MCP Basics

MCP (Model Context Protocol) is a JSON-RPC based protocol for exposing tools, resources, and prompts to clients. A client does not need to know your internal code structure; it only needs MCP methods and structured metadata.

Think of MCP as a standard contract between:

• A server that provides tools, resources, and prompts.
• A client that discovers and calls those capabilities.

Core ideas

Tools

Tools are functions a client can call with structured arguments. They have:

• A name.
• A description.
• An input schema (arguments).
• An output schema (structured results).

Resources

Resources are static content (typically Markdown) that clients can read. They are identified by resource://<name> URIs and include a media type.

Prompts

Prompts are templated Markdown files with optional input definitions. Clients request a prompt and fill in inputs.

Typical flow

1. Client calls initialize to negotiate protocol and discover capabilities.
2. Client calls tools/list, resources/list, prompts/list.
3. Client calls tools with tools/call.
4. Client reads resources with resources/read.
5. Client fetches prompts with prompts/get.

Example: list tools

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list"
}

Response contains an array of tools with names and schemas.

Example: call a tool

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "user-profile",
    "arguments": { "userId": "123" }
  }
}

Responses return a result payload with both a human-readable content and a structuredContent value when possible.

Why dzx

dzx implements MCP end-to-end. You define tools and content in files, and dzx handles:

• Discovery and metadata.
• Schema inference and validation.
• HTTP + JSON-RPC serving.
• Build output and import manifests.