Most of the web treats AI agents as intruders: blocked, cloaked, or tolerated as scrapers. We think the next decade of visitors will be people and their agents, working together. So we built this site for both, as equals, and we are writing down the pattern so any site can adopt it.
Starwell is, as far as we know, the first site designed end to end on this principle. Every choice below is live on this domain today; none of it is a roadmap.
A human page and its data are the same resource in two costumes. Every public data page here carries a rel="alternate" link to its exact API equivalent: the catalog twins to /v1/catalog, a source page to its datasets endpoint, a series page to its observations. The { } control in the header of those pages takes you to the twin in one click, so people can see precisely what agents see.
An agent should never have to parse our HTML. Request any data page with Accept: application/json (or ?format=json) and it redirects straight to the structured answer. The prose is for people; the data was never trapped inside it.
/llms.txt is the site in one screen: what this is, where the docs are, where the API lives. /llms-full.txt is the complete documentation as a single markdown file, generated from the same source as the human docs so the two can never disagree.
Every series page publishes schema.org Dataset markup with its license, temporal coverage, and CSV and JSON distributions; the catalog publishes DataCatalog. Machines get commitments, not vibes.
The store speaks Model Context Protocol natively. Point any MCP client at https://starwell.dev/api/starwell/mcp and verified statistics become tools in the agent's hands: search, observations, computed answers, deep analysis, monitors. Same data, same verification, same citations as this website.
Agents ride the same rate limits and the same free keys as people. No cloaking, no agent-only paywalls, no serving machines different numbers than humans. If a value is cited on the page, the same citation rides the API envelope.
The machine surfaces expose exactly what the human pages expose: the world's public record. Accounts, keys, and usage stay behind authentication and out of every machine manifest. Being agent-friendly is not an excuse to leak.
{ } on the catalog or any series page: the machine twin, one click away.curl -H "Accept: application/json" https://starwell.dev/catalog: negotiation in action.The pattern is deliberately small enough to adopt in a weekend, and none of it requires a framework:
rel="alternate" links from every content page to its structured equivalent.Accept: application/json on those pages with a redirect to the data./llms.txt, and /llms-full.txt if you have real documentation.If you adopt the pattern, or think it should say something it does not, write to hello@starwell.dev. We would like the agent-native web to be much bigger than one statistics company in Toronto.