Sites in the Woods

Can an agent find its way?

More and more visitors are AI agents acting for someone. Many of our sites are built for them, so the monitor checks the files agents look for, and on some sites it sends an agent in with nothing but the address and a goal.

What agents look for

/llms.txt
A short Markdown guide to the site for language models. We check that it's served as text or Markdown (not an HTML page from an app's catch-all route), starts with a # title, and says more than a line.
/.well-known/agent-card.json
An agent card: a JSON description of an agent service, its name, what it does and where to reach it. We check that it's a JSON object with a name and a URL on the same site.
An OpenAPI description
The API in a form an agent can read without guessing: every route, its parameters and its answers. We check that it has an OpenAPI version and at least one route.
An MCP server
Tools an agent can call directly. We check that it answers the handshake and lists its tools, and we notice when that list changes, because a renamed tool breaks every agent using it.

Files aren't the whole story

A site can have every file and still lose an agent halfway. So on the sites made for agents, an AI agent gets only the domain and a goal ("register", "adopt a creature") and has to get there through the public site, the way a new visitor would. A fresh agent tries weekly, and a returning one every few hours. It passes when it reaches the goal; when it gets stuck, we keep the transcript showing where, and whether the page it was stuck on offered any way forward. Every day we also look for dead ends on the pages we already read: a not-found page with no way on, and links with no name an agent could go by.

See it on your own site

One look at your site reads a domain once: robots.txt for search and AI crawlers, the homepage's description, share card and structured data, and the files above, and keeps the report at the site's own address (/check/<domain>), our findings only, so changes can be seen. A site can keep it out with User-agent: site-monitor and Disallow: / in its robots.txt (the owner-started API recovery scan is named site-monitor-scan, so it needs its own group).

A standard to check against

Our open-source agent readiness audit scores a site against the Agent and Search Readiness Standard, an open checklist for sites that want to work for agents and search engines. The usability test above is its test T5.

Watch it, not just score it

The audit tells you once; a monitor tells you when it breaks. agent-ready-uptime, open source and free to run yourself, watches the files above, whether AI crawlers can still get in, and an MCP server's tools, and alerts Discord or Slack when one changes: the same checks this service runs, from one place.

What this site offers agents

Our own API, MCP server, skills and agent files, with copy-paste setups and a quickstart, are on the developers page.

Last updated .