#Primitives

The most important part of MCP is its primitives: the kinds of context a server can offer a client. Officially there are three, split by who decides to use them.

If you know HTTP: a resource is like GET (load data; do not change the world). A tool is like POST (do work; may have side effects). A prompt is closer to a saved query the user runs by name.

#Tools: actions the model calls

Each tool has a name, a description, and a JSON Schema for arguments. The model decides when to call; the host should still approve writes and other risky work.

A conceptual list request (the current spec also sends _meta on every call; only the method matters here):

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/list",
  "params": {}
}

A list result might include an adder:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "tools": [
      {
        "name": "add",
        "description": "Add two numbers.",
        "inputSchema": {
          "type": "object",
          "properties": {
            "a": { "type": "integer" },
            "b": { "type": "integer" }
          },
          "required": ["a", "b"]
        }
      }
    ]
  }
}

A call:

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "add",
    "arguments": { "a": 1, "b": 2 }
  }
}

With the official Python SDK you do not write this JSON by hand: type hints become the schema and the docstring becomes the description. See Python Server.

Design notes: one job per tool; stable names (weather_current beats a vague get); mention side effects in the description so the host can ask the user first.

#Resources: data the app loads

A resource has a URI and a MIME type. The host chooses what to read and how much to pass to the model.

Two shapes:

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "resources/read",
  "params": { "uri": "greeting://Ada" }
}

The body is the resource content (text or encoded bytes). It is not a guarantee the model “already read it”—the host may attach only a summary.

Do not expose secrets, full production connection strings, or raw personal data as resources. Resources are often pasted wholesale into context and leak more easily than a tool return value.

#Prompts: templates the user invokes

A prompt is triggered from a slash command, palette, or button—not silently by the model. prompts/get fills arguments; the server returns one or more messages; the host sends them to the model.

{
  "jsonrpc": "2.0",
  "id": 5,
  "method": "prompts/get",
  "params": {
    "name": "summarize",
    "arguments": { "text": "MCP is USB-C for AI." }
  }
}

Good uses: a domain workflow (“review this PR against our checklist”) or a demo of how to combine this server’s tools and resources. Poor use: stuffing an employee handbook into a template—that belongs in a resource.

#Discovery sequence (mental model)

sequenceDiagram
    participant C as Client
    participant S as Server
    C->>S: server/discover
    S-->>C: versions / capabilities / identity
    C->>S: tools/list
    S-->>C: tool schemas
    C->>S: tools/call
    S-->>C: result

server/discover is optional but returns capabilities and cache hints (ttlMs, cacheScope) in one shot. List methods may paginate with cursor. The SDK and Inspector walk this path for you.

#Deprecated: Sampling (awareness only)

Older revisions let a server call sampling/createMessage and ask the host to run the model. As of 2026-07-28, Sampling (and Roots / Logging as core features) is deprecated, with at least a twelve-month window. Do not implement Sampling in new tutorials or new projects. If you need a model, call a provider API from your app or host. The rest of this course does not teach Sampling.

Clients may still support elicitation (ask the user to confirm or fill a gap). The current spec expresses that with Multi Round-Trip Requests, not a permanently open bidirectional stream.

#Next

• Transports
• Quick Start
• Python Server