What Is MCP? The Model Context Protocol Explained for Businesses
MCP connects Claude, ChatGPT and Copilot to your systems. What an MCP server is and what your IT will ask about it: installation, data, permissions.
MCP Servers
Connect AI agents to your systems under control – with sign-in, clear permissions and traceable access.
dectria builds MCP servers through which AI applications such as Claude, ChatGPT, Microsoft Copilot Studio or your own agents access your systems: reading data, finding documents, creating records. Every access runs through defined tools, with sign-in and the permissions of the respective user.
We run MCP in production ourselves: via MCP, Claude or your own agents can connect securely to NetCero, an ESG platform in which dectria holds a stake. Access is secured with OAuth and role-based permissions.
We start with a few narrowly scoped tools and only expand once it is clear how agents actually use them.
What is MCP – and when is it worth it?
The Model Context Protocol (MCP) is an open standard through which AI applications access external data and functions. An MCP server provides tools, resources and prompts that any MCP-capable AI application can use without its own integration.
MCP is worth it when several AI applications or agents need to access the same systems, or when users want to work with company data directly in their AI assistant. It is not worth it when a fixed process should always run the same way – a classic interface is more predictable there. And if agents would write to critical systems, you first need human approval, not more tools.
MCP explained in detail, with the questions IT departments ask before they connect a server: Guide "What is MCP?"
MCP or classic API integration?
| MCP server | Classic API integration | |
|---|---|---|
| Who decides on the call | the model, based on the tool description | hard-coded logic |
| Connecting a new AI application | no extra effort if it supports MCP | separate integration per application |
| Security | OAuth 2.1, permissions per tool | API keys or OAuth, depending on the system |
| Predictability | lower – approval needed for critical steps | high |
| Suited for | assistants and agents with open-ended tasks | fixed, recurring processes |
Capabilities
What We Build with MCP
MCP Server Development (TypeScript) OAuth 2.1 & Protected Resource Metadata Integration with Existing Identity Providers Permissions per Tool (RBAC) Tool Design & Descriptions MCP for Legacy Systems Without Connectors Approval Workflows (Human-in-the-Loop) Logging & Traceability Code Mode & Sandbox Execution Testing with Common AI Clients Operation in the EU
Use Cases
Typical Use Cases
AI Access to Company Data
Employees ask their AI assistant about customers, orders or documents – the MCP server returns only what that person may see.
Agents for Recurring Tasks
Agents create tickets, prepare reports or reconcile data; critical steps go to a person for approval first.
Making Legacy Systems AI-Ready
For systems without a ready-made connector, the MCP server becomes the adapter – with narrowly scoped tools instead of direct database access.
MCP directory
Which software already has an MCP server?
For software that companies use every day, we check the vendors’ own sources: is there an MCP server, an interface, an AI feature?
FAQ
MCP FAQ
What is an MCP server?
A service that provides an AI application with defined tools and data through the Model Context Protocol. The AI only sees what the server offers and calls functions only through these tools. This keeps access to your systems limited and traceable.
How is an MCP server secured?
With OAuth 2.1: the MCP server checks every access token and only accepts tokens issued for it. We grant permissions per tool and follow the principle of least privilege; sign-in runs through your existing identity provider, such as Microsoft Entra ID. Write actions can require human confirmation.
Which AI applications support MCP?
MCP is an open standard. Claude, ChatGPT and Microsoft Copilot Studio can connect custom MCP servers, as can the agent tooling from Anthropic, OpenAI, Microsoft and Google. An MCP server therefore works with several AI applications and your own agents, without an integration per application.
Does MCP work for legacy systems without an API?
Yes. The MCP server then becomes the adapter: it reads from existing interfaces, databases or files and gives agents only narrowly scoped tools, no direct database access. This way, even systems without a ready-made connector can be connected.
What is Code Mode in MCP?
Code Mode is a pattern in which the model does not call MCP tools one by one but writes short code against a TypeScript interface generated from them. The code runs in an isolated sandbox, and only the result goes back to the model. With many tools and multi-step tasks, this saves a lot of context and cost. With few tools and simple queries, direct calls are simpler – we decide based on the scope of the MCP server.
Every project starts with a conversation.