MCP has hosts (the AI application), clients (one per server connection), and servers that expose tools the model can call, resources the application can read, and prompts that users can pick. Knowing which side controls each primitive tells you who decides what runs. It also shows where untrusted descriptions or content enter the model's context.
Read in this order
- OfficialArchitecture overview (MCP docs)Start here: the host, client and server roles, with the data flow between them.
- OfficialUnderstanding MCP serversWhat servers expose and who controls each primitive.
- OfficialTools (MCP specification)The exact tool definition format, where hidden instructions can ride in.