Why it matters
- When an AI agent reaches a site, it looks at fixed addresses (
/.well-known/…) for what it can do there: an MCP server to connect to, an A2A agent to talk to, skills to install, or described APIs. - A site that publishes none of them stays, for the agent, pages to read and nothing more.
- This is information: most sites offer agents none of this, and need none.
How to fix
Publish what fits what you offer, one at least:
json
{
"name": "example-store",
"version": "1.0.0",
"description": "Product search and order tracking",
"remotes": [{ "type": "streamable-http", "url": "https://example.com/mcp" }]
}- An MCP server card at
/.well-known/mcp/server-card.json(the example above). - An A2A agent card at
/.well-known/agent-card.json. - A skills index at
/.well-known/agent-skills/index.json. - An API catalog at
/.well-known/api-catalog(RFC 9727).
How we detect
In the agent-readiness tool alone: we ask for each address as ArablyzerBot with Accept: application/json, where robots.txt allows, and with it an address no site has, to know a site that answers every address. A file counts as published when it is JSON with the fields its specification requires. When robots.txt keeps us from every address, or none answered, we say we could not check.