MCP development
Talk to your data. And to your machines.
Without your data, a language model is an adviser without the file. MCP (Model Context Protocol) is the standard through which a model accesses your systems without you exporting data, copying it or moving it into someone else's cloud. We build this connection, on both sides of the company: in administration and in production.
What MCP actually is
MCP is an open standard that allows a language model to call tools. An MCP server provides a handful of clearly defined functions, such as “Find receipts for client X in period Y”, and the model decides when to call them. The decisive difference from a conventional chatbot: nothing is trained, nothing is uploaded in advance and nothing is duplicated. The data stays in your system and is read per request.
For you, this means: your people ask a question in the chat in plain language and get an answer from the actual records. For IT, it means: a manageable, auditable component with a defined attack surface instead of a copy of the data that nobody controls any more.
Two arms, two specialists
This combination is the real point. Business consultants rarely get to the machine, automation engineers rarely get to the accounts. When a manufacturing company wants to know what a stoppage cost, it needs both in one answer.
- Finance and administration — Dr. Maher Hamid
- Abacus, Bexio, SAP, DATEV, ERP and document management systems. Typical questions: contribution margin per customer, variance against budget, open items, searching for documents across years. Principle: read-only.
- Production and engineering — Schady Hamid, B.Eng
- OPC UA, PLC, MES and control systems. Typical questions: why did the line stop, how is OEE developing, which batch deviates. Here, control is also possible, but only with an approval chain and audit log.
The mistake almost every first draft contains
LLM platforms call MCP servers server-side. The request comes from the platform's fixed outbound address, not from the computer of the employee who is typing. Anyone who restricts access to the company network for that reason locks out their own solution. The browser is not in the data path, so a browser bridge does not solve the problem.
We settle this question at the start, not after the security audit. Instead, protection is provided by mutual certificate verification, an assurance from the platform regarding its fixed outbound address, and a permission check in the server itself.
Approach
- 1. Feasibility, 1 to 2 weeks
- Which interface actually supports the requirement? We test the available APIs against your specific questions and tell you which of them cannot be done. The result is a document you can continue working with, even without us.
- 2. Pilot, 2 to 4 weeks
- A working, read-only server with a limited set of functions, connected to your platform, in a separate environment. From here on, we talk about concrete answers rather than slides.
- 3. Go-live
- Permission model per user or permission group, logging, operational handover to your IT or your hosting provider, documentation in German or English. Your people should be able to run the server, not us.
Data protection is part of the architecture, not a chapter in the proposal
A data processing agreement (DPA) with the platform provider is required as a rule. In short paper no. 13, the German Data Protection Conference (DSK) focuses on the possibility of access, not on actual access. The statement “no data is stored, so we don't need a DPA” does not hold up to scrutiny.
Technically, we adhere to the principle that only the MCP server is reachable from outside. The source system, whether DATEVconnect or a PLC, stays in the internal network. In Switzerland, the same pattern applies under the Swiss FADP (revFADP).
What we don't do
- No training and no fine-tuning on your data. MCP reads at runtime; it learns nothing.
- No copy of the data into a secondary database just to make search faster.
- No write access to financial systems. Postings stay with your people.
- No connection whose permission model we cannot reproduce. If a system only offers all-or-nothing access, we say so before the proposal.
Frequently asked questions
What is an MCP server?
An MCP server is software that offers a language model defined functions on your systems, such as a document search or a key figure query. The model calls these functions as needed. The data stays in the source system and is read per request, not copied or trained on.
Which AI platforms does this work with?
With all that support MCP. In practice, these are Claude, ChatGPT and enterprise platforms widely used in the DACH region, such as Langdock. Because MCP is an open standard, the same server remains usable if you switch platforms.
What does an MCP project cost?
The effort is almost always determined by the source system and its permission model, not by the AI side. Shared access is quick to build; mapping per user or permission group is not. That is why we do not quote list prices, but provide a proposal after the initial meeting.
Do we need a data processing agreement for this?
As a rule, yes. What matters is the platform provider's possibility of access, not whether data is actually stored. This corresponds to short paper no. 13 of the German Data Protection Conference (DSK). We provide the necessary information on the architecture; you conclude the agreement yourself.
Can we run the server ourselves?
Yes, and that is the normal case. The server runs in your environment or with your hosting provider. You receive the source code, documentation in German or English and an operational handover. Kapitel H works towards you no longer needing us afterwards.
How long does it take to get a first usable version?
From the feasibility study to a pilot with which your specialist department can ask real questions typically takes four to six weeks. The bottleneck is almost always the approval of the interface in the source system, not the development.
Let’s check whether it holds up for your systems.
The feasibility study takes one to two weeks and ends with a document you can keep working with even without us. If an interface cannot meet your requirement, it says so.
Request a first call