MCP (Model Context Protocol).

MCP solves an integration problem that grew with every new model and every new business system. Without a standard, each combination of AI application and data source needs its own connection; ten applications and ten systems mean a hundred individual integrations, all maintained separately. MCP inverts that: each system is exposed once as an MCP server, each AI application speaks MCP once, and a hundred integrations become twenty. The frequently used comparison with a common plug type fits well, because the real gain is not technical elegance but the disappearance of special cases.

The architecture is deliberately plain. A host — a chat client or an agent application — starts one MCP client per connection, and that client talks to exactly one MCP server. The server offers three kinds of building block: tools, meaning callable actions with a described input schema; resources, meaning readable content such as files or records; and prompts, meaning ready-made templates for recurring tasks. When a connection opens, both sides negotiate which capabilities they support. Transport is standard input and output locally, and an HTTP variant with streaming over the network.

Anthropic published MCP in November 2024 and handed governance of the standard to the newly founded Agentic AI Foundation under the Linux Foundation in December 2025, together with other companies as founding and platform members. MCP is therefore no longer one vendor's product but a vendor-neutral standard with an open process. For companies the practical relevance is investment protection: a connection built once remains usable when the model is swapped. The major assistant products and development environments support MCP natively today, and public directories list several thousand servers.

When introducing MCP inside a company, security is the first question. An MCP server holds real permissions on real systems, and a model decides when it gets called. Sensible practice: grant only the minimum permissions needed, review third-party servers before use, separate writing actions from reading ones, and treat anything a server returns as potentially manipulated content. A typical first use case in an SME is an internal server that exposes quotes and project status read-only — useful from day one and incapable of breaking anything.