DenseReader

Summarize any webpage using Chain-of-Density with adjustable density (1–5). On-device AI, privacy-first.

DenseReader — Turn long articles into dense, accurate summaries. Powered by Chain-of-Density. Stop scrolling. Stop skimming. Stop pretending you read it. DenseReader is an AI-powered reading companion that distills dense articles, research papers, YouTube videos, and uploaded documents into information-rich summaries you can actually trust. Built on the Chain-of-Density (CoD) prompting technique — the same method published by Salesforce AI Research — DenseReader produces summaries that keep adding entities, facts, and context across multiple passes, so you get the substance of a piece, not just the surface. If you've ever watched a "summarize this page" extension spit out three generic bullet points that could describe any article on earth, you already know the problem. DenseReader fixes it. WHY DENSEREADER IS DIFFERENT Most summarizers compress by deleting. DenseReader compresses by densifying. Here's the difference in practice. A typical AI summarizer reads a 5,000-word investigation into, say, a biotech scandal, and returns: "The article discusses a biotech company facing regulatory challenges. It covers their response and the market impact." Technically correct. Practically useless. DenseReader runs the same article through a multi-step Chain-of-Density loop. Each iteration forces the model to add missing entities — names, numbers, dates, places, relationships — without increasing the length. The result is a summary at the same word count as the lazy one, but packed 3-5x more facts per sentence. You walk away actually knowing what happened. That's what we mean by "dense." FIVE DENSITY LEVELS, ONE SLIDER DenseReader gives you a single slider labeled 1 to 5: Gist (1) — One tight paragraph, ~150 words. The "tell me in 30 seconds" setting. Perfect for triaging your reading list. Short (2) — ~400 words. The "I need to sound smart at a meeting" setting. Covers the main argument plus 2-3 supporting points. Medium (3) — ~600 words. The balanced default. Readable in 3 minutes, retains the article's structure and nuance. Detailed (4) — ~1,100 words. For when you actually need the facts. Every named entity, every key number, every cited source preserved. Comprehensive (5) — ~2,000 words. The "I have to present on this tomorrow" setting. Near-exhaustive coverage with all supporting detail. Because CoD is iterative, each level isn't just "more words" — it's more information per word. Level 5 doesn't pad; it densifies further. FOUR WAYS TO FEED IT CONTENT DenseReader is not locked to the current tab. Open the side panel and pick your input: Page — The article on your active tab. DenseReader extracts clean reader-view text using a Readability-style parser, stripping ads, nav, comments, and boilerplate. Selection — Highlight any text on any page, and summarize just that. Perfect for long Hacker News threads, Reddit discussions, or when you only care about one chapter of a long post. Paste — Drop in plain text, a transcript, meeting notes, emails — anything you can copy. No page needed. Upload — Summarize .md, .txt, .docx, and .pdf files up to 10 million tokens. Yes, ten million. Academic papers, legal documents, book chapters, whitepapers — DenseReader handles them with a hybrid chunk-and-reduce pipeline that never truncates your source. YOUTUBE TRANSCRIPT SUMMARIES Watching a 90-minute lecture in 1.25x still takes 72 minutes. DenseReader reads it in under 30 seconds. When you're on a YouTube video page, a banner appears in the panel: "YouTube video detected — Get transcript." One click grabs the official auto-generated transcript and feeds it into the same Chain-of-Density pipeline. You get a structured summary with timestamps preserved, speakers identified where possible, and all the facts, numbers, and claims the video actually made. Works on talks, interviews, tutorials, earnings calls, podcasts uploaded to YouTube — anything with a transcript available. THREE AI BACKENDS, AUTO-SELECTED DenseReader is designed for reliability. It uses up to three AI backends, picked in order: Chrome's built-in Prompt API (on-device) — If you're on a supported Chrome build with Gemini Nano enabled, summaries run entirely on your device. Zero network traffic, zero cost, full privacy. Chrome's built-in Summarizer API (on-device) — Fallback on-device option when Prompt API isn't available. DenseReader cloud proxy — For devices that don't support on-device AI, or when you want higher-quality models (DeepSeek V4 Flash with 1M token context), DenseReader proxies your request through our Cloudflare Worker. No API key required on the free tier. You never have to configure model endpoints, paste API keys, or babysit fallbacks. DenseReader detects what's available and uses the best option automatically. The panel shows a live status dot: green when ready, red when the backend fails, so you always know what's running. HALLUCINATION VERIFICATION, BUILT IN Every DenseReader summary passes through a post-generation verification step. We extract named entities from the summary, cross-check them against the source, and flag any entity that appears in the summary but not the source. When that happens, you see a yellow "Potential hallucination detected" badge on the output. You'll rarely see it — Chain-of-Density is inherently grounded because each step explicitly pulls entities from the source — but when you do, you know to double-check before citing. No other browser summarizer gives you this. A visible "Entities Mentioned" tag list accompanies every summary so you can scan who and what the article is actually about in one glance. EXPORT TO YOUR SECOND BRAIN A summary is only useful if you can find it again. DenseReader ships with one-click exports to the tools knowledge workers actually use: Readwise — Send any summary straight to your Readwise highlights. Paste your Readwise API token once in Settings, then every summary is one click away from joining your daily review. Notion — Push summaries into a Notion database of your choice. Works with both full database URLs and raw 32-character IDs. DenseReader auto-extracts the ID so you don't have to fight with Notion's URL format. Copy to clipboard — Plain text, markdown-friendly, nothing weird. Download — Save as .md for your Obsidian, LogSeq, or plain-file workflow. Summaries include the article title, source URL, and density level used, so exports are self-describing when you revisit them months later. Summaries themselves are produced in the language best suited to the source content. THE READER OVERLAY Sometimes you want summary and source side by side. Click the expand button on any output, and DenseReader opens a distraction-free reader overlay: original article on one side, summary on the other, synced scrolling. Read with comprehension instead of skimming with anxiety. Prefer a popup tab instead of an overlay? Toggle it in Settings. Some sites (mainly those with aggressive CSP headers) block overlays — DenseReader remembers which sites fail and automatically falls back to a popup tab for those. Your "Sites where expand didn't work" list is visible in Settings if you ever want to audit or clear it. FREE vs PRO DenseReader is free to use. The free tier includes: Density levels 1 (Gist), 2 (Short), and 3 (Medium) 10 summaries per day via the cloud proxy Full on-device summarization (unlimited, if your Chrome supports it) Copy and download exports English-only interface DenseReader Pro unlocks: All five density levels including Detailed and Comprehensive Bring your own key. Use your own apis including ollama. Unlimited summaries per day via the cloud proxy Readwise and Notion exports Priority support Upgrade anytime from the Settings panel. Your Pro license is tied to a key, so you can use it across devices. Past-due or expired licenses gracefully fall back to the free tier — no surprise lockouts. PRIVACY THAT ISN'T MARKETING Privacy is a technical decision, not a promise. Here's how DenseReader actually handles your data: On-device mode — When your browser supports Chrome's Prompt API or Summarizer API, your page content never leaves your device. No network request is made. No server sees your reading. Cloud proxy mode — When on-device AI isn't available, DenseReader sends the article text to our Cloudflare Worker, which forwards it to DeepSeek for summarization. We do not log article content. We do not train on your data. We do not share it with anyone except the AI provider required to generate the summary. Rate-limit counters are the only per-request telemetry we keep, and they are tied to an ephemeral request UID, not to you. Settings and license — Your license key, Readwise token, and Notion credentials are stored in Chrome's extension-local storage.sync API. They never leave your browser except when you explicitly export to Readwise or Notion. No analytics, no trackers, no ads. DenseReader has zero third-party scripts outside the fonts loaded from Fontshare for the UI. Full privacy policy: densereader.com/privacy DESIGNED FOR ACTUAL READING DenseReader's side panel is typographically obsessive on purpose. We use Boska for display type and Satoshi for body text — two typefaces designed for reading, not for looking like a SaaS dashboard. Line lengths, contrast ratios, and hierarchy are tuned for 3-minute reading sessions rather than 3-second glances. Dark mode follows your OS preference automatically. WHO IS DENSEREADER FOR Researchers and PhD students who need to triage 30 papers a week Lawyers and analysts who read long documents professionally Product managers drowning in user-research transcripts and competitive-intel reports Founders and VCs who can't afford to miss the third paragraph of anything Anyone who has a "read later" list that has quietly metastasized into a "read never" list If you read for a living, DenseReader gives you an hour of your day back. If you read for pleasure, it finally lets you keep up with the backlog. GETTING STARTED Install DenseReader and pin it to your toolbar. Click the icon to open the side panel. Navigate to any article. DenseReader detects the page automatically. Pick a density level with the slider. Hit Summarize. Your first summary arrives in under 10 seconds on most articles. Longer pieces (academic papers, full books) take 30-90 seconds due to chunked processing. SUPPORT & FEEDBACK DenseReader is actively developed. We respond to every bug report, feature request, and typo correction. Email support@densereader.com or use the feedback link in Settings. Known limitations we're working on: no Firefox version yet (Manifest V3 differences), no voice output (planned), no offline mode outside on-device AI (by design). Made with insomniac devotion in Hong Kong. Start reading denser today.