Build an AI Agent That Uses MCP

The basic architecture has three layers: the agent runtime manages reasoning and tool selection; the MCP client discovers and calls external capabilities; the MCP server owns the business capability.

User → Agent Runtime → MCP Client → MCP Server → API/DB/Service

Start with one narrow capability

Do not begin by connecting ten servers. Give the agent one MCP server with a small, well-described tool set. Verify tool selection, arguments, errors, approvals, and traces before expanding.

Keep authorization outside the prompt

The server must enforce permissions. The model deciding to call a tool is not proof that the user may perform the action.

Normalize outputs

Prefer structured results with stable IDs. That allows the agent to chain tool calls without parsing prose.