MCP (Model Context Protocol): definition
MCP (Model Context Protocol) is an open protocol introduced by Anthropic in late 2024. It standardizes how an AI assistant connects to external sources: tools, databases, documents, internal services. Before MCP, each integration between a model and a data source required a custom connector. The protocol offers a common interface, which reduces integration work and makes connections reusable from one assistant to another.
How MCP works
MCP is built on a client-server architecture. An MCP server exposes resources (data to read) and tools (actions to trigger) in a standardized format. An MCP client, embedded in an assistant such as ChatGPT, Claude, Cursor or Gemini, discovers those resources and uses them during a conversation. The model can then read a piece of content, call a tool, and fold the result into its answer.
This separation has a concrete benefit: the data source stays under the control of whoever exposes it. The server defines what is readable, what is writable, and within which limits. The model only queries that interface.
Why MCP matters for a company
The main benefit of MCP for an organization is portability. A company can plug its business memory into several AIs that read and enrich it, without locking it inside a single model. If the assistant in use changes, the source stays the same and remains accessible.
This neutrality connects to the question of data sovereignty: the company's memory lives in a system it controls, rather than inside a given AI vendor. That logic is developed in the article why your memory should not live inside any AI. Verbasil relies on MCP to make its memory readable and enrichable from these different assistants.
FAQ
Is MCP tied to a single AI vendor?
No. MCP is an open protocol. It was introduced by Anthropic, but it is designed to be implemented by different clients. Several assistants (ChatGPT, Claude, Cursor, Gemini) can query the same MCP server.
Does MCP give full access to my data?
No. An MCP server exposes exactly the resources and tools it chooses to make available, within the limits set by whoever operates it. The model only reaches that interface, not the entire underlying system.