From chatbot to agent: what MCP changes
A chatbot talks. An agent searches, calculates, drafts and prepares. The difference is not in the language model but in what it is allowed to reach.
Why that first chatbot did not change your work
Almost everybody has tried a chatbot by now, and most organisations stopped afterwards. The conversation was fun, but nothing changed about the work. That is not down to the language model. It is down to the chatbot not being able to reach anything: an agent that cannot search your catalogue, does not know your volume tiers and may not use your quoting system can only talk.
So the question that matters is not how clever a system sounds, but what it is allowed to touch. That is exactly what MCP arranges. This piece is about that layer: what an agent may request, what it may write, and why the approval step in between is not a tick box in a settings screen but a decision you take at the start.
What an agent is, and what it is not
The word agent is used for three very different things, and that difference decides whether anything about your work changes. So with every request we first say which level something sits at. That avoids the disappointment that starts with a demonstration at level one and an expectation at level three.
The three levels differ in what they may touch, not in how clever they sound. An assistant that knows your whole archive sounds more impressive than an agent allowed to carry out one action — but only the second takes work out of your day. So with any supplier, watch what the system may do, not what it can say.
Level 1 — the model
It can understand and produce language, but it knows nothing about your company. It is a good writer without access to your archive. Anything it says about your prices is a guess that sounds plausible.
Level 2 — the assistant
It knows your documents and answers questions about them. Useful, and it saves searching, but it changes nothing about the work itself: you still have to build the quote, find the item and apply the volume tier.
Level 3 — the agent
It may carry out actions in your systems: search, calculate, create, prepare. That way part of the work disappears instead of moving. This is the level at which MCP becomes necessary, because without an agreement about access it gets nowhere.
One agreement about what an agent may request and may do
MCP is the standard for connecting an agent to a system safely. Not a separate bespoke integration per question, but one agreement about what an agent may request and may do. We have MCP servers for our PIM, the quoting module, the ERP and the mailbox, and we use them to connect systems you already have. Per connection we set out what is readable and what has to pass a person.
That is not only a safety measure, it is also the reason it works. An agent that can reach your real prices and volume tiers over MCP gives an answer that is correct; an agent that has to take it from a manual invents something that sounds plausible. With us the MCP servers sit in Magic Orb — not a separate product, but the layer that makes the rest of Magic Studio reachable. It is also how we connect to an ERP you already run.
Two very different questions
In practice every connection falls apart into two questions that have to be answered separately. The first is what the agent may request. There the answer is usually generous: an agent allowed to read your catalogue, your stock and your price agreements can answer a question without anybody having to search. Reading does no harm as long as the separation per customer is right.
The second question is what it may write, and there the answer is almost always narrow. Creating a draft quote is not the same as placing an order; requesting a sample is not the same as changing a price. Lump those two questions together and you get either an agent that can do nothing useful, or an agent that can do too much. So per connection it is set out separately what may be read and what may be written.
The approval step is a design decision
It disappears easily in a demonstration: the moment the agent stops and asks for confirmation. Yet it is the most important part of the design. The agent prepares, the person decides. Nothing goes to your customer without approval, and your sales desk approves or rejects line by line — the order is then built from the approved lines.
That is not a setting you switch on later, because it determines how the flow around it is built. Which steps the agent may take is set out in Flowbuilder: in a flow you can read and adjust, not in code you have to trust. What it may not do is in there just as firmly. Every step goes into a log, so you can see why the agent did something without having to call us.
What we do not do
No agent between you and your customer without an approval step. No quote that goes out automatically. No agent changing prices on its own. And no company data of yours going to third-party models. That costs a little convenience; it is also why your sales desk still trusts the agent after two weeks.
What happens when it gets stuck
If a conversation stalls or gets delicate, a member of staff takes over — including in the middle of a WhatsApp conversation. Sensitive actions also get a second check. An agent that honestly says it cannot work something out is worth more than one that carries on.
What this does today
At the Brinxx group the quoting agent sits between the request and the sales desk, for around 380 quotes a month across Brinxx, Promotionalz and Moxz together. It reads the request, finds the item in a catalogue of 74,136 products, picks the variant, applies the print volume tier, has Magic Logo produce the proof and prepares the quote. The searching and the calculating are gone; the checking and sending are not.
That it works is easier to see than to read. Three recordings show the MCP layer on real figures: a call list rolling out of the outstanding quotes, a revenue report over the current year, and a request for the top five products against the top five customers. In all three somebody asks a question in plain language and the answer comes from the data that is in the system today — including the moments where the agent asks for confirmation.
What this is about
The MCP servers sit in Magic Orb; the agent running on them is Magic Agent. The agentic AI page has the full story, including the four-step approach and the limits we keep.
Agentic AI and MCP
The three levels, the four-step approach and what we deliberately do not do.Magic Agent — the agent that does the groundwork
From request to draft quote, with a person approving line by line.The demo recordings
Twenty recordings of our own environments, three of them on the MCP layer.Ask your question
Send us last week's requests and we will set the agent on them.Curious what an agent would do at your place?
We would rather start with a data scan than with a demonstration of a chat window. A chat window always looks good; a count of your variants, your print areas and your missing copy tells you what an agent can really do for you.
Rather watch first? The recordings are with the demo videos.