AI protocol · in regular use
MCP agency in Nice
The Model Context Protocol: the standard that lets an AI assistant such as Claude or ChatGPT use your tools and data, with defined permissions.
What we do with it
MCP at our clients, concretely
We build MCP servers: on Ordelya, an assistant manages the restaurant's menu, with confirmation before deletion and the ability to undo; on LLM Monitor, it queries a brand's visibility data. We also run WordPress sites through MCP with the Novamira and IATO plugins, always on staging first.
The limits
When we recommend something else
Giving an AI agent write access with no log and no undo, or access to sensitive data without scoping, is a risk: without those safeguards, we do not connect it.
Our work
Projects we delivered with MCP
LLM Monitor
SaaS that measures how visible brands are in AI answers.
MCP: MCP server: an assistant queries a brand's visibility, competitors and reputation.
Read the LLM Monitor case studyOrdelya
Restaurant management platform: back office, tablets, kiosks, screens and website.
MCP: MCP server: an AI assistant manages the menu, with confirmation before deletion and undo.
Les Sens
Redesign of a Nice restaurant's website, in three languages, with online booking.
MCP: Site run through MCP (Novamira, IATO): changes made on staging, with rollback.
Site d'une entreprise de recherche de fuites
Nice-based showcase website built on our WordPress base theme.
MCP: Deployment and maintenance run through MCP with Novamira.
Site d'un architecte d'intérieur
Bilingual website showcasing projects and furniture.
MCP: Elementor content edited through MCP, with dry runs and rollback.
Method
How we bring MCP into a project
The Model Context Protocol (MCP) is an open protocol that lets an AI assistant such as Claude or ChatGPT use a piece of software's functions: reading data, creating it, changing it. We build MCP servers for our own platforms. On Ordelya, an assistant can manage a restaurant's menu: view it, search for dishes, add them, edit them, mark them unavailable, delete them after confirmation or undo a change. Access is by personal key or OAuth. On LLM Monitor, an assistant queries a brand's visibility data in AI tools, covering competitors, reputation and conversations, protected by API keys and OAuth.
We also operate WordPress sites over MCP, using two third-party plugins that we did not develop: Novamira, which runs code and WP-CLI commands with files kept in a sandbox, and IATO MCP, which edits Elementor content with a dry run beforehand and rollback. Our rule: identifiable content goes through the tool that allows undoing, global settings through Novamira, and every intervention starts on staging, with production only changed once the result has been confirmed.
For your business
What it changes in practice
For users, MCP replaces a series of clicks with a request in everyday language. A restaurant owner can ask their assistant to mark a dish unavailable or find every dessert on the menu, without opening the Ordelya interface. A marketing team can query LLM Monitor data about a specific competitor. As with any AI project, we start nothing without a use case costed in hours saved per week.
For management, the issue is control. An assistant acting on your data must only be able to do what has been planned. Our MCP servers restrict permissions to defined operations, ask for confirmation before any deletion and allow a change to be undone; access goes through a personal key or OAuth, which can be revoked. On WordPress, the dry run, rollback and mandatory pass through staging keep an assistant's mistake from reaching the live site. The decision to publish stays with a person.
Decision guide
Should you choose MCP?
An MCP server or a conventional API integration?
An API integration suits cases where one piece of software calls another automatically, always following the same scenario. An MCP server is justified when people want to act on your data through an AI assistant, with varied requests phrased in everyday language. Both often rely on the same internal functions. On Ordelya, the MCP server reuses the menu management operations already present in the platform.
What permissions should an AI assistant have over our data?
The minimum the intended use requires. We define a closed list of operations rather than general access: on Ordelya, the assistant works on the menu, and a deletion requires explicit confirmation. Every change can be undone. Access goes through a personal key or OAuth, which shows who is acting and lets one person's access be withdrawn without affecting anyone else. Read-only operations are sometimes enough, as when querying LLM Monitor data.
Should our WordPress site be operated by an AI assistant?
Only if the gain is real and the safeguards are in place. We do it with third-party plugins, Novamira and IATO MCP, under a strict rule: identifiable content is changed through the tool offering a dry run and rollback, global settings through Novamira, and always on staging before production. For a few changes a month, the standard admin area is often simpler. We talk it through in a free first conversation.
Client questions
What clients ask us about MCP
Which AI assistants can connect to an MCP server?
Any assistant that supports the protocol, such as Claude or ChatGPT. That is the point of an open protocol: the Ordelya or LLM Monitor MCP server is not written for a single provider. Your staff use the assistant they already know, and the connection is made with a personal key or through OAuth, depending on what the assistant supports.
Could an assistant delete data by mistake?
We make that as hard as possible. On Ordelya, deleting a dish requires confirmation, and marking a dish unavailable remains a less final option. A change can be undone. The assistant only has the planned operations, with no general access to the database. On WordPress, we work through tools offering a dry run and rollback, and always on staging first.
Did you develop Novamira and IATO MCP?
No. They are third-party WordPress plugins that we use to operate sites over MCP. Novamira runs code and WP-CLI commands, with files placed in a sandbox; IATO MCP edits Elementor content with a dry run beforehand and rollback. Our work lies in how they are used: which tool for which change, and staging first.
Does our data pass through the assistant's provider?
Yes, in part: whatever the MCP server returns to the assistant is read by the model, and therefore processed by its provider. That is why we limit the exposed operations to what the use case needs. For sensitive data, we review with you what can be passed on, the chosen provider and its contractual terms before access is opened.
Where do we start with MCP?
With a free first conversation, in which you describe the operations your staff repeat in a piece of software. We check whether an assistant would save hours each week, and what permissions it would need. You then receive a written, costed action plan setting out the exposed operations, the confirmations required and the access method. The server code belongs to you.
Going further
MCP sits within our Artificial intelligence
A technology does not make a project. The method around it decides the outcome: scoping, prototypes, short cycles, and a clean way out.
Neighbouring technologies
A MCP project in mind?
The first conversation is free. We will tell you plainly whether this technology is the right one, and if not, which is.