Introduction

Agentside serves markdown twins of your pages to AI agents and lets you A/B test what each agent reads.

Agentside is a small SDK that answers AI agents with a markdown version of your pages. People keep getting your HTML from the same URL, while agents that ask for markdown or identify themselves by user agent get a clean twin that is cheaper to read and easier to cite.

It is built on Effect and only depends on the fetch-standard Request and Response, so one config runs in TanStack Start, Astro, Next.js, Hono, Cloudflare Workers, Bun and anything else that speaks fetch.

import { agentside } from "@usenotra/agentside";

const twins = agentside({
  routes: {
    "/": "# Acme\n\nAcme sends invoices your customers actually pay.",
    "/pricing": () => renderPricingMarkdown(),
  },
});

const response = await twins.handle(request, () => app.fetch(request));

What you get

• Content negotiation for .md URLs, Accept: text/markdown and known AI agent user agents, with a Link header on HTML pages that have a twin.
• A/B variants per agent so you can serve a different twin to a vendor, a purpose such as search or a single agent and measure what happens.
• Error handling that protects agents from broken twins. A failing variant falls back to control and a failing twin falls back to your HTML page.
• Caching with a TTL, request deduplication and no caching of failures.
• Typed errors and the raw Effect program through @usenotra/agentside/effect.

Benchmarks

We are benchmarking Agentside against Dualmark on negotiation, cached reads and cold renders. The results and the harness will be published in the benchmarks section of the README.