How can AI agents query The Website Specification?

AI agents can query The Website Specification through three separate channels: an open, read-only MCP server at https://mcp.specification.website/mcp (HTTP transport, no authentication), a published Agent Skill that teaches compatible agents when and how to use the spec, and per-page Markdown served via /llms.txt or an Accept: text/markdown request header on any spec URL. All three are documented on the site itself, so the fastest path is to pick the channel your agent already supports rather than building a custom integration.

The three access channels

The MCP server is described as "open" and "read-only, no auth" — the site does not state rate limits, quotas, or uptime guarantees, so treat those as unknown rather than assuming they are unlimited.

Connecting an agent over MCP

The site publishes the exact client configuration, so no discovery step is needed:

{
  "mcpServers": {
    "specification-website": {
      "transport": "http",
      "url": "https://mcp.specification.website/mcp"
    }
  }
}

Input: the JSON block above, placed in your MCP client's server configuration. Action: the client opens an HTTP transport connection to that URL. Expected result: the agent can query the spec's topics through MCP tools without supplying credentials.

If your client expects a different config key or a command-based transport, the site's "Connect MCP" link is the authoritative source — the snippet above is the only form given in the page evidence.

Using the Agent Skill

The site also publishes an Agent Skill, described as teaching "any compatible agent when and how to use it." This matters because an MCP connection alone tells the agent how to fetch spec content but not when it is relevant. The skill supplies that judgement layer.

Practical implication: if your agent framework supports skills, load the Agent Skill alongside the MCP connection rather than relying on the MCP server alone. If your framework has no skill mechanism, the MCP server and the Markdown endpoints still work — you just lose the routing guidance.

Fetching spec pages as Markdown

For agents that do not speak MCP, every spec URL can return Markdown instead of HTML:

• /llms.txt — a published index of the spec in the llms.txt convention.
• Accept: text/markdown — send this header on any spec URL to get that page as Markdown.

Input: an HTTP GET to a spec URL, with Accept: text/markdown set. Action: the server negotiates content type and returns Markdown. Expected result: clean Markdown for that single topic, suitable for chunking into a retrieval index or feeding directly into a context window.

This is the lowest-friction option: no client library, no protocol handshake, just an HTTP request with a header.

What the agent can actually retrieve

The spec covers 173 topics across ten categories, all mapped to external standards (WHATWG, W3C, IETF RFCs, WCAG, MDN). The categories are:

• Foundations — 20 topics (HTML, head, document basics)
• SEO — 14 topics (robots.txt, sitemaps, canonicals, structured data)
• Accessibility — 30 topics (WCAG-aligned rules)
• Security — 20 topics (headers, transport, policies)
• Well-Known URIs — 13 topics (standard paths under /.well-known/)
• Agent Readiness — 22 topics (legibility to AI agents and crawlers)
• Performance — 26 topics (Core Web Vitals, caching, images, fonts, network)
• Privacy — 7 topics (consent, signals, visitor choice)
• Resilience — 8 topics (error pages, offline, redirects)
• Internationalisation — 13 topics (language, localisation)

Each topic is structured as a yes/no checklist item, which makes the output easy for an agent to act on: the agent can retrieve a topic, evaluate the target site against it, and report a pass/fail plus the linked standard.

Choosing between the channels

• Agent already supports MCP → connect the MCP server; add the Agent Skill if your framework loads skills.
• Agent is a plain HTTP client or a RAG pipeline → use /llms.txt to enumerate pages, then Accept: text/markdown per page.
• You want the agent to audit a site → the spec's own "How to use this site" flow is audit → learn → improve, with each item as a yes/no check. Point the agent at the relevant category rather than the whole 173-topic set to keep context manageable.

Common sticking points

• Assuming the MCP server is the only route. It is not — the Markdown endpoints exist precisely so agents without MCP support can still read the spec.
• Skipping the Agent Skill. Without it, a compatible agent may have access to the spec but no trigger telling it to consult the spec during a task.
• Treating the spec as opinion. Every topic links back to a source standard, so when an agent reports a failure, the citation is the thing to check — not the spec page alone.
• Expecting pricing or account details. The page evidence shows no pricing signals and no payment platforms; the MCP server is stated as no-auth, but nothing on the page addresses commercial terms or limits.