I Made My Site Usable by AI Agents. Then I Counted What It Cost.
Motion Garden now has an index for agents, a Markdown copy of every page and an assistant that can work the page. The first version ran three server functions on every page view. Here's how I got that to zero.
Motion Garden is my catalog of 154 animated React components. People open a page, tune a component with live controls and copy the source. But more and more, the reader is not a person. It's a coding agent that someone told to "add that wave background from Motion Garden".
An agent is a weak reader. It fetches one address and reads what comes back. It runs no script, it can't guess a URL, and a long answer gets cut short. My pages were built for eyes and a mouse.
So this week I made the site usable by agents, with Claude Code doing much of the typing. It took four steps. Then I counted what each one cost me to serve, and the first count was bad.
Four steps, in order
| Step | The site is | What ships | What runs on my server |
| 1 | Findable | An index at /llms.txt, a robots.txt that lets agents in | Nothing |
| 2 | Readable | A Markdown copy of every page, the facts as JSON | Nothing, apart from search |
| 3 | Operable | An assistant that can work the page | Nothing. It runs in the visitor's browser |
| 4 | Generous | Two free questions a day | One guarded endpoint |
The order matters. An assistant with nothing true to read makes things up. So reading comes first.
Readable first
Step one is an index. /llms.txt is a plain text file, in the shape from llmstxt.org. It has one line per component and the exact shape of every data address. Mine has 157 links and 36,282 characters. It stays an index. The content lives in each page's own document.
Step two is that document. Every component page has a twin at the same address with .md added:
curl https://motion-garden.netlify.app/llms.txt
curl https://motion-garden.netlify.app/components/waves.mdThe twin for Waves is 19,482 characters. It holds every source file, the install command, the props with their defaults and ranges, a usage example and the integration rules. That's what an agent needs to finish the job without opening the page. It's not the page turned into Markdown. It's the same text the page's "Copy for Agent" button gives, and one function builds both.
The same facts are also there as JSON, open to any origin. All of it comes from the one registry entry that the page itself is built from. I never write a second copy by hand, because a second copy goes out of date on the first new component.
Then operable
Step three is an assistant called Ask the Garden. I didn't invent the hard part. It's built on the open agentic web from Cognizant AI Labs, which gives you a browser runtime and a spec for how agent sites talk to each other.
My part was the site's side: two agents and eight actions. The actions are the site's own functions, like searching the catalog, reading a component's code, opening a component and setting a control. An action calls the same function the page's own control calls. A bad value is refused with data, {ok: false, reason, field}, and nothing changes.
The assistant runs in the visitor's browser, on the AI key they bring. Their messages go straight to their provider and never touch my server.
Two free questions a day
Most visitors don't have a key ready. So step four gives everyone two free questions a day, and I pay. That puts a key of mine behind a public address, and that's the part that needs care.
So the endpoint is narrow. It answers only my own pages. It accepts only the exact instructions of my two agents, and only my own tools. The server picks the model, and it picks free models only, so the key can't be billed. It counts two questions per visitor and 300 across the site each day, and no model call runs before a question is counted.
It doesn't stop everything. A visitor can still type anything in the message, and a model may follow it. The limits bound how much anyone gets.
Then I counted
Motion Garden runs on Netlify, where function calls are the metered cost. A static file is free to serve. A function is not.
When the assistant first worked, I opened one component page in a real browser and listed every request that reached my server. Each page view ran three functions:
- The page itself.
- A view count.
- A 404 for a file the assistant's runtime asked for by convention.
The first one surprised me most. The build said the page was prebuilt. But two file reads in a server component left a part that a function had to fill on every request, so the host could not keep the page. Next.js gave no warning. The fix was to move each read into a cached function:
async function storedBrief(slug: string) {
"use cache"
cacheLife({ revalidate: Infinity, expire: Infinity })
return buildBrief(slug)
}The view count now goes out once per browser, not once per view. The missing file became a real file.
Then the same problem showed up on the agent side. I had cached the agents' documents with cacheLife("max"), Next's 30 day life, which seemed generous. After each deploy, all 616 of them came out already stale, and the host ran two function calls for each one before it kept the answer. A life that never revalidates fixed it. Only a deploy can change those files anyway.
| What happens | Before | After |
| A page view | 3 function calls | 0, or 1 the first time a browser sees a component |
| An agent reads a document | A function call every time | A file |
| A document after a deploy | 2 function calls before it's kept | Kept after 1 request |
Opening the chat costs one call, to read how many free questions are left. A question costs the model calls and nothing else.
Count with a browser, not by reading code.
I would not have found any of this by reading the code. The code looked right. The build output looked right. Only the list of requests was wrong. Now a check runs at the end of every build. It stops the deploy if a page needs a function again, or if a prebuilt file would be rebuilt by time.
What it still can't do
- An address that doesn't exist still runs the 404 page, which is a function. I fixed the one file the runtime asks for, not the general case.
- 31 of the 154 briefs are over 30,000 characters. The assistant doesn't send those through the chat. It gives the install command, the file list and the Markdown link.
- People behind one address, like an office, share one count of free questions.
- The assistant only changes the page: which component is open, and its control values. It can't create a component or edit your project, and it says so.
The method is a skill now
I packed the method into an open skill, agentify-any-website, under MIT:
npx skills add kleenpulse/agentify-any-websiteIt has two scripts. One audits a live address and reports which agent reads reach your server. The other opens a page in headless Chrome and counts the server calls per view. The notes inside are the traps above, with the numbers. For the assistant step it points to Cognizant's own agentify-website skill and doesn't copy it.
Quick answers
Do I need an AI model for this?
Not for steps one and two. An index and a Markdown copy of each page are plain files, and most of the value is there.
Does it need Next.js?
No. The rule is the same on any stack: write the files during the build. Next.js 16 only made it harder to see when I hadn't.
What does it cost to serve?
For reading, nothing. Every document is a file. For the assistant, one call when the chat opens, and the model calls for a free question.
Motion Garden is live at motion-garden.netlify.app, and the code is on GitHub. The numbers in this post are from runs on 21 November 2026.
Open to full-time roles