How an MCP server works
An MCP server offers an AI client three kinds of building blocks: tools, which are callable functions such as a customer search; resources, such as documents or records; and reusable prompts. Local servers usually communicate over stdio, while shared servers use Streamable HTTP with OAuth-based authorization.
For an ERP, a server might expose functions to search items, check stock or show open orders. A CRM server could return contact history and account owners. The assistant picks the right tool based on the question and turns the result into an answer.
Internal knowledge such as manuals and policies is often better served by a RAG platform like kalAIDoCreate. The two approaches can be combined.
Security questions to settle first
An MCP server is a new way into your data and needs to be secured accordingly. We apply least privilege, start with read-only functions and add write actions only with explicit human confirmation.
Prompt injection deserves special attention: text inside your data sources can contain instructions that mislead the model. We therefore validate inputs, limit how much data is returned and keep tenants and roles strictly separated. Every server goes through a code security review before production use.
We also recommend authorizing access per user rather than through one shared service account. That way the permissions from your ERP or CRM carry over, and a sales rep cannot see more through the assistant than in the original application.
From use case to production server
If you want your developers to build and maintain MCP servers themselves, we can pair the project with a workshop on Claude Code, MCP and secure review practices.
- In a free 15-minute intro call we choose a use case with clear value and manageable risk.
- We analyze the interfaces of your ERP, CRM or database and define tools, permissions and logging.
- Our AI code engineers build the server, test it with your clients and run a security review.
- After the pilot you run it yourself, or we operate it as part of neooutput with 24/7 support.
Frequently asked questions
Which AI clients support MCP?
Claude Desktop, Claude Code and Cursor among others, and the number of MCP-capable clients keeps growing. We clarify which clients are approved in your organization during the project.
Can the AI change data in our ERP?
Technically yes, but we recommend a gradual path. We start read-only and add write actions only with human confirmation and clear logging.
Do we need a modern API?
A documented REST or GraphQL API makes things easier. Older systems can often be connected through restricted database access or existing export interfaces.