MCP and AI basics
MCP host
The AI application a person actually uses, such as a chat assistant or a coding assistant. The host runs one or more MCP clients and decides what the model is allowed to do.
The MCP host is the application a person actually uses: a chat assistant, a coding assistant or an agent runtime. It contains the model, runs one MCP client per connected server, and decides what the model is allowed to do. Approval prompts before a tool runs, limits on which tools are enabled and logging of calls all live in the host.
This matters for governance. The same MCP server can behave very differently in two hosts: one asks you to approve every write, another runs tools automatically once you have connected the server. When you evaluate a setup for a store or an agency, look at the host's permission settings as closely as at the server's access level.
In everyday language the host, the client and the assistant are often the same product. The distinction is useful when reading vendor documentation, which tends to use the precise terms, and when deciding where a safeguard such as human approval should sit.