArt text warps (a:prstTxWarp, §20.1.9.19 — all 40 presets, per-glyph envelope fit incl. Follow Path)	✅
	Text highlight / marker (a:highlight — §21.1.2.3.4)	✅
	Math equations (OMML m:oMath / m:oMathPara, incl. a14:m / mc:AlternateContent; STIX Two Math via MathJax — opt-in @silurus/ooxml/math)	✅
Text — paragraphs	Horizontal alignment (left / center / right / justify)	✅
	Vertical anchor (top / center / bottom)	✅
	Line spacing (spcPct, spcPts)	✅
	Space before / after paragraph	✅
	Bullet points (character, auto-numbered, and picture a:buBlip §21.1.2.4.2)	✅
	Tab stops	✅
	Indent / margin	✅
	Vertical text (bodyPr@vert — vert / vert270 / eaVert)	✅
	Right-to-left text — UAX#9 bidi engine, pPr@rtl, RTL bullets, bodyPr@rtlCol column order, tblPr@rtl tables	✅
Text — body	Text padding (insets)	✅
	normAutoFit (shrink to fit)	✅
	spAutoFit (expand box; suppresses wrap when text fits in one line)	✅
	Word wrap / no wrap	✅
	Japanese kinsoku line breaking (a:pPr@eaLnBrk, §21.1.2.2.7 — 行頭/行末禁則, shared core engine)	✅
	Multi-column text body (numCol / spcCol — balanced flow)	✅
	Theme object-default inheritance (<a:objectDefaults><a:txDef|spDef> bodyPr fallback)	✅
Tables	Cells, rows, columns	✅
	Cell merges (horizontal / vertical)	✅
	Cell borders	✅
	Cell fills (solid / gradient)	✅
	Cell diagonal lines (lnTlToBr / lnBlToTr)	✅
	Table theme styles (74 built-in PowerPoint presets)	✅
Theme	Scheme colors (dk1/lt1/accent1–6)	✅
	Font scheme (+mj-lt, +mn-lt)	✅
	lumMod / lumOff / alpha transforms	✅
Interaction	Text selection, including table-cell text (transparent overlay, native copy)	✅
	Bounded text/element selection context (getSelectionContext(), element selection, master/layout/slide provenance, main + worker)	✅
	Comments — opt-in slide-side cards (comments: true), authored target markers and highlights, replies and resolved state; also available per slide through the presentation model	✅
	In-document find (findText / findNext / findPrev / clearFind — matches tagged with slide)	✅
	Runtime zoom (getScale / setScale / fitWidth / fitPage)	✅
	Clickable hyperlinks (onHyperlinkClick; internal slide-jump navigation)	✅
	Continuous scroll viewer (PptxScrollViewer — virtualized slide list, desk background / shadow, Ctrl/⌘+wheel zoom, borrowed-engine factory)	✅
Loading	Password-protected files ([MS-OFFCRYPTO] Agile Encryption — load(bytes, { password }), decrypted client-side via WebCrypto; legacy Standard / Extensible encryption → typed unsupported-encryption)	✅

---

A note on text selection. Across DOCX / PPTX / XLSX, text selection is currently implemented by rendering glyphs to the canvas while overlaying a transparent DOM layer that mirrors the canvas text positions for native browser selection. This dual-layer approach is a deliberate stop-gap: once the Canvas drawElement API (proposed in WICG/html-in-canvas, currently in Chromium Origin Trial) ships across browsers, the project plans to migrate to a single DOM-as-source-of-truth pipeline where the canvas mirrors the DOM directly — eliminating the duplication while keeping z-order correctness and native selection / a11y.

---

Companion packages

• packages/markdown/ — internal workspace adapter and ooxml-md development CLI for the same GitHub-flavoured Markdown projection exposed by each format model's toMarkdown() method.
• packages/node/ — the implementation behind the public Node-only @silurus/ooxml/node subpath. Its canonical APIs are the explicitly owned, bounded openPptxPresentation, openDocxDocument, and openXlsxWorkbook sessions. Async materializePptxPresentation, materializeDocxDocument, materializeXlsxWorkbookIndex, materializeXlsxWorksheet, and materializeXlsxWorkbook are provided when a complete caller-owned graph is actually needed. Each open* call returns an explicit, idempotent close()-able session; PPTX streams slides(), DOCX completes format-required sequential pagination before streaming pages(), and XLSX parses its workbook index once before sequential worksheetRows(sheetIndex) streams reuse the retained archive. Useful for CI checks and headless rendering pipelines; canvas rendering accepts a user-supplied backend such as skia-canvas without making it a runtime dependency. See the 0.75 to 0.76 migration guide for every removed synchronous helper and its replacement.
• packages/vscode-extension/ — VS Code extension (ooxml-viewer) that registers CustomEditorProviders for .docx, .xlsx, and .pptx, and (opt-in) auto-installs and registers the ooxml-mcp-server for GitHub Copilot Chat in Agent mode, including active Viewer selection. Claude Code and Codex can configure the same binary separately for path-based file tools, but do not receive the active selection bridge. The preview is offline by default; an opt-in ooxmlViewer.useGoogleFonts setting (off, and force-disabled in untrusted workspaces) surfaces the library's metric-compatible font substitution, widening the webview CSP to the Google Fonts CDN only while enabled.
• packages/mcp-server/ — Rust MCP server (ooxml-mcp-server) exposing the parsers as tools for AI agents (Claude, Copilot, Codex, etc.). Provides structured queries (docx_get_structure, xlsx_get_cell_range, pptx_get_slide_structure, …) so agents can inspect OOXML files without shelling out to unzip. Prebuilt binaries are attached to each GitHub Release for macOS / Linux / Windows; the VS Code extension downloads them on demand.

---

Development

# Install dependencies
pnpm install

# Build all WASM parsers (requires Rust + wasm-pack)
pnpm build:wasm

# Start Storybook dev server (port 6006)
pnpm storybook

# Type-check all packages
pnpm typecheck

# Run visual regression tests (local only — not run in CI)
pnpm vrt
# Adopt the current rendering as the new reference baseline
UPDATE_REFS=1 pnpm vrt

# Build the library
pnpm build

WASM build (individual packages)

cd packages/docx/parser && wasm-pack build --target web && cp pkg/docx_parser_bg.wasm  pkg/docx_parser.js  ../src/wasm/
cd packages/xlsx/parser && wasm-pack build --target web && cp pkg/xlsx_parser_bg.wasm  pkg/xlsx_parser.js  ../src/wasm/
cd packages/pptx/parser && wasm-pack build --target web && cp pkg/pptx_parser_bg.wasm pkg/pptx_parser.js ../src/wasm/

Error handling

Headless APIs (DocxDocument, XlsxWorkbook, and PptxPresentation) and Viewer APIs report failures from awaitable operations by rejecting the returned Promise. This includes viewer.load() parsing and its initial render, whether or not the Viewer has an onError(error) callback. A failure is never delivered through both channels.

Use onError for later Viewer-managed work that has no directly awaitable result, such as virtualized scroll-view rendering or embedded-media playback. Those failures are logged with console.error when the callback is omitted. PptxPresentation.presentSlide() follows the same boundary: initialization rejects its Promise, while PresentSlideOptions.onError observes only media decode or playback failures after the presentation handle has been returned.

Stable failures can be narrowed without parsing message strings:

• OoxmlError — container failures. Its code is encrypted, invalid-password, unsupported-encryption, legacy-binary-format, or not-ooxml.
• OoxmlResourceLimitError (code === 'ooxml-resource-limit') — a measured package or format resource crossed a configurable limit or hard ceiling. details.violation contains the resource, metric, limit, observed value, and usage snapshot.
• OoxmlDecodedImageLimitError (code === 'ooxml-decoded-image-limit') — a raster crossed an image pixel or active decoded-byte ceiling. Its metric, limit, and observed properties are stable.
• TiffDecodeError (code === 'ooxml-tiff-decode') — a recognized TIFF part is malformed, uses a class the configured codec does not support, or fails during bitmap handoff. Its message is diagnostic rather than a stable subtype. Omitting the optional TIFF codec is not an error; the affected image is shown as an unavailable-image placeholder while the rest of the document renders.
• An otherwise ordinary Error may carry code === 'parser-crashed' for a recognized WASM trap. This does not mean “OOM”: panic, allocation failure, stack overflow, and other traps can be indistinguishable at the current WASM boundary.

All other configuration, fetch, parser, renderer, worker, and media failures remain Error, TypeError, or RangeError values without a stable code. Their messages are diagnostic text, not a programmatic API.

import {
  DocxViewer,
  OoxmlDecodedImageLimitError,
  OoxmlError,
  OoxmlResourceLimitError,
  TiffDecodeError,
} from '@silurus/ooxml/docx';

const viewer = new DocxViewer(canvas, {
  // Background failures after an awaited operation has completed.
  onError(error) {
    if (error instanceof OoxmlResourceLimitError) {
      const { limit, observed } = error.details.violation;
      showTooLargeMessage({ limit, observed });
    } else if (error instanceof OoxmlDecodedImageLimitError) {
      showImageTooLargeMessage(error);
    } else if (error instanceof TiffDecodeError) {
      showUnsupportedImageMessage();
    } else if (error instanceof OoxmlError) {
      handleContainerError(error.code);
    } else {
      reportUnexpectedError(error);
    }
  },
});

try {
  await viewer.load(file);
} catch (error) {
  reportUnexpectedError(error);
}

Security & Privacy

• Canvas-only rendering. Documents are decoded and drawn to an HTMLCanvasElement. No script, link, form, or other active content from the source file is executed or injected into the DOM.

• Bounded OOXML package expansion. DOCX, XLSX, and PPTX use the same resource policy. By default, one archive entry may inflate to at most 128 MiB, the distinct entries visited during one package session may inflate to at most 256 MiB, and an archive may contain at most 4,096 entries. Override these budgets with a plain resourceLimits object on a viewer or load(...) call:

  new XlsxViewer(container, {
    resourceLimits: {
      maxArchiveEntryBytes: 64 * 1024 * 1024,
      maxTotalInflatedBytes: 192 * 1024 * 1024,
      maxArchiveEntries: 2048,
    },
  });

  maxArchiveEntryBytes applies to every XML, text, image, media, and other package part that the parser reads. maxTotalInflatedBytes counts the largest amount actually read from each distinct part during the lifetime of the loaded package; reading the same part again does not consume that budget twice. maxArchiveEntries bounds central-directory entries before the ZIP library allocates its owned index. Set an individual field to null to disable that configurable budget. Internal hard safety ceilings still apply, so disabling a budget does not make arbitrary archives acceptable. Values other than null must be positive safe integers; byte fields are expressed in bytes and the entry field is a count.

  A violation rejects with OoxmlResourceLimitError (code === 'ooxml-resource-limit'). Its structured details.violation reports the resource, metric, limit, observed value, usage snapshot, and part name when a particular part caused the failure. The deprecated maxZipEntryBytes option remains as a compatibility alias for resourceLimits.maxArchiveEntryBytes, but is scheduled for removal in a future breaking release; new code should use resourceLimits.

  Applications can collect the same data as a machine-readable OoxmlResourceMetrics report without enabling console output. This is useful for choosing limits from representative files in the application's own domain:

  new DocxViewer(canvas, {
    onResourceMetrics(metrics) {
      // Send only under your application's own consent/privacy policy.
      analytics.record('ooxml-resource-usage', metrics);
    },
  });

  After resource options validate, the callback also runs for failed document loads, for which no document/viewer engine is returned. It is not awaited; callback exceptions and rejected promises are ignored and never change the load result. On a successful load, every DOCX/XLSX/PPTX engine and Viewer also exposes await getResourceMetrics(); unlike the one-shot callback, this probes the archive owner and returns a fresh snapshot after lazy sheet, slide, image, font, or media access. A failed or unresponsive probe rejects within a bounded timeout rather than presenting stale counters as current. Collection for that getter is always active—debug controls only console presentation. The library itself does not transmit or persist the report.

  For interactive inspection, pass debug: true. The same report is formatted as one compact, Ratatui-inspired console card with the compressed source size, largest observed inflated entry, distinct inflated total, entry count, configured limits, and timing checkpoints:

  new DocxViewer(canvas, {
    debug: true,
    resourceLimits: {
      maxArchiveEntryBytes: 128 * 1024 * 1024,
      maxTotalInflatedBytes: 256 * 1024 * 1024,
      maxArchiveEntries: 4096,
    },
  });

  Browser DevTools receive typography-only %c styling (a fixed-width font stack, fixed size and line height, disabled ligatures, zero letter spacing, and preserved whitespace) so Unicode borders and gauges stay aligned. The library does not set foreground or background colors. Node and Worker consoles receive the same card as one plain argument without CSS or ANSI escapes.

  The report is content-free by construction: it does not include source URLs, filenames, package paths, document text, passwords, or raw error messages. It still contains document-derived sizes, counts, and timings, so applications remain responsible for consent, retention, and telemetry policy. The initial browser callback covers the underlying document/workbook/presentation factory and does not wait for a Viewer's first canvas paint; use getResourceMetrics() for the latest observed package counters. Bounded Node sessions accept both onResourceMetrics and debug. DOCX/PPTX report successful terminal metrics when their one-pass stream completes or the session is explicitly closed; XLSX reports success when the reusable workbook session is explicitly closed. Open-time and session-operation failures report immediately.

  Image decoding uses a separate adaptive resource policy shared by all three formats. Target-resizable rasters—browser-decodable formats and TIFF when its optional codec is configured—keep their native decode while the complete visible paint fits the default 128 MiB decoded budget and the per-surface ceilings, preserving established output. When native ownership would cross either boundary, the renderer decodes those rasters at the current canvas/DPR requirement instead of retaining source pixels the paint cannot show. If even those display targets exceed the aggregate budget, it applies one uniform quality reduction across them; a later zoom requests a sharper cache variant. Image-bearing paints for the same loaded document are serialized so overlapping paints cannot each consume the full budget, while the admitted paint still runs up to two decodes concurrently. Formats or sources that cannot be decoder-resized remain on their format-specific path under the hard surface ceilings. Applications with a known environment can override the aggregate budget on Viewer or per-render options:

  new PptxViewer(canvas, {
    imageResources: {
      decodedByteBudget: 256 * 1024 * 1024,
      strategy: 'adaptive', // or 'strict' to reject instead of reducing quality
    },
  });

  decodedByteBudget accepts a positive safe integer from 4 bytes through 512 MiB. This configures planned and retained decoded RGBA ownership; it does not measure browser decoder intermediates or disable the encoded-source, per-axis, or per-surface hard safety ceilings. A strict aggregate crossing or any hard-ceiling crossing rejects with OoxmlDecodedImageLimitError (code === 'ooxml-decoded-image-limit') instead of silently omitting the image.

  The package counters and raster-image guards are deterministic admission limits, not exact JavaScript/WASM process-memory accounting. XML trees, document models, canvas backing stores, browser decoder overhead, renderer state, and browser-managed SVG/vector parse or decoded storage can still require several times the measured input. SVG has no portable decoded-byte measure or explicit browser release primitive; the library count-bounds its cache and revokes owned object URLs, but cannot charge it as RGBA bytes. The defaults therefore reduce risk but cannot promise that an OOM is impossible on every device. Running parse and render work in mode: 'worker' can contain many failures away from the main UI thread, but a Worker is not a separate operating-system process or a strict memory sandbox.

  A measured limit crossing is reported as OoxmlResourceLimitError. A residual WASM failure that reaches a recognized trap-shaped boundary is reported conservatively as parser-crashed, not parser-oom: with the current aborting Rust/WASM boundary, panic, allocation failure, explicit unreachable, and stack overflow can lose their distinct causes and converge on the same generic runtime error. Inferring OOM from an exception class or message would misclassify some parser defects as large-file failures. Reliable OOM classification would require preserving a structured cause before the trap across every relevant allocation path; it cannot be recovered from the generic trap afterward. The WebAssembly JavaScript embedding also permits implementation-defined stack/OOM failures, including an indistinguishable plain Error or process termination, so converting and poisoning every engine-level failure cannot be guaranteed.

• No network by default. The library does not send telemetry or analytics, and does not contact third-party services unless you ask it to. In particular, theme webfonts, Office font metric substitutes (Carlito/Caladea), and the script fallback fonts are not loaded from Google Fonts unless you pass useGoogleFonts: true to the relevant Viewer / load(...) options — supported uniformly by DocxViewer, PptxViewer, XlsxViewer, and XlsxSheetViewer. When enabled, fonts for non-Latin scripts are supplied on demand from Noto families so text does not fall back to tofu: Arabic (Noto Naskh/Sans Arabic), CJK (Noto Sans/Serif KR · SC · TC · JP, picked per document language so shared Han glyphs take the right shapes), Cyrillic (Noto Sans/Serif), Hebrew (Noto Sans/Serif Hebrew, RTL), Thai (Noto Sans Thai) and Devanagari (Noto Sans Devanagari). No font binaries ship in the bundle. Enabling this option causes the end-user's browser to send an HTTP request (IP and User-Agent) to fonts.googleapis.com, which may have GDPR implications for your application — consider self-hosting the required fonts via @font-face instead.

• XML parsing. Uses roxmltree, which does not resolve external entities (XXE-safe by default).

• Encrypted OOXML ([MS-OFFCRYPTO] Agile Encryption). Password-protected .docx / .xlsx / .pptx files are OLE2/CFB containers, not ZIPs. Pass password to load(...) and the file is decrypted client-side via WebCrypto — no bytes and no password leave the browser:

  const doc = await DocxDocument.load(bytes, { password: 'secret' });

  Key derivation (SHA-512 spin, commonly 100,000 iterations) and AES-CBC segment decryption run on the main thread and add roughly a second before parsing. Failures are typed OoxmlErrors: no password on an encrypted file → encrypted, wrong password → invalid-password, a non-Agile scheme (legacy Standard / Extensible encryption, or an encrypted legacy binary .doc/.xls/.ppt) → unsupported-encryption. Note: decryption recovers the plaintext but does not verify the file's HMAC data-integrity tag ([MS-OFFCRYPTO] §2.3.4.14), so tampering with the ciphertext is not detected — treat decrypted output from untrusted sources with the same care as any other input.

License

MIT

Third-Party Notices

The library's own code is MIT-licensed. It also bundles a small set of permissively-licensed third-party components — see THIRD_PARTY_NOTICES.md (included in the npm tarball) for the full list and license texts. Highlights:

• MathJax + STIX Two Math (Apache License 2.0) — the equation-rendering engine behind the opt-in @silurus/ooxml/math entry described in Rendering equations. It ships in the tarball as a standalone ~3 MB asset but is never loaded by a consuming app unless that app imports @silurus/ooxml/math and the viewer is handed a document that actually contains an equation.
• Rust crate dependencies of the WASM parsers (docx/pptx/xlsx) — all MIT / Apache-2.0 (or compatible permissive licenses), no copyleft.