Case study · Custom MCP server
AI fault diagnosis: service knowledge from four systems
- Area
- Tools & knowledge
- Scenario
- Machine-building SME, eastern Switzerland (demo data)
- Technology
- Model Context Protocol, Node.js
- Deployment
- Hosted in Switzerland, login via Microsoft Entra ID
Starting point
In our reference scenario, a machine builder in eastern Switzerland services packaging and filling lines for customers across the country. When a machine reports a fault, the service technician pulls the information together from four systems: error messages and readings in remote maintenance, past service calls in the service database, spare parts and prices in the ERP, response times in the service contract. That takes time, especially at night or for new staff. At the same time, an AI assistant should only be introduced if customer data stays protected and nobody hands the AI a master key to the ERP and remote maintenance. There is no ready-made MCP server for the industry software in use.
The solution
andyou developed a custom MCP server that bundles the four systems behind six domain-specific tools. Instead of passing on raw data, each tool answers a concrete service question. The status tool condenses 24 hours of readings into trend, minimum and maximum and tolerance violations. The diagnosis weights possible causes using the fault manual, current readings and identical cases across the whole fleet, and returns the check to perform, the spare part with its storage location and the cost under the service contract. Because MCP is an open standard, the same server works in Claude, ChatGPT and Microsoft Copilot.
Security & privacy
The server gives the AI only what it needs for the task. Phone numbers and email addresses of customer contacts never leave the system. Technicians see no prices, sales support does; the role comes from the login, not from the request. Customer reports can contain hidden instructions. The server therefore explicitly marks them as data and does not release contact details even when a report asks for them. The AI only creates service calls as drafts, after explicit confirmation. Every call is logged.
Result and cost
The diagnosis that used to require opening four systems is ready after one question. Because the tools condense rather than pass on raw data, the language model reads around 940 instead of 71,600 tokens per diagnosis. That cuts AI cost by 99 per cent: at 600 diagnoses a month, from around CHF 103 to CHF 1.40. The bigger lever is at the customer: less time to the right diagnosis and therefore shorter downtime. The proof of concept is covered by eight automated tests, including privacy, roles and approvals.
Technology in brief
The server uses the official MCP SDK on Node.js and runs either locally or as a stateless remote service over Streamable HTTP. For operation, a gateway with sign-in via Microsoft Entra ID is planned, hosted in Switzerland. The data connection is encapsulated in one place: in the customer project it is connected to remote maintenance (OPC UA), the service database and the ERP. Tools, rules and tests stay unchanged.
- Model Context Protocol
- Node.js
- MCP SDK
- Zod
- Streamable HTTP
- OAuth · Entra ID
- Swiss hosting
- Claude · ChatGPT · Copilot
Connect your systems securely to AI?
We check whether existing MCP servers are enough or whether a custom server reflects your processes better, and show you in a proof of concept.
Let’s talk about your project