As of August 2026, two small open-source projects on Hacker News are making the same argument, and it reaches every business that owns a website. Accept-Markdown pulled 175 points and 108 comments for one plain idea: when an AI agent requests your page, hand it clean Markdown instead of the HTML you built for people. WebMCP, published August 4 with 56 points and 59 comments, goes a step further — it says your site should publish named actions an agent can call, like "check availability" or "get a quote," not just text to read.
What they are actually proposing
Accept-Markdown uses a mechanism that has been in browsers for two decades: the Accept header. Your browser tells a server "I want text/html." An agent can instead say "I want text/markdown," and the server returns a stripped version — no navigation, no cookie banner, no 40 kilobytes of JavaScript. The agent parses it faster and misquotes you less often.
WebMCP is a bigger idea. The page itself declares tools through the Model Context Protocol, so an agent that lands on your site can call "book_appointment" or "check_inventory" directly instead of guessing which button to click. One project is about being read correctly. The other is about being used.
The operator's read
You now have two audiences: the people who visit your site, and the models that summarize your site to people who never visit it. When a homeowner asks ChatGPT or Perplexity for the "best roofing company near me," the model often fetches a few candidate pages before it answers. If your page is a wall of scripts with your phone number rendered inside an image, the model takes what little it can read — or leaves you out.
Here is the part owners miss: this is not a Google ranking you wait months to earn. It is a parsing problem you can fix in an afternoon. A dentist in Winter Park with hours locked inside a JavaScript scheduling widget is invisible to a model that reads the raw page. A competitor two miles away with "Open Mon–Fri 8–5, Saturday 9–1, walk-ins welcome" written in plain text gets quoted. Same services, different markup.
The plumbing here is getting fast and cheap. A related project this week — an open-source model gateway that hit 201 points — adds under 1 millisecond of overhead per request. The cost of an agent reading your site is rounding to zero. The only bottleneck left is whether your site gives a clean, factual answer when it does.
Do this week
- See what an agent sees. Load your top three pages with JavaScript disabled, or run curl against them. If your hours, service area, and starting prices are not in the raw text, that is your first fix.
- Put facts in text, not pixels. Address, phone, service area, price ranges, and booking link belong in real HTML — not baked into a graphic or a widget that only renders after a script runs.
- Add a plain summary block to each key page: who you serve, where, what it costs, how to book. Three sentences an agent can lift verbatim.
- Then consider a Markdown feed. A parallel clean copy of your core pages is a small build, and it removes all doubt about what the model reads.
We covered the Accept-header mechanics and the thread behind them in an earlier post on Markdown feeds and your website's second audience. The direction has not changed; the number of projects pushing it has grown.
Where to be careful
Reading is safe to open up. Actions are not. WebMCP-style tools mean an anonymous agent can trigger a booking, request a quote, or change a cart on your site. That needs authentication, rate limits, and an approval step before anything is committed — the same gates you would put on a new front-desk hire who cannot yet be left alone with the calendar. Make your site readable first. Wire up agent-callable actions only after the read layer works and you have someone watching the queue.
What this means if you're weighing AI marketing or an agent build
If you sell to local customers, making your site cleanly readable by models is the cheapest visibility work available right now, and you can start it this week with no vendor. If you are weighing a full agent stack, treat the readable site as step one and agent-callable actions as a later phase with real guardrails — not a same-quarter project.
Want to know what ChatGPT, Claude, and Perplexity actually say about your business today? Get your free AI Visibility Report — results in 24 hours.