โ ๐ฆ๐๐ป๐ฑ๐ฎ๐ ๐๐ผ๐ณ๐ณ๐ฒ๐ฒ & ๐๐ผ๐ฑ๐ฒ: ๐๐ถ๐ฟ๐๐ ๐๐ต๐ฒ ๐ฟ๐ผ๐ฏ๐ผ๐๐ ๐ฐ๐ผ๐๐น๐ฑ ๐๐ถ๐๐ถ๐. ๐ง๐ต๐ฒ๐ป ๐๐ผ๐บ๐ฒ๐ผ๐ป๐ฒ ๐ฎ๐ป๐๐๐ฒ๐ฟ๐ฒ๐ฑ ๐๐ต๐ฒ ๐ฑ๐ผ๐ผ๐ฟ. ๐ก๐ผ๐: ๐๐ต๐ฎ๐ ๐ถ๐ณ ๐ฒ๐๐ฒ๐ฟ๐ ๐๐ฒ๐ฏ๐๐ถ๐๐ฒ ๐ต๐ฎ๐ฑ ๐๐ต๐ฒ ๐๐ฎ๐บ๐ฒ ๐ฑ๐ผ๐ผ๐ฟ?
A few weeks ago I asked whether our websites are ready for non-human visitors. I added an MCP server at www.ai4enterprise.ca/mcp so AI agents could use the siteโs content as tools. The following week I gave the site an agent of its own at /agent, and showed a Microsoft Copilot Studio agent delegating a task to it over A2A.
Building those two front doors left me with a nagging question. My /mcp has get_services and search_insights. The next siteโs might have offerings_list and find_articles - that doesnโt scale.
So this week Claude and I took a stab at writing down how it could work: the MCP Website Profile (MCP-WP) v0.1, a proposed draft.
๐ง๐ต๐ฒ ๐ถ๐ฑ๐ฒ๐ฎ:
- A known front door: every participating site serves MCP at https://<your-site>/mcp
- Core (mandatory, public, read-only): five standard tools for site info, services, contact, search and page retrieval. No AI model needed, just a thin layer over your CMS.
- Extended (optional): modules for articles/insights, service details, people and events, plus sector modules.
- Transactional (optional, provisional): booking, registering and requesting a quote. These use a prepare โ human approves โ submit lifecycle with idempotency keys and an audit trail. Payment details never pass through the tools.
Itโs really just an attempt at standardising MCP for websites. With /mcp, the visiting agent uses your tools and keeps control of its own workflow. With /agent, it hands the whole job to your agent. MCP-WP standardises the first.
My www.ai4enterprise.ca/mcp is the reference implementation. To be honest, it doesnโt fully conform to my own spec yet, so the next step is aligning it and building a conformance checker.
The repo has the spec, JSON schemas for all tools, worked examples and a validation script: github.com/steveh250/website-mcp-standard
This is one practitionerโs attempt to put something on paper, not a finished standard. Iโd genuinely like to hear where itโs wrong. Is Core too heavy for a small site? Should discovery lean on the new MCP Server Cards rather than a fixed path? Would you trust an agent with the Transactional level?
Think. Model. Act.