acceptmarkdown.com makes the case for serving Markdown to AI agents via Accept headers

What the site proposes

A project called acceptmarkdown.com, which reached the front page of Hacker News, is promoting a simple idea: use ordinary HTTP content negotiation to hand AI agents a Markdown version of a page. When a client requests a URL with Accept: text/markdown, the origin responds with the article as Markdown instead of the full HTML document, letting agent clients read the content directly rather than working around navigation, scripts and layout markup.

According to the site, publishers already have the content — a Markdown variant simply exposes it, keeping a single canonical URL that serves both humans and machines.

The pitch: tokens, retrieval quality, speed

acceptmarkdown.com lays out three arguments. The first is context efficiency: a Markdown representation strips out navigation, styles, scripts and structural wrappers, so an agent's limited context window is spent on the writing itself rather than on markup. The second is retrieval quality: with no advertising, recommendation sidebars or overlay dialogs in the payload, the text a RAG pipeline embeds is cleaner, raising the signal-to-noise ratio. The third is speed: a much smaller response means less to download and process before the model starts generating, shortening the wait for the first output token.

Tooling for checking and deploying

The site ships a checker titled "How AI-ready is your URL?". A user enters a URL, the service issues the request from its own edge, and it reports back how the origin responds. The test covers four criteria: whether Markdown is served for Accept: text/markdown, whether the response sets Vary: Accept, whether unsupported media types are rejected with HTTP 406, and whether q-value preferences are honored.

For implementation, there is a quick start describing a minimal setup — a single server that produces two representations — plus guides on the correctness details: Vary, q-values, 406 responses, caching, and a zero-configuration shortcut on Cloudflare. A recipes section provides copy-paste configurations for much of the mainstream web stack, including Nginx, Apache, Caddy, Cloudflare Workers, Next.js, Astro, SvelteKit, Nuxt/Nitro, Express, Go, Django, Rails, Laravel, WordPress and Discourse.

It depends on the clients, too

The pattern only pays off if agents actually ask for Markdown. To that end, the site maintains an AI agent support matrix tracking which AI agents include Accept: text/markdown in requests made by their browse or fetch tools — an acknowledgment that client adoption, not server configuration, is the open variable. A reference section grounds the approach in existing standards, citing RFC 9110 for HTTP semantics and RFC 7763, which registered the text/markdown media type, alongside related specifications.

Why it matters

Content negotiation is one of the oldest features in HTTP, and this proposal reuses it instead of inventing a new format, endpoint or sidecar file. That keeps the adoption cost low: the same URL serves browsers and agents alike, and existing caches, CDNs and frameworks can be configured to support it. The correctness requirements are real — omitting Vary: Accept can leave caches serving the wrong representation to the wrong client, and strict 406 handling can break clients that send unusual Accept headers — which is precisely what the site's guides address. If automated traffic becomes a meaningful share of visits, sites that negotiate cleanly get cheaper agent queries and more accurate answers about their own content; if it never does, the change costs very little. As a combination of checklist, verifier and a large pile of server recipes, acceptmarkdown.com lowers the barrier for anyone who wants to try the pattern on their own infrastructure.