One Accept header serves agents Markdown and cuts tokens ~80%
Content negotiation for agents is real: Claude Code, Cursor and Copilot already send the header. Here's how to serve it — and a plugin trap to avoid.
Copy markdown
Your coding agents already ask for Markdown
The live acceptmarkdown.com tracker shows Claude Code, Cursor, GitHub Copilot CLI/Chat, Microsoft Copilot, OpenCode and OpenClaw all send `Accept: text/markdown` today; OpenAI's Codex CLI finds it via <link> tags. Serve that header and the tools you already use read your pages clean instead of scraping HTML.
The 80% token cut, measured
A page that costs 16,180 tokens as HTML drops to 3,150 as Markdown — about 80% fewer. WordPress.org logged an 84% byte cut and roughly 6x fewer tokens. Practically, an agent that could read 5 HTML pages of your docs can read ~25 in Markdown.
Ship it from the same URL, no new routes
It's plain RFC 7231 content negotiation: detect `Accept: text/markdown` on your canonical URL, return Markdown, and set `Vary: Accept` so caches don't cross the wires. No new protocol, no .md mirror. Cloudflare added a one-toggle "Markdown for Agents" that converts HTML to Markdown at the edge — free in Beta on Pro, Business and Enterprise.
The plugin trap: a .md URL is not content negotiation
Watch for SEO plugins that claim "Markdown for AI" but only mint .md variants and never read the header — AIOSEO still returns text/html on the canonical URL when an agent asks for Markdown. Roots' post-content-to-markdown and WordPress.org do it right, honoring the header on the same URL.
Elsewhere: GLM 5.2 free on Vercel ends today
Z.ai's GLM 5.2 (1M-token context) has been free for eve-framework agents via Blackbox on Vercel's AI Gateway — the promo closes today, Aug 27. Switch before it reverts to paid with `eve set --model zai/glm-5.2` (Fast mode and the -fast variant are excluded).