
MCP servers for businesses: should you connect your software to AI?

Google is opening Gemini and its search to third-party apps through MCP. What an MCP server is, what it is for in an SME and the precautions to take before getting started.
An MCP server is a standardised entry point that allows an artificial intelligence assistant, such as Gemini or ChatGPT, to look up data in your software or trigger an action in it, with the user's consent. For an SME, building one makes sense when your customers or your teams often query your tool and you know precisely what the AI is allowed to read or change. In other cases, a well-designed application and a clean API remain the priority.
The question is becoming concrete. On 28 October, Google opened a preview of its Connected Apps platform, which lets publishers plug their service into the Gemini app and into AI Mode in Google Search. The connection goes through an MCP server hosted in the cloud, or through Android's AppFunctions. Google already claims more than 50 connected apps, including Adobe, Experian, Localiza and Peloton, and the announcement circulated widely on X last week.
What is an MCP server, in practice?
MCP stands for Model Context Protocol. It is an open standard published by Anthropic in November 2024 to connect AI assistants to the systems where data lives, whether a customer database or a management tool. Before it, each AI publisher required its own connector. With MCP, you write the integration once and several assistants can use it.
The MCP server describes a list of tools the assistant can call. Each tool has a name and a description, for example "search for an order" or "create a quote request". The assistant reads this list and then chooses the tool that matches the request. It is your server that carries out the action, with its own rules.
Let us take a fictitious example. A trading company in Chambéry manages its orders in an internal software tool. With an MCP server, a sales rep on the road asks their assistant which of a given customer's orders are late. The assistant calls the search tool and then summarises the answer. The sales rep has not opened a single screen.
What does Google's announcement change for a French company?
In the short term, little in terms of everyday use. Access is being opened in stages and, for now, custom apps connected to Gemini only work in English. Google plans to widen access to other developers and other countries over the coming months.
The signal, however, is clear. After Anthropic and OpenAI, Google is adopting the same protocol to bring third-party services into its assistants. Part of search will go through a conversation rather than a list of links, as we explained in relation to Google's AI Overviews and their links. A service that exposes its functions cleanly will be easier to use from these assistants.
There is no rush, though. The platforms are changing quickly and an American preview says little about what your customers in Annecy or Geneva will see. On the other hand, preparing your software for this kind of connection is a worthwhile project right now.

Photo: Scott Rodgerson on Unsplash
What can an MCP server be used for in an SME?
For your teams
This is often the most cost-effective use. Your staff already use an AI assistant to draft emails or summarise documents. If it can also query your management tool, it saves them going back and forth between screens. A common example: looking up a customer's history just before calling them, without opening the software.
The value depends on the quality of the tool behind it. An MCP server connected to poorly maintained data produces wrong answers with great self-assurance.
For your customers
A company selling an online service, such as bookings or order tracking, can let its customers act from their assistant, for example to book a slot or track a delivery. This is the model Google is aiming for with Connected Apps. For an SME, it will mainly be relevant if some of its customers already use these assistants every day.
What security risks should you anticipate?
An MCP server opens a door to your data. Google says so in its own help pages: it does not control third-party MCP servers and it is up to whoever operates them to ensure their security. A few precautions avoid unpleasant surprises:
- Start with read-only tools. Looking up an order carries little risk, cancelling it carries more.
- Require proper authentication. Each user logs in with their own account and only sees what their rights allow them in the software. Google's documentation relies on OAuth 2.1 for this.
- Ask for confirmation before any action that changes data or commits to a payment.
- Keep a record of every call: who asked for what and with what result.
- Know where the data travels. A response sent to an assistant passes through its publisher's servers.
This last point matters for companies that handle sensitive data or have obligations towards Swiss customers. We have detailed the location issues in our article on hosting data in Switzerland or in France.

Photo: FLY:D on Unsplash
Do you need an MCP server for your software or your mobile app?
Start by looking at actual use. If your teams or your customers often query your tool to get simple information, the idea is worth exploring. The data still has to be reliable, and you need to be able to say, action by action, what the AI can do on its own and what requires human approval. Without that, it is better to sort those points out first.
If your business relies on custom software, adding an MCP server means exposing some of its functions in a form that assistants understand. The work is reasonable when the software already has a documented API and well-defined access rights. When that is not the case, you have to start there, and it is often an opportunity to modernise ageing business software.
For a custom mobile app, the logic is the same. Android's AppFunctions allow Gemini to trigger actions directly in the app installed on the phone. It is an avenue to keep in mind during design, without making it a priority while the feature remains in preview.
How much does setting up an MCP server cost?
There is no standard price. The cost depends mainly on the state of your software: an existing API and clean data greatly reduce the work. The MCP part itself often comes down to describing a few well-chosen tools, then securing them and testing them with the assistants your users actually use.
The most sensible approach is to start small, with two or three read-only tools reserved for your teams. You will quickly see whether people use it and which data causes problems. Opening the service to your customers afterwards will then be done on a tested foundation.
If you are wondering what your tool could gain from talking to AI assistants, show it to us. From Rumilly, we look with you at how it is built and what it already exposes, then tell you honestly whether it is time to go ahead or better to wait.
Frequently asked questions
What is an MCP server?
It is a program that exposes a piece of software's functions to AI assistants according to the Model Context Protocol, an open standard published by Anthropic at the end of 2024. The assistant can then look up data or trigger an action in that software, within the limits of the rights granted.
What is the difference between an API and an MCP server?
An API is designed to be called by other programs written by developers. An MCP server generally builds on that API and adds descriptions that the AI assistant reads to know which tool to use and with which parameters.
Is an MCP server secure?
It is as secure as you design it to be. You need per-user authentication, limited rights, confirmation before sensitive actions and a log of calls. Google points out that the security of a third-party MCP server is the responsibility of whoever operates it.
Can you already connect your app to Gemini in France?
Not yet for the general public. Google's Connected Apps platform is in preview, opened in stages, and custom apps currently only work in English. Google plans to add markets in the coming months.
How much would your project cost?
Project type, features, timeline: a costed estimate in two minutes, with no commitment.
Is your software ready for AI assistants?
Show us your tool: we look together at how it is built before advising you. Ask for a quote or book 15 minutes with Jérémy.