Let AI agents do more than read about you —
let them order from you
Being found is layer one. The real head start belongs to businesses where an agent can actually act: check availability, reserve, order, request a quote. One MCP server, and every compatible assistant can reach you.
From €3,950 excl. VAT · permissions, logging and handover included
The Model Context Protocol (MCP) is an open standard that lets AI models use tools and data sources. A business publishes one MCP server exposing its functions — check availability, book, order — and every assistant that speaks the protocol can connect to it. One integration instead of one per assistant.
Why this matters more than it sounds
Every time a new assistant appeared, businesses faced the same choice: build yet another custom integration, or sit this one out. That is precisely what kept AI saying things rather than doing things.
MCP removes that barrier. You describe once what your system can do, in a form a model understands, with clear permissions and limits. From then on it is up to the assistant to connect — not up to you to build per platform.
The consequences are concrete. A restaurant whose availability is queryable becomes an option inside an agent planning an evening. A wholesaler whose stock is queryable enters the comparison a buyer asked for. A business with only a contact form drops out of both — not because it is worse, but because it is unreachable to the party doing the choosing.
What you typically expose
- Availability and capacity — free tables, seats, time slots, stock, planning room.
- Prices and terms — current rates, packages, discounts, delivery conditions.
- Actions — reserve, order, book an appointment, request a quote, buy a ticket.
- Status queries — where is my order, is my appointment confirmed, what is my balance.
- Internal knowledge — for agents your own team uses: procedures, files, history.
Outward or inward?
Outward: reachable for your customers’ agents
You make a defined slice of your system available to external assistants. That is the offensive side: you become an option in decisions that would otherwise be made without you. The security layer is critical here — you decide exactly what is allowed, how much, and with what confirmation.
Inward: agents that take over your own work
You build agents for your own team, with access to your own systems. Qualifying enquiries, drafting quotes, reading documents, proposing rosters, sorting orders, producing reports. This is usually where payback is fastest, because you see the effect in hours saved.
How an implementation proceeds
Start small, expand safely. Never the other way round.
Scoping and risk analysis
Which functions are worth exposing, and which absolutely are not? We set the boundaries upfront: what an agent may read, may change, and where a human must always sit in between. This is the most important conversation of the whole engagement.
Building the MCP server
Functions are described in a form models understand, with strict input validation, per-call permissions and full logging. Everything an agent does is traceable afterwards to who, when and why.
Testing with real agents
Not just unit tests, but actual assistants performing the tasks your customers would give them — including the strange ones. Agents do things no developer imagines; you want to find that out before your customer does.
Cost ceilings and monitoring
Limits per user and per period, alerts on abnormal behaviour, and a kill switch. An agent stuck in a loop should become an incident, not an invoice.
Handover and documentation
You get the code, the configuration and the documentation. Your own developer or a next partner must be able to take this over without me. That is not a nicety, it is the acceptance condition.
An example from my own practice
I build and run Cilurzo Hospitality, a platform for reservations, ordering, events, staff scheduling and food costing used by real restaurants. AI does not have a demo role there: a voice assistant takes reservations by phone, Deliveroo and UberEats orders are sorted automatically into the kitchen flow, and rosters are proposed based on expected footfall.
That is exactly where this topic stops being theory: permissions that hold, failures you can absorb, costs that are bounded, and a human with the final word at the right moment. What I build for clients, I run in production myself.
Price
An MCP implementation starts from €3,950 excluding VAT for a defined set of functions on a system with a usable existing integration. When an API has to be built first on a closed or dated system the price is higher; I estimate that after a technical exploration and fix it in writing beforehand.
Frequently asked questions about MCP
What is MCP (Model Context Protocol)?
MCP is an open standard that lets AI models use tools and data sources. A business publishes one MCP server describing what its system can do — check availability, reserve, order — and any assistant supporting the protocol can connect. It replaces the situation where you had to build a separate integration for every AI platform.
Isn’t that just an API?
There is an API underneath, but the difference is the description. A classic API is written for a developer reading documentation. An MCP server describes itself in a form a model can interpret directly: what each function does, when to use it, what the input means and what can go wrong. That lets an agent use it correctly without anyone writing code for it.
Is it safe to let agents into my systems?
Only if you design it that way. Safety comes from scoping (only what is needed), per-call permissions, strict input validation, human confirmation on irreversible actions, full logging and hard cost ceilings. An MCP server that opens everything is dangerous; a well-scoped one is manageable. That design is most of the work.
Which businesses benefit today?
Businesses where availability or stock drives the decision: hospitality, hotels, practices with appointment books, rental, wholesale and logistics. And businesses with a lot of repetitive internal work around enquiries, quotes and documents — that is where the inward gain sits.
What if the standard changes?
It certainly will, this is a young field. So I always build the business logic separately from the protocol layer: the functions live in your own system, MCP is only the wrapper. If the standard shifts or another arrives, you replace the wrapper, not the contents.
Can an agent do anything with you today?
Tell me briefly which systems you run. Within 48 hours I will tell you whether a useful MCP layer fits on top and roughly what it would cost. Free and without obligation.