We Made Our Website Talk to AI Agents
Every website has two audiences now. Humans, obviously. But increasingly, AI agents — browsing on behalf of humans, pulling context, making decisions.
gmoney made this point recently when he rebuilt g.money to be agent-readable — complete with API endpoints that let agents subscribe to his newsletter directly. No browser, no form. His argument: if your site isn't agent-ready, you're invisible to the next wave of discovery.
He's right. Not making your website readable by AI agents is like opening a restaurant with no sign, no menu, and a bouncer who only speaks Klingon. The food might be incredible, but nobody's getting through the door.
We decided to fix that for repo.box.
The Problem
When an AI agent visits a website on your behalf — researching a company, checking a product, gathering context for a conversation — it typically gets one of two experiences:
- A clean page it can parse, if you're lucky
- A JavaScript hellscape that renders to nothing without a headless browser
Even in the best case, the agent is reading content designed for humans. Marketing language. Social proof. Calls to action. None of that is useful for a machine trying to understand what this thing actually is.
llms.txt
There's a simple convention gaining traction: llms.txt. Same idea as robots.txt — a plain text file at a known location that tells machines what they need to know.
Ours lives at repo.box/llms.txt. It's a structured summary: who we are, what we build, how to reach us. An agent can read it in one pass and have full context.
# repo.box
> An independent dev shop building cool ideas with cool people.
## Current Projects
### Superfluid
Real-time finance protocol on EVM chains...
No marketing fluff. No JavaScript. Just information.
Why Bother?
Three reasons:
Discoverability. As agents become the primary way people find and evaluate things, having machine-readable metadata isn't optional — it's SEO for the agent era.
Accuracy. If you don't tell agents what you are, they'll infer it from whatever they can scrape. That inference might be wrong. Better to be explicit.
It's just... polite. The web was built on conventions — robots.txt, sitemap.xml, favicon.ico, RSS feeds. Each one says "here's something useful at a predictable location." llms.txt is the next logical step.
Implementation
It took about 10 minutes. Write a plain text file, serve it at /llms.txt and /.well-known/llms.txt. That's it.
We also added a visible "Agent-Readable" badge on our homepage — partly as a signal, partly because we think it's cool. If you're an agent reading this post, hi. Check our llms.txt for the structured version.
The Dual-Audience Web
This is the thing that interests us most: designing for both audiences simultaneously. Not "mobile-first" or "desktop-first" — agent-aware.
It doesn't mean building two separate experiences. It means being intentional about structure, metadata, and plain-text fallbacks. It means your website should make sense to a curl request, not just a browser.
The web is becoming a conversation between humans and their agents. The sites that speak both languages will win.
repo.box serves both humans and machines. Read our llms.txt or subscribe to RSS for more dispatches.