In a production MCP architecture, where are tool implementations hosted?

In Model Context Protocol (MCP) production architectures, tool implementations reside and execute on dedicated MCP servers. MCP is designed as a client-server protocol where client applications, such as AI agents or host applications, communicate with external MCP servers to discover and invoke tools, context resources, and prompt templates. Hosting tools on an independent MCP server isolates executable logic, external API integrations, and local system access from the client application. This separation allows developers to deploy, scale, and secure tool services independently without modifying the core agent client code or embedding business logic into the language model architecture.

B. Large language model weights store parameter knowledge only and cannot directly execute or host external code logic.
C. Hardcoded tool decorators tightly couple execution logic to local applications rather than leveraging standard MCP server architectures.
D. The MCP client only orchestrates requests and invokes remote tool definitions hosted and processed by the MCP server.

- Model Context Protocol Architecture Documentation, "MCP Architecture: Clients and Servers". https://modelcontextprotocol.io/docs/concepts/architecture

- Anthropic / Model Context Protocol Specification, "Server Concepts: Tools and Resources". https://spec.modelcontextprotocol.io