Parse the change. Not the conversation.
HyperMarkdown is a streaming-native Markdown renderer built for LLM output. It caches settled content down to code lines, table rows, and list items, so growing responses do not keep paying to parse and render work that is already finished.
1.6×–10.6× faster than the nearest streaming renderer across our benchmark suite.
| Workload | HyperMarkdown | Markstream | Streamdown | DeepSeek Harness | react-markdown |
|---|---|---|---|---|---|
| Large code block | 216 ms | ||||
| 769 ms | 3,242 ms | 4,416 ms | 2,054 ms | ||
| Mixed prose | 153 ms | ||||
| 313 ms | 559 ms | 249 ms | 2,629 ms | ||
Captured AI code stream (real-code-os ) |
|||||
| 659 ms | |||||
| 4,417 ms | 12,511 ms | 14,621 ms | 8,008 ms | ||
Captured AI table stream (real-table-head ) |
|||||
| 654 ms | |||||
| 4,948 ms | 11,621 ms | 19,644 ms | 10,236 ms | ||
| Large table | 874 ms | ||||
| 9,276 ms | 33,142 ms | 55,918 ms | 29,747 ms |
The captured model fixtures are not generated stress cases: their content is real AI output, replayed in controlled 8-character frames. HyperMarkdown renders the code stream in 659 ms versus 4,417 ms for the next closest streaming renderer, and the table stream in 654 ms versus 4,948 ms.
On a large streaming table, HyperMarkdown completes the workload in under one second.
- Streamdown: 33 seconds - DeepSeek Harness strategy: 56 seconds - react-markdown: 30 seconds
Same Markdown. Same stream. Very different architecture.
Production React benchmark on an Apple M2 Max, including chunk processing and synchronous render/commit. Absolute timings vary; the ratios are the useful comparison.
Read the methodology · View the full benchmark results
Most streaming Markdown renderers optimize at the document or block level. HyperMarkdown goes further.
Traditional streaming renderer
new token
↓
growing active block
↓
parse the active block again
↓
render again
HyperMarkdown:
new token
↓
active block
│
├── settled code lines → cached
├── settled table rows → cached
├── settled list items → cached
└── changing frontier → parse
A 1,000-line code block does not become a 1,000-line parsing problem every time another token arrives.
Completed work stays completed.
- Sub-block caching for code, tables, and lists
- Streaming-safe handling of incomplete Markdown
- GFM tables, task lists, autolinks, and footnotes
- Reasoning blocks written as
<think>
,<thinking>
, or<reasoning>
- KaTeX math
- Mermaid diagrams
- Syntax highlighting
- Raw HTML with sanitization
- React 18 and React 19
- SSR and hydration, including a Next.js App Router client boundary
- Lightweight core with heavy features loaded as optional plugins
- Production integration with DeepSeek Harness / DSH architecture
HyperMarkdown is the official Markdown component used by Æven and integrates with the DeepSeek Harness (DSH) architecture. It was built around the demands of an agent harness—long answers, dense code, wide tables, reasoning traces, and many small deltas—not adapted from a finished-document renderer after the fact.
The captured AI workloads in the benchmark suite come from that environment. They exercise the content shapes a renderer encounters in a real model response, with controlled chunk sizes that keep comparisons reproducible.
npm install @aeven-ai/hypermarkdown
React 18 or 19 is required as a peer dependency.
Import the component stylesheet once in your application entry point:
import "@aeven-ai/hypermarkdown/styles.css";
Then import the component:
import {
HyperMarkdown,
type HyperMarkdownHandle,
} from "@aeven-ai/hypermarkdown";
Use the md
prop when the complete document already exists:
<HyperMarkdown md={markdown} />
Updating md
replaces the document. This mode works naturally for stored messages, previews, and server-rendered content.
Mount one renderer for the active response and write each incoming delta to its imperative handle:
import { useRef } from "react";
import {
HyperMarkdown,
type HyperMarkdownHandle,
} from "@aeven-ai/hypermarkdown";
function Chat() {
const renderer = useRef<HyperMarkdownHandle>(null);
async function generate(prompt: string) {
// Reuse the mounted component for a new response.
renderer.current?.reset();
const deltas = await createResponseStream(prompt);
try {
for await (const delta of deltas) {
renderer.current?.write(delta);
}
} finally {
// Flush the final open paragraph, fence, list, table, or reasoning block.
renderer.current?.write("", true);
}
}
return (
<HyperMarkdown
ref={renderer}
streaming
animation
/>
);
}
createResponseStream()
above represents your SDK or transport. It only needs to yield the text fragments produced since the previous event.
write()
appends. Pass the new fragment exactly once:
renderer.current?.write(delta); // correct
renderer.current?.write(fullText); // wrong: repeats everything already written
When the stream ends, finalize it exactly once. Either form is valid:
renderer.current?.write("", true); // separate finalization
renderer.current?.write(lastDelta, true); // final delta and finalization
Finalization matters even when the visible text looks complete: it settles the active frontier and lets the renderer finish incomplete-block bookkeeping.
Before using the same mounted component for another response, call reset()
.
Keep the component and its ref
mounted during a response; changing its React
key
creates a new, empty renderer.
Some APIs emit "Hello"
, then "Hello world"
, rather than "Hello"
, then
" world"
. Convert those snapshots to deltas at the boundary:
let previous = "";
function startSnapshotStream() {
previous = "";
renderer.current?.reset();
}
function writeSnapshot(next: string, final = false) {
const handle = renderer.current;
if (!handle) return;
if (!next.startsWith(previous)) {
// The provider revised an earlier prefix. Rebuild from the new snapshot.
handle.reset();
previous = "";
}
handle.write(next.slice(previous.length), final);
previous = next;
}
Call startSnapshotStream()
before the first snapshot of each response.
Do not put the growing Markdown string in React state just to feed it back as a prop on every token. In streaming mode, HyperMarkdown owns that buffer so your component tree does not have to.
HyperMarkdown supports Next.js server rendering and hydration. Its public
component entry includes "use client"
, so an App Router Server Component can import it directly. The Client Component is still prerendered into the initial HTML and hydrated in the browser.
Import the stylesheet once in the root layout:
// app/layout.tsx
import "@aeven-ai/hypermarkdown/styles.css";
export default function RootLayout({ children }: { children: React.ReactNode }) {
return (
<html lang="en">
<body>{children}</body>
</html>
);
}
Then render finished Markdown from a Server Component:
// app/page.tsx
import { HyperMarkdown } from "@aeven-ai/hypermarkdown";
export default async function Page() {
const markdown = await loadMarkdown();
return <HyperMarkdown md={markdown} />;
}
The md
prop is serializable and can cross the Server Component boundary. Create plugins, component overrides, refs, and callbacks inside a Client Component because they contain functions. Streaming through the imperative handle also belongs in a Client Component.
HyperMarkdown does not normally need dynamic(..., { ssr: false })
; disabling SSR removes the rendered Markdown from the initial HTML.
See the full SSR, hydration, and Next.js guide for App Router plugins and streaming, Pages Router SSR, and hydration-mismatch guidance.
For finished documents, the migration is a component swap:
// Before
<ReactMarkdown>{markdown}</ReactMarkdown>
// After
<HyperMarkdown md={markdown} />
For streaming, the architectural change is more important. A typical prop-based loop rebuilds an accumulated string and reparses it on every chunk:
// Before: full document goes back through React on every delta.
const [markdown, setMarkdown] = useState("");
for await (const delta of stream) {
setMarkdown((current) => current + delta);
}
<ReactMarkdown>{markdown}</ReactMarkdown>
Replace that state loop with one stable HyperMarkdown
instance:
// After: only the new fragment enters the renderer.
const renderer = useRef<HyperMarkdownHandle>(null);
for await (const delta of stream) {
renderer.current?.write(delta);
}
renderer.current?.write("", true);
<HyperMarkdown ref={renderer} streaming />
Common migration mappings:
| Previous pattern | HyperMarkdown |
|---|---|
Markdown passed as children |
|
md={markdown} for finished content |
|
| Accumulated text prop updated per token | write(delta) with streaming |
| Clear state before a new answer | ref.current?.reset() |
| End-of-stream state flag | write("", true) |
| GFM remark plugin | Built in |
| Custom element renderers | components={{ ... }} |
| Math, highlighting, Mermaid | Optional plugins slots |
| Raw HTML plugin | Built in; sanitized by default |
HyperMarkdown does not accept arbitrary remarkPlugins
or rehypePlugins
through the component API. Use its typed feature plugins and component overrides; if you depend on a custom AST transform, verify that transform before replacing the old renderer.
MarkdownStream
remains exported for compatibility but is deprecated. Mount
HyperMarkdown
, hold a HyperMarkdownHandle
, and replace direct engine calls
with write(delta)
, write("", true)
, and reset()
. The component owns the store and subscribes React to it safely.
Math, syntax highlighting, diagrams, and CJK-friendly emphasis are optional. Install only what your application uses:
npm install katex remark-math rehype-katex
npm install rehype-highlight
npm install mermaid
npm install remark-cjk-friendly
js
import { katexPlugin } from "@aeven-ai/hypermarkdown/plugins/math";
import { highlightPlugin } from "@aeven-ai/hypermarkdown/plugins/code";
import { mermaidPlugin } from "@aeven-ai/hypermarkdown/plugins/mermaid";
import { cjkPlugin } from "@aeven-ai/hypermarkdown/plugins/cjk";
import "katex/dist/katex.min.css";
// Build this once. A new plugin object rebuilds the processing pipelines.
const plugins = {
math: katexPlugin(),
code: highlightPlugin(),
diagram: mermaidPlugin({ theme: "neutral", fontFamily: "Inter" }),
cjk: cjkPlugin(),
};
<HyperMarkdown md={markdown} plugins={plugins} />
Each missing plugin degrades gracefully:
| Missing plugin | Behavior |
|---|---|
math |
|
$x$ remains literal text |
|
code |
|
| Code blocks retain caching, controls, and line numbers but are not highlighted | |
diagram |
|
A mermaid fence renders as an ordinary code block |
|
cjk |
|
| Standard CommonMark emphasis rules apply |
Mermaid is dynamically imported. With preload
off, starts when an
opening Mermaid fence appears, overlapping the rest of the stream. Set
preload
when a view is very likely to contain diagrams and should begin the download on mount.
LLM chunks end in inconvenient places. HyperMarkdown treats partial syntax as a normal state, not an error:
- Half-written links, autolinks, HTML tags, and math are withheld until safe to render.
- Emphasis resolves eagerly when the CommonMark delimiter rules make it unambiguous.
- Open code fences, tables, and lists render their stable content while the unfinished frontier continues changing.
- A finalized stream matches the whole-document parse across the correctness fixture suite.
Model reasoning wrapped in <think>
, <thinking>
, or <reasoning>
becomes a collapsible block. It stays open while tokens arrive and collapses when the block finishes.
<think>
Checking the constraints first.
</think>
The answer is 42.
Markdown inside the reasoning block is rendered normally. A partial opening
tag such as <thi
is withheld instead of flashing as text.
To place reasoning outside the answer container, provide a portal target:
const reasoning = useRef<HTMLDivElement>(null);
return (
<>
<div ref={reasoning} />
<HyperMarkdown
ref={renderer}
streaming
reasoningTarget={() => reasoning.current}
/>
</>
);
Set controls={{ reasoning: false }}
to render the content without the
collapsible wrapper. Override translations.thinking
and
translations.thoughtFor
to localize its labels.
Markdown produced by a model is untrusted input. HyperMarkdown defaults to
html="sanitize"
: raw HTML is parsed, then cleaned before math, syntax highlighting, diagrams, or animation run. Scripts, styles, iframes, forms, and event-handler attributes are removed.
Links and images are checked separately. By default, http
, https
,
mailto
, and tel
protocols are allowed, as are data:
images.
Choose the policy explicitly when needed:
<HyperMarkdown md={markdown} html="literal" />
html mode |
Behavior |
|---|---|
"sanitize" |
Default. Parse raw HTML and remove anything outside the schema. |
"literal" |
Render raw HTML as visible text. Strongest option for untrusted output. |
"raw" |
Parse without sanitization. Use only for content you control. |
To widen the default policy without disabling it:
<HyperMarkdown
md={markdown}
allowedTags={{ mention: ["data-user-id"] }}
linkSafety={{ allowedLinkPrefixes: ["https://docs.example.com/"] }}
/>
sanitize={false}
is retained for compatibility and selects raw mode when
html
is not set. Prefer the clearer html
prop in new code.
The shipped stylesheet is scoped under .hypermarkdown
. Customize it with CSS variables rather than overriding internal selectors:
.assistant-message {
--hm-font: Inter, sans-serif;
--hm-font-mono: "Geist Mono", monospace;
--hm-color: #171717;
--hm-background: #f5f5f5;
--hm-link-color: #2563eb;
--hm-radius: 16px;
--hm-max-width: 100%;
}
<HyperMarkdown className="assistant-message" md={markdown} />
The root always receives hypermarkdown
; className
is added alongside it. KaTeX requires its own stylesheet when the math plugin is enabled.
Replace rendered tags with stable React component references:
const components = {
a: AppLink,
img: ProxiedImage,
code: Code,
};
<HyperMarkdown md={markdown} components={components} />
HyperMarkdown already provides specialized renderers for links, images, code blocks, tables, and diagrams. An override wins over the built-in component. Keep the object and component functions stable—recreating them on every render can remount rendered elements.
| Prop | Type | Description |
|---|---|---|
md |
||
string |
||
Finished Markdown. Ignored while streaming is true. |
||
streaming |
||
boolean |
||
| Receive content through the imperative handle. | ||
animation |
||
boolean |
||
| Fade arriving words in. | ||
plugins |
||
PluginConfig |
||
| Optional math, code, diagram, and CJK plugins. | ||
preload |
||
boolean |
||
| Begin the configured diagram engine on mount. | ||
components |
||
RendererComponents |
||
| Stable tag-to-component overrides. | ||
html |
||
| `"sanitize" | "literal" | "raw"` |
Raw HTML policy. Defaults to "sanitize" . |
||
allowedTags |
||
Record<string, string[]> |
||
| Additional sanitized tags and attributes. | ||
linkSafety |
||
LinkSafetyConfig |
||
| Allowed URL protocols and prefixes. | ||
reasoningTarget |
||
| `HTMLElement | null | () => HTMLElement |
| Optional portal target for reasoning. | ||
controls |
||
ControlsConfig |
||
| Configure or hide reasoning, code, table, and diagram controls. | ||
translations |
||
Partial<Translations> |
||
| Override UI strings. | ||
icons |
||
Partial<IconMap> |
||
| Override toolbar icons with inline SVG strings. | ||
lineNumbers |
||
boolean |
||
Show code line numbers. Defaults to true . |
||
codeBlockMaxHeight |
||
| `number | string` | |
| Height at which code blocks scroll; numbers are pixels. | ||
tableMaxHeight |
||
| `number | string` | |
| Height at which tables scroll; numbers are pixels. | ||
scrollDown |
||
() => void |
||
| Runs after each committed update for host scroll management. | ||
onFullscreenChange |
||
(fullscreen: boolean) => void |
||
| Reports code, table, or diagram fullscreen changes. | ||
onAlert |
||
(alert: HyperMarkdownAlert) => void |
||
| Lets the host present block alerts. | ||
className |
||
string |
||
Additional class on the .hypermarkdown root. |
| Member | Description |
|---|---|
write(delta, finalize?) |
|
Append one delta; pass true once at end of stream. |
|
reset() |
|
| Discard rendered content and start a new stream. | |
store |
|
| The component's rendering store for advanced integrations. | |
stream |
|
Deprecated alias for store . |
CommonMark plus GitHub Flavored Markdown: tables, task lists, strikethrough, autolinks, and footnotes. Optional plugins add KaTeX math, Mermaid diagrams, syntax highlighting, and CJK-friendly emphasis. Raw HTML is supported under the selected safety policy.
npm install
npm test
npm run test:coverage
npm run typecheck
npm run lint
npm run build
npm run benchmark
npm run website:dev
The documentation site and playground live in website/
and deploy to
GitHub Pages from
.github/workflows/pages.yml
. Set the repository Pages source to GitHub Actions.
Pull requests and pushes to main
run lint, typecheck (React 18 and 19), unit
and full coverage, and a production build. Both coverage gates are 100%
statements, lines, functions, and branches. Coverage reports are uploaded as
artifacts and posted on pull requests. Mark the CI job as a required status
check on main
if you want GitHub to block merges on a red build.
The correctness suite compares finished streaming output with whole-document rendering across the fixture corpus. The benchmark harness validates DOM output and measures chunk processing plus synchronous React commits.
Publishing to npm is triggered by a GitHub Release whose tag matches
package.json
:
- Bump
version
inpackage.json
(and the lockfile). - Commit, tag
vX.Y.Z
, and push. - Create a GitHub Release from that tag.
The publish workflow re-runs the CI gates and
publishes @aeven-ai/hypermarkdown
with npm trusted publishing
(OIDC, no long-lived token, provenance generated automatically). Do not set
NODE_AUTH_TOKEN
or setup-node
's registry-url
; those force classic auth
and the publish fails with E404
.
One-time setup on npmjs.com → package Settings → Trusted Publisher:
| Field | Value |
|---|---|
| Organization or user | Aeven-AI |
| Repository | HyperMarkdown |
| Workflow filename | publish.yml |
| Environment name | npm |
| Allowed actions | npm publish |
Create a GitHub Environment named npm
if it does not exist. Optional required reviewers on that environment add a human approval step before npm sees the package.
Project layout #
index.tsx public React component and types
lib/
renderer.tsx buffers, caching, and incremental rendering
processors.ts Markdown and HTML processing pipelines
stream/ block detection and boundaries
repair/ safe handling of incomplete inline syntax
cache/ code-line, table-row, and list-item caches
plugins/ optional math, code, Mermaid, and CJK adapters
code/ table/ mermaid/ rich block renderers and controls
reasoning/ streamed reasoning UI
sanitize.ts HTML and URL safety policy
styles/hypermarkdown.scss scoped component stylesheet
tests/ correctness, streaming, API, UI, and security
benchmarks/ fixtures, competing renderers, and results
example/ browser example using the built package
MIT