Everyone is talking about MCP. Almost nobody shows what actually happens.

The Model Context Protocol (MCP) is the plumbing that lets an AI assistant reach beyond its training data and call a real tool: your compliance library, your CRM, your internal knowledge. It is one of the most important building blocks in applied AI right now, and also one of the most hand-waved. People say "the AI calls a tool" and leave it there.

So let us pull the cover off and trace a single call from end to end - the user's question, the exact JSON the assistant sends, the request that hits the server, the response that comes back, and the answer the user finally reads. To keep it concrete and a little fun, we will use one of our own MCP servers: a free, educational handbag and fashion knowledge server we are building under our Visionaire brand. It is in preview while our subject-matter expert signs off the content, but the mechanics below are exactly how it works.

The cast

• The user - a person typing into an AI chat client.
• The MCP client - the AI assistant (Claude, ChatGPT and others) that holds the conversation and decides when to call a tool.
• The MCP server - our service. It advertises a couple of tools and answers calls over plain HTTP.
• The knowledge bundle - a packaged, cited corpus the server reads from, so every answer traces back to a real source.

Step 1 - The user asks

The user is standing in a shop and types:

"I'm about to buy a handbag - how can I tell if it's actually good quality?"

Step 2 - The assistant decides to use a tool

When the client connected, it asked the server for its tool list, so the assistant already knows a tool called query_visionaire_knowledge exists and what arguments it takes. It recognises this question is a fit, and emits a tool call. On the wire that is a JSON-RPC message - note it lets the server auto-pick the knowledge area and asks for the default cited answer:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "query_visionaire_knowledge",
    "arguments": {
      "query": "I'm about to buy a handbag - how can I tell if it's good quality?",
      "requested_output": "cited_answer"
    }
  }
}
                        

Step 3 - The call hits the server

That message is sent as a single HTTPS POST. Access is gated by a bearer token, and the body is size-capped - small, boring, deliberate security:

POST /mcp HTTP/1.1
Host: visionaire-mcp.appgenie.com.au
Authorization: Bearer <token>
Content-Type: application/json
Accept: application/json, text/event-stream
                        

Step 4 - What the server does

The server does not call out to a large language model. It is deterministic. In a few milliseconds it:

• Guards against abuse - it refuses "dump the whole knowledge base" style requests, so the corpus cannot be scraped wholesale.
• Routes the question - keywords in the query select the right knowledge area (here, handbag quality) from the available profiles.
• Selects sources - it scores the documents in the bundle and picks the best-matching ones.
• Builds a cited answer - it returns the relevant expert excerpts, each tagged with the document it came from.
• Attaches the guard-rails - an educational disclaimer and a "Powered by" attribution line are added to every single response, structurally. The model cannot strip them.

Step 5 - The response comes back

The server returns a structured result: a human-readable text block plus a machine-readable structuredContent object the assistant can reason over. Trimmed for clarity:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "structuredContent": {
      "status": "ok",
      "selected_profile": "visionaire-handbag-quality",
      "answer": {
        "mode": "cited_answer",
        "guidance": "Use the excerpts as the expert framework; keep the educational framing, never turn it into a purchase directive.",
        "excerpts": [
          {
            "document": "Quality at a Glance (KB-VIS-HQ-002)",
            "heading": "The in-store quality scan",
            "content": "Look at the material, the hardware and findings, the stitching, and the lining. Price alone does not decide quality."
          }
        ]
      },
      "sources": [
        { "documentId": "KB-VIS-HQ-002", "title": "Quality at a Glance", "citation": "Quality at a Glance (KB-VIS-HQ-002)" }
      ],
      "educational_disclaimer": "General education, not advice for a specific purchase. Physical inspection and your own judgement remain essential.",
      "attribution": "Powered by Visionaire - Fashion Product Development Experts | visionaire.com.au"
    }
  }
}
                        

Step 6 - The assistant writes the answer

The assistant now has an expert, cited framework instead of a guess. It composes a natural reply, keeps the educational framing, and surfaces the disclaimer and attribution exactly as supplied. The user sees something like:

• Material - leather or fabric is fine either way; check the quality of whatever it is, and the inner tag (or the maker's site) if you are unsure.
• Hardware - look at the seams and joins, and whether the stitching is neat and in place.
• Lining - a thin lining versus a heavier, considered one tells you a lot.
• Price is not quality - a good bag does not have to be expensive.

General education, not advice for a specific purchase - your own inspection still matters.
Powered by Visionaire - Fashion Product Development Experts | visionaire.com.au

Why this pattern matters

Look at what the user never had to see: the routing, the source scoring, the guard-rails, the citations. They asked a question and got a grounded, attributed answer. That is the whole point of a well-built MCP server - it makes expertise callable, with the controls baked in rather than bolted on.

And here is the part we love: the engine is the same one behind our AppGenie Compliance MCP. Same retrieval, same citation discipline, same fail-closed gating - we just swap the corpus. ISO and NIST controls for enterprises; handbag and fashion know-how for everyone else. Cited, governed, and honest about what it does and does not do.

The Visionaire fashion MCP is still in preview while the expert content is reviewed and signed off, and it is educational only - never product or purchase advice. When it opens up, you will be able to point your own AI assistant at it. In the meantime, if you have deep expertise trapped in people's heads and want it turned into a callable, governed AI tool, that is exactly what we build.

Want one for your domain?

Drop a note to our team and we will show you what an MCP server built on your expertise looks like.

• Back to Blog