What is an MCP server?
An MCP server is the side that owns tools. The agent connects as a client and calls those tools over MCP.
An MCP server advertises tools, accepts calls, and returns results. Your agent runtime is usually the client. The server can run on your laptop, in your cluster, or as a hosted endpoint.
One server or many
You can run a small MCP server per tool. That is fine for experiments. In production you often want fewer servers, clearer auth, and one place to see usage.
Servers own tools. Clients own the agent loop.
What operators ask
Who can call this tool? Which key was used? How much did it cost? Those answers live next to the server, outside the model weights.