Let AI work with your systems.
Your team has ChatGPT, Claude or Copilot. It can write an e-mail. It can not check an order, a stock level or an invoice.
I build custom MCP servers: the connection between your AI assistant and your CRM, ERP, e-shop or database.
You decide what the AI can read, what it can change and what needs a person's approval.
Price: From €3,000. First version in 1 to 2 weeks.
What MCP is, in two sentences
MCP (Model Context Protocol) is an open standard for connecting AI assistants to business software. ChatGPT, Claude, Gemini and Microsoft Copilot all support it.
An MCP server is a small, controlled doorway into one of your systems. It tells the AI exactly which questions it can ask and which actions it can take. Nothing else is open.
You build the connection once. It works with the AI tools you use today and the ones you switch to later.
You probably need one if
- People copy data between systems by hand.
- Your AI rollout stopped at writing texts.
- Your software has no AI connector, or the connector does too little.
- You run an internal system nobody else will connect for you.
What changes in the company
Five examples. The times are targets we measure, not promises.
| Team | Today | With the connection | What we measure |
|---|---|---|---|
| Sales | Five tabs open before every customer call | One question, one account summary from CRM and ERP | Preparation time per call |
| Support | Agent searches e-shop, ERP and tickets for one order | AI finds shipping, payment and return status | Handling time per ticket |
| Quotes | Stock, discounts and prices checked by hand | AI drafts the quote, a person approves it | Time from request to quote |
| Finance | Overdue invoices matched to accounts by hand | AI prepares a prioritised list of who to call | Hours per week |
| Management | A report request waits for an analyst | A question in plain language, answered from data the person may see | Time to answer |
Three packages
Start small. Read-only first, actions second.
Connect
One system. Up to five read-only functions. One AI assistant. Login and a log of every request.
1 to 2 weeks. You get the running server, the source code, tests and documentation.
From €3,000
Act
One system. Up to eight functions, three of them can write. Permissions by role, approval before every change, audit log.
3 to 5 weeks. Includes deployment, security tests and handover.
From €7,000
Operate
Two systems. One defined workflow from start to end. Error handling, monitoring, acceptance tests.
6 to 9 weeks. Includes production rollout and team training.
From €15,000
I am not a VAT payer. The price is final.
You pay in your own currency: the payment page shows CZK in Czechia, EUR in the eurozone fixed prices in USD and GBP in the US and the UK, and your local currency elsewhere (CHF, PLN, AED and more, converted by Stripe).
All packages need a system with a documented API or database access. Old systems without one are scoped separately. Maintenance after handover: from €300 a month.
Security comes first
MCP does not make a connection safe by itself. The way it is built does.
Least access
The AI gets only the functions on the agreed list. No general access to the database.
Read before write
Every project starts read-only. Write actions come later, one by one.
A person approves changes
Anything that changes data, sends a message or costs money waits for a human yes.
Every request is logged
Who asked, what the AI did, what came back. You can check it any time.
A map of where data goes
Before the build you see which data leaves your systems and to which AI provider.
We try to break it
Before handover we test whether the AI can be talked into going past its permissions.
How it works
Pick one workflow
A free 20-minute call. We choose one task with a clear owner and a number to measure.
Scope and access map
A written list of functions, permissions, approval points and acceptance tests. You sign it off.
Build
The server is built and connected to your AI assistant in a test environment.
Test
Acceptance tests, security tests and a trial with real users.
Hand over
Source code, documentation, training. You own everything.
Who builds it
I scope the project, define the tests and check the result. Dominik, a senior developer I work with, builds the server.
My own brand runs on AI agents connected to my tools. The playbook and the failure log are public.
There is no client case study for this service yet. The first one will be published here, with numbers.
The first three clients
A reduced pilot price for the Connect package, in exchange for measured results and an anonymised case study.
When you do not need this
Your software already has an AI connector that does what you need. Use it. I will tell you so on the call.
The task happens twice a month. A custom connection will not pay back.
Not sure about the numbers? Try the AI automation ROI calculator first.
Frequently asked questions
Does our data leave the company?
The server can run on your own infrastructure. The answers the AI reads still go to your AI provider. Before the build you get a map of what goes where, and we set it up to match your rules.
Our software already has an MCP connector. Why build a custom one?
Often you should not. Use the vendor's connector when it covers your permissions and actions. A custom server fills the gaps.
Are we locked in to you or to one AI vendor?
No. You own the source code. MCP is an open standard, so the same server works with several AI assistants.
Who maintains it?
APIs change, so somebody has to own it. Your developers can, with the documentation. Or we do, from €300 a month.
Which systems can you connect?
Anything with a documented API or database access: CRM, ERP, e-shop, ticketing, BI, internal systems. We check it before quoting a price.
Do I need technical knowledge?
No. You describe the task and what a correct result looks like. Your IT person joins for access and security.
Start with one system
Tell me which system and which task.
In 20 minutes you will know whether a custom connection makes sense, and which package fits.