A translator between the messy web and your token-starved agent

What it does

Give it a URL or upload a file and it hands back clean Markdown: web pages, PDFs, Office docs, EPUB, images, audio, and YouTube transcripts. It runs as a self-hosted service with a web UI, REST API, MCP server, and a Claude Code skill. Metadata lives in YAML frontmatter so the body stays pure content.

The interesting bit

The extraction is a fallback chain — Cloudflare’s native Markdown, then Readability and Trafilatura with the best result winning, then headless Chromium for JavaScript-heavy sites. There’s also query-scoped extraction: pass a question and get only the relevant sections. The README’s own numbers show a Reddit thread that plain curl can’t even fetch (bot check) coming back as ~6,400 tokens, or ~640 with a query attached.

Key highlights

• Reddit and Hacker News threads with full comment trees, depth and count configurable
• YouTube transcripts with clickable timecodes, no API key needed
• Live share links (/s/:id) that re-fetch stale caches and serve the last good snapshot if the source dies
• Every sidecar is optional — drop Playwright or MarkItDown and it degrades gracefully
• SSRF protection, rate limits, and auth modes from open homelab to multi-user

Caveats

• The Playwright sidecar adds ~3.7 GB of image cache (Chromium, Firefox, WebKit) — optional, but the JS-heavy-site fallback is why you’d want it
• Reverse-proxy setups need PULLMD_TRUST_PROXY configured correctly or all clients share one rate-limit bucket
• Token-savings numbers are the project’s own measurements on three pages, not an independent benchmark

Verdict

If you run agents that read the web — especially via Claude Code or MCP — this is a genuinely useful middleman. If you just need one-off HTML-to-Markdown, a library will do the job with less ceremony.