AI Assistant · October 2026

What is an MCP server?

For the past few months, the word "MCP" has been coming up in every tech conversation. For wine professionals who manage a cellar, this quiet protocol is silently changing the way we query software. Here is what you need to understand.

MCP, in one sentence

MCP, for "Model Context Protocol", is an open standard released by Anthropic in November 2024 and rapidly adopted by OpenAI, Mistral, Google and Perplexity. Its role: to let a conversational AI assistant (Claude, ChatGPT, Mistral, Perplexity) directly connect to a business software and read or write data there, without a hacked-together integration, without copy-paste, without a bespoke API to build.

Concretely: you paste a URL into your AI assistant’s settings, you authenticate once, and from that moment, your assistant understands your cellar as well as you do. It sees your bottles, your inflows, your outflows, your prices, your valuation. It can read. It can write. It can analyse.

Why it is a turning point

Before MCP, an AI assistant was an isolated box: it knew a lot about the world but nothing about your business. For it to work with your data, you had to either copy-paste manually, or build a specific integration via API, a project that often costs several thousand euros for every software you want to link.

MCP standardises this connection. A software exposes it once, and every AI assistant on the market can connect to it. No SDK to install. No developer access to configure. No API keys to manage in a text file. Just a URL and your usual credentials.

For a restaurant or a wine shop, the equation changes radically: your AI assistant is no longer a wine encyclopaedia, it is your colleague who knows your stock by heart. It does not tell you "here are the great Burgundies of 2020", it tells you "here are the 14 bottles of 2020 Burgundy still in your cellar, three of which reach their peak this quarter".

What your assistant can do with your cellar

As soon as Caves Explorer is connected to your AI assistant via MCP, you have about twenty ways to query your cellar and about fifteen ways to record something in it, all in natural language. A few of the most frequent examples:

  • Which bottles reach their peak this quarter?
  • Which reds under €30 do I still have in stock for tonight?
  • Give me the BevCost of the month by wine region.
  • Compare the rotation of my white Burgundies over the last six months.
  • Log the delivery: 24 bottles of 2019 Côte-Rôtie at €42 each.
  • Put this Pouilly-Fuissé on the list at €48.
  • What should I order this month to keep my pinot noir catalogue stocked?

Your assistant reads the answers in Caves Explorer and returns them in the conversation. If you dictate an inflow, an outflow or a price, it writes it directly in the software. You no longer open the interface: you talk to your cellar.

What MCP does not do

MCP is not an AI. It is not a new assistant to download, it is not an additional subscription. It is simply the pipe that connects the assistant you already use to the software you already use. The intelligence stays on the Claude, ChatGPT, Mistral or Perplexity side: you choose.

MCP is not an open door onto your data either. Caves Explorer explicitly exposes what your assistant can read and what it can modify. Only three sensitive actions allow deletion, and each requires explicit confirmation. Your assistant does not "guess" anything: it asks, you confirm.

Who can use it?

Any restaurant professional, wine merchant, hotelier, consulting sommelier or wine estate manager who already has a Caves Explorer account and a subscription to a mainstream AI assistant. The plans that expose MCP include Solo, Domaine and Signature.

On the assistant side, compatible versions are: Claude Desktop and Claude.ai (Anthropic, natively), ChatGPT Plus and Enterprise (OpenAI, via connectors), Mistral Le Chat Pro (Mistral AI), and Perplexity Pro. The list grows every quarter, as the standard is supported by all major AI models on the market.

How long does it take to connect?

Five minutes, once. You open your AI assistant’s settings, you add a new connector, you paste the address of your Caves Explorer MCP server, you authenticate via your browser with your usual Caves Explorer credentials. That is it. Your cellar is connected.

No local installation, no software to download, no update to plan. MCP server upgrades are rolled out on the Caves Explorer side: your assistant benefits from new features as soon as they are released, without any action from you.

Security and confidentiality

MCP relies on OAuth 2.1 for authentication. In other words: the same standards as your online bank. You never give your password to the AI assistant; you grant it a revocable access authorisation, which you can withdraw at any time from your Caves Explorer area or from the assistant’s settings.

The data queried by the assistant travels encrypted (HTTPS) and is not stored on the AI publisher’s side beyond your session. Caves Explorer, for its part, logs your assistant’s requests like any other action on your account: you keep a full trace of what was read, written or modified.

Why Caves Explorer adopted MCP

Our customers have been telling us the same thing for years: "I want to be able to ask my software what I have left, not to search in three menus". Before MCP, we would have had to develop our own conversational interface, train our own AI, maintain models. A technical and financial pit, for a result worse than the market’s best assistants.

MCP lets us do what we do best (manage your cellar data accurately) and lets Claude, ChatGPT, Mistral or Perplexity do what they do best (converse in natural language). Everyone sticks to their trade. The result: the best augmented sommelier experience on the market, without dependence on a single AI supplier.

Going further

Convinced? Head to the "Your AI assistant (MCP server)" page for the full setup procedure: the five connection steps, the exhaustive list of 20 reads and 16 writes available, the technical guarantees, the FAQ. If you prefer to see before acting, our teams offer a video demonstration on request.