HTML in, Markdown out, tokens down 92%

What it does

mdream converts HTML to GitHub Flavored Markdown, tuned specifically as LLM input rather than for human prettiness. It ships as a Rust NAPI engine (with WASM for edge runtimes) plus a pure-JS engine with zero native dependencies, and it powers Cloudflare Browser Run’s /markdown and /crawl endpoints — a decent hint it holds up under real traffic.

The interesting bit

It converts as HTML streams in, and the joined output matches a one-shot conversion — so you can start feeding a crawler or LLM pipeline before the page finishes downloading. The token math is the real pitch: up to 92% fewer tokens than raw HTML on their measured pages, which directly cuts cost and context waste.

Key highlights

• Dual engines: Rust NAPI for performance, @mdream/js for tree-shakable, dependency-free use
• Streaming conversion with output identical to one-shot mode
• clean mode strips tracking params, empty links, and redundant anchors
• Ecosystem around the core: crawler for llms.txt generation, Vite plugin, GitHub Action, Docker images, and a Rust crate
• Extraction hooks let you pull out headers or images mid-conversion

Caveats

• The Rust engine loads native bindings, so bundlers may need it marked external (Next.js, Webpack, etc.)
• The 92% token-reduction figure applies only to their measured pages, not a universal guarantee
• The package sprawl (seven-plus ways to install it) takes a moment to navigate

Verdict

If you’re building anything that feeds web pages into an LLM — RAG, crawlers, agent tooling — this is squarely aimed at you. If you just want pretty Markdown for a human reader, a simpler converter will do.