What an MCP server does and where its limits are
An MCP server provides tools, resources and prompts to an AI client. A tool might search for a ticket, run a database query with fixed parameters or fetch a build status. The agent decides when to call a tool; the server decides what is possible at all.
That leads to the core design rule: as little as possible, as much as necessary. A tool that runs arbitrary SQL is convenient but risky. Narrow tools with clear inputs, permission checks and logging are better. Prompt injection must be considered too, because content from tickets or documents can contain instructions the agent should never follow.
From workshop to running server
neooutput's AI code engineers work with your team remotely or on site, combining learning and building:
Operations deserve special attention. A local MCP server runs on the developer's machine and uses their credentials. A central remote server needs its own authentication, tenant separation and monitoring. We weigh both options with you and take your information security and data protection requirements into account.
- In the workshop we collect use cases and rank them by value and risk.
- For the most important ones we define tools, inputs, outputs and permissions.
- We decide on transport and hosting, either local via stdio or as a remote server over HTTP with authentication.
- The first MCP servers are implemented, tested and connected to Claude Code and other clients.
- Your team receives documentation and templates to build further servers on its own.
Documents and sensitive data
Many use cases revolve around knowledge held in documents: manuals, policies, project reports. For company-wide knowledge search with access control, kalAIDoCreate is often a better foundation than a home-grown server, and an MCP server can then act as the bridge to your agents.
If an MCP server returns content containing personal data, for example from support tickets, Textwash Pro can sit in front of it and anonymise the text automatically before it reaches the agent's context.
Frequently asked questions
Which programming language suits MCP servers?
Official SDKs exist for TypeScript, Python, Java, Kotlin and C#, among others. We choose the language that fits your team so the server can be maintained internally after handover.
Do custom MCP servers only work with Claude Code?
No. MCP is an open standard now supported by many clients, including Claude Code, Claude Desktop, OpenAI Codex and Cursor. A well-built server can therefore serve several tools.
How do we secure an MCP server?
With narrowly scoped tools, authentication, per-request permission checks, logging and tests against prompt injection. Write operations should also require human confirmation.