Model Context Protocol (MCP)
Also: MCP · model context protocol
An open protocol that standardizes how apps expose tools, data, and prompts to LLMs over a client/server link.
Why CCAR-P tests this
It is the backbone of P3 integration and candidates are expected to know its architecture and its security model cold.
MCP standardizes the connection between an LLM application and external capabilities. The host (the AI app, e.g. Claude Desktop or an agent runtime) embeds one or more clients, and each client maintains a 1:1 connection to an MCP server. The server exposes three primitives: tools (model-invocable functions with side effects), resources (readable data the host attaches as context), and prompts (reusable templated workflows the user triggers).
The critical routing fact: the model decides which tool to call almost entirely from the tool's `description` and input schema, not from its name alone. A vague or misleading description causes wrong or missed tool calls, so descriptions are effectively prompt-engineering surface.
Servers should follow least-privilege scoping — expose only the tools a given task needs, with narrow parameters. Remote MCP servers authenticate via OAuth / delegated tokens tied to the acting user, never a shared static secret embedded in the server. Transport is JSON-RPC 2.0 over stdio (local) or streamable HTTP/SSE (remote).
Exam trap
Tools are model-invoked while resources and prompts are app/user-controlled — treating all three as things the model freely calls is wrong, and assuming the model routes on tool name rather than description is the classic miss.
Commonly confused with
Related concepts
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.