Decant - In-Browser Document & Web Page to Markdown for AI, Cut Token Cost

Decant converts documents — and whole web pages — to Markdown before they reach an LLM chat, locally, in your browser, without interrupting how you already work. THE PROBLEM Uploading a PDF to an LLM is often the most expensive way to share it. Many chat backends render every page as an image and bill you for that alongside the text, so a document whose actual content is modest can balloon in cost. Converting to Markdown first cuts that cost, keeps headings, lists, and tables intact, and — because the conversion happens on your own machine — never hands your document to a third-party conversion service. HOW IT WORKS Attach a file the way you always do: the paperclip, drag-and-drop, or paste. Decant catches it on the way out, converts it, and swaps in the Markdown version in place. You get a .md attachment instead of a 15 MB PDF. No extra steps. No separate app to open. No copy-paste dance. CAPTURE THE PAGE YOU'RE READING The same conversion runs in the other direction. Click Decant's toolbar button on any page — or press Alt+Shift+C, or right-click and pick a chat — and the page becomes Markdown and lands in the composer of the LLM chat you last used. Because it reads the page as rendered, single-page apps and logged-in views capture correctly, rather than the empty shell a re-fetch would return. Site furniture — nav bars, banners, hidden elements — is left behind. The page's own images come along as separate attachments; a checkbox sends text only instead. If the chat you're sending to accepts no file attachments, the Markdown goes to your clipboard with a paste hint. Either way the page you clicked on tells you what happened — captured, delivered, or why it failed. Reading a page takes no standing permission: the click is the consent. Decant reads the one page you invoked it on, once. No browsing history, no background access, no permission to run everywhere. SUPPORTED FORMATS • PDF (digital / text-based) • Word — .docx • Excel — .xlsx, .xls • PowerPoint — .pptx • HTML — saved web pages • Live web pages — captured from the tab you're on Charts in Office files are recovered as data tables rather than lost as pictures. WHERE IT RUNS Enabled out of the box: claude.ai, ChatGPT, Gemini, Perplexity. Add them yourself in options: Microsoft Copilot, kimi.com. claude.ai, ChatGPT, Perplexity, and Copilot get full coverage — file picker, drag-and-drop, and paste. Gemini and kimi.com convert through the file picker; their uploaders reject synthetic drops, so on those two, drag-and-drop and paste deliberately send your original file through unchanged rather than lose it, and say so on screen. Page capture works on any site you're reading. The Markdown it produces is delivered to one of the chat sites you've enabled — those are the only places Decant can put it. IT WON'T QUIETLY WRECK YOUR DOCUMENT The image layer isn't pure waste — it's what lets a model read charts, scans, and complex tables. So Decant never converts blindly: • Scanned and image-only PDFs pass through untouched. No prompt, no degradation. • Documents mixing text with real charts or photos ask you first, offering three choices: convert to Markdown, convert and attach the figures alongside it, or send the original. • A hotkey (Alt+Shift+O) forces the next upload through untouched. • Anything Decant doesn't handle uploads exactly as it would if the extension weren't installed. PRIVACY • Conversion happens in your browser. Decant operates no servers. • No analytics, no telemetry, no tracking, no data sold or shared. • Default-deny: Decant does nothing on any site until you enable that site yourself, and enabling one asks Chrome for permission to run there. • Page capture reads only the tab you invoke it on, only when you invoke it. Decant holds no permission to watch what you browse. • Settings sync through Chrome's own storage, like your bookmarks. They never reach a Decant server, because there isn't one. Full policy: https://github.com/jtrachtenberg/decant/blob/main/docs/privacy.md WHAT TO EXPECT Across six real-world PDFs — annual reports, government FAQs, financial guides — conversion cut estimated token cost by roughly a third to a half, and file size by 60–99%. Several 15 MB print-production PDFs became under 160 KB of Markdown. Each converted in under about three seconds. Token figures are estimates: actual billing depends on the model and backend, and savings shrink on figure-heavy documents where you choose to attach the visuals. The size reductions are measured. Full methodology and per-document numbers are in the project README. OPTIONAL: THE HIGHER-FIDELITY TIER For scanned documents needing OCR, or image-heavy files you'd rather not simplify, Decant can hand off to an optional companion service you run on your own machine. It talks to localhost only — nothing leaves your computer. Entirely opt-in, and the extension is fully useful without it. CONFIGURABLE • Which sites Decant runs on • What happens per file type — convert, pass through, or route to your own endpoint • The passthrough hotkey (the capture shortcut rebinds through Chrome's own keyboard-shortcuts page) • Whether captured pages bring their images along • A default answer for the convert-or-not prompt, so it stops asking IN YOUR LANGUAGE The interface is available in English, Spanish, French, German, Brazilian Portuguese, and Japanese, following your browser's language automatically. The translations are new — if a phrase reads wrong, please submit a bug report using the bug report option in the options menu. OPEN SOURCE Source, documentation, and design notes: https://github.com/jtrachtenberg/decant This is an early public release. If something converts badly, the options page has a "Report a bug" button that opens a pre-filled GitHub issue — a report naming the document that broke is the single most useful thing you can send.