Model Context Protocol (MCP).

The Model Context Protocol, or MCP for short, is an open standard used to connect AI applications—such as chat assistants or AI agents—to data sources and tools. Anthropic introduced MCP in November 2024 because, until then, each data source required its own connection. In December 2025, according to its own announcement, Anthropic handed the protocol over to the Agentic AI Foundation, which operates under the umbrella of the Linux Foundation and was founded by Anthropic, Block, and OpenAI. MCP matters for your company because it determines how easily and how securely an AI can access your CRM, ticketing system, or document management system.

How the Model Context Protocol Works

MCP connects three roles via a fixed message format. According to the specification—currently in the version dated July 28, 2026—there are hosts (i.e., the AI application itself), clients (which act as connectors within this application), and servers that provide data and capabilities. The messages follow the JSON-RPC 2.0 format. MCP is thus a specialized form of an API tailored to language models.

According to the specification, an MCP server can provide three things: resources (i.e., data and context), prompts (i.e., pre-prepared templates for users), and tools (i.e., functions that the model is allowed to perform). For example, a server for your ticket system could read open tickets and create new ones.

What You Should Keep in Mind at Work

Treat every MCP server as software that has access to your systems. The specification itself states that tools can execute arbitrary code and should be handled with caution. Descriptions of a tool are considered untrustworthy unless the server comes from a trusted source. Servers of unknown origin therefore do not belong on work computers.

Grant only the permissions that a task actually needs. According to the specification, users must explicitly consent to every data access request, and hosts must obtain consent before each tool call. However, the protocol itself cannot enforce these rules—the application must do so. So make sure your AI tool properly requests permissions and verify whether a server is allowed only read access or also write access.

Clarify the data flow before you integrate a business system. Whatever an MCP server provides ends up in the context of the language model and thus with the model provider. In AI projects, we often see that the technical integration is set up quickly, but issues regarding access rights and order processing take significantly longer to resolve. Plan for this time from the very beginning—ideally in conjunction with an AI policy for your employees.

Frequently Asked Questions About the Model Context Protocol

Is MCP just for Claude? No. Anthropic developed MCP, but the protocol is open-source and is now maintained by a cross-vendor foundation, of which OpenAI is one of the founders.

What is the difference between MCP and an AI agent? MCP is the connection; the AI agent is the system that uses these connections to perform tasks in multiple steps.

Do we need our own MCP servers? Right from the start, Anthropic has released ready-to-use servers for popular systems such as Google Drive, Slack, GitHub, and Postgres. For internal systems, it’s worth having your own server if multiple teams need the same access. We outline which workflows are best suited for implementation first in “AI Agents for Small and Medium-Sized Businesses 2026.”

If you want to connect your systems to AI via MCP without losing control over your data and rights, take a look at how we approach AI for businesses.