MD-This-Page

A Chrome/Firefox extension (adem-kouki, MIT, 1.3k★/127 forks) that turns the page you are looking at into clean Markdown and hands it to a model. Alt+M or right-click, preview tab, then copy, download, or copy-as-prompt. Toggles for images, links, metadata, source URL and page maps; settings sync across browser profiles.

The stack is four off-the-shelf parts: Plasmo (extension framework), React + Tailwind for the UI, Mozilla’s Readability to find the article inside the page, and Turndown for HTML → Markdown. No model, no API key, no server.

Tier: T2 — first-party README, primary on itself; nothing measured. The token-efficiency claim is asserted, not benchmarked.

The fifth lever on the token-budget corner

The spoke’s context-cost taxonomy had four moves, all operating on context already in flight inside a harness: compress (headroom), remove (safe-prompt-pruning), swap modality (pxpipe), route to a cheaper model (tokenmining). MD-This-Page adds a fifth that sits upstream of all of them — strip at capture. Ads, nav, cookie banners and script tags never become context, because Readability drops them before the text reaches a clipboard.

It’s the cheapest possible version of the move, which is what makes it worth a page: no proxy, no MCP server, no LLM in the loop deciding what to cut. Readability is a heuristic Mozilla shipped for Reader Mode in 2010, and it is doing the same job safe-prompt-pruning does with three deterministic passes. The lesson the expensive tools keep re-learning is that a lot of context waste is structural and can be dropped by rule.

Where it sits against today’s other two

Same day, same principle, three different points on it: mobilerun reads a phone’s accessibility tree instead of its pixels, video-use reads a transcript instead of frames, and this reads an article instead of a rendered page. All three take a representation the platform already maintains — Readability is exploiting the fact that HTML carries semantic structure the visual rendering throws away — and refuse the raw surface. See text-first-agent-design.

The difference is who drives. Mobilerun and video-use are agents choosing their own input; MD-This-Page is a human doing the curation by hand, one page at a time, and pasting the result into a chat. That makes it the manual, pre-agentic form of the same discipline, and a reminder that context-rot‘s remedy doesn’t require any machinery — someone deciding what not to paste is the whole intervention.

What it isn’t

Not an agent tool. Nothing autonomous consumes its output; there’s no MCP surface, no CLI, no way for a harness to invoke it. It’s a human’s clipper that happens to be aimed at a prompt box, which is also its adjacency to ../research-wiki (the markdown-PKM / web-clipper lineage — the logged runner-up for this ingest). Routed here because the README’s stated purpose is LLM context efficiency, and that corner lives in this spoke.

Related

text-first-agent-design · context-rot · headroom · safe-prompt-pruning · pxpipe · tokenmining · mobilerun · video-use · adem-kouki