MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard that connects AI applications — LLMs, AI assistants and AI agents — with external data sources and tools. Published by Anthropic at the end of 2024, the protocol is now supported by a broad range of AI tools. Instead of building a separate integration for every combination of model and system, one MCP server per system is enough.
What problem does MCP solve?
An LLM only knows what was in its training data and what is supplied in the request. For real tasks, it needs access to live systems: databases, ticketing systems, engineering tools, business applications. Before MCP, this meant point-to-point integrations — every AI application multiplied by every system resulted in its own connection. MCP standardizes this interface: a system provides an MCP server once, and any MCP-capable AI application can use it. N×M individual connections become N+M building blocks.
How does MCP work?
MCP follows a client-server architecture: the host is the AI application — such as a chat interface, an IDE assistant or an agent — which maintains connections to one or more MCP servers via clients. A server provides three kinds of building blocks:
| Building block | Function | Example |
|---|---|---|
| Tools | Actions the model can invoke | run a query, create a ticket, generate a document |
| Resources | Data and context to read | files, data sets, model and requirements data |
| Prompts | Predefined templates for recurring tasks | review checklist, analysis workflow |
Which tool is useful when is decided by the model during the conversation — within the permissions and approvals set by the server and the host application.
MCP in engineering toolchains
In engineering organizations, knowledge is spread across tool boundaries: requirements in DOORS or Jama, architecture models in Enterprise Architect, code in GitLab. While standards such as OSLC connect engineering tools to each other, MCP opens them up to AI assistants: a model can query requirements and model data directly — without manual export steps — and answer questions such as “Which requirements are affected by this architecture change?”
MCP in business processes and for data sovereignty
Outside engineering, too, MCP is the usual way to connect AI to enterprise systems in a controlled manner. Two patterns from practice: firstly, models can be coupled to internal systems via MCP without data leaving the company — the server runs in your own environment and only exposes what is meant to be exposed. Secondly, a project-specific MCP server filled with context can give the AI a precise semantic frame — for example in legacy modernization, where it provides system knowledge, architecture rules and dependencies and thus rules out structural errors in code generation.
MCP in practice: connectivity is not the same as context
The protocol solves the technical connection — nothing more. Whether an MCP server is useful is determined by the context it provides in a structured way: poorly maintained data does not get better through a standard. And because a server gives the model real permissions, it deserves the same care as any other interface — clearly limited access rights, approvals for write actions and traceability of what the AI did and when.