MCP Tools vs Resources vs Prompts: When to Use Each

MCP servers expose three types of capabilities: tools, resources, and prompts. Developers often use "tool" to mean any of them — but they behave differently because they're controlled by different actors. Getting this right affects how your server integrates with AI clients.

The short answer

The key question for any capability: who should decide when this is used?

Tools — the AI decides

Tools are functions the AI model can call autonomously based on the conversation. When a user says "check the weather in Tokyo," the model decides to call your get_weather tool — the user never explicitly invokes it.

Use a tool when:

• The capability performs an action or has side effects (writing data, calling an API, sending a message)
• The model should be able to discover and use it automatically based on context
• The output feeds back into the conversation

import { z } from "zod";
import { type InferSchema } from "xmcp";

export const schema = {
  to: z.string().email(),
  subject: z.string(),
  body: z.string(),
};

export const metadata = {
  name: "send_email",
  description: "Send an email to an address",
};

export default async function sendEmail({ to, subject, body }: InferSchema<typeof schema>) {
  // send the email
  return { sent: true };
}

Resources — the application decides

Resources are read-only data sources that a client application loads into context. The AI model doesn't invoke resources directly — the host app (Claude Desktop, Cursor, etc.) fetches them and makes their content available.

Think of resources as the "what the AI can read" layer, not the "what the AI can do" layer. They have stable URIs and are fetched by address.

Use a resource when:

• You're exposing static or semi-static data (a file, a config, a knowledge base entry)
• The content should be readable but not modified
• The client application — not the model — decides when to load it

export const metadata = {
  uri: "docs://company/handbook",
  name: "Company Handbook",
  description: "The employee handbook",
  mimeType: "text/markdown",
};

export default async function companyDocs() {
  return "# Company Handbook\n\n...";
}

Prompts — the user decides

Prompts are reusable, parameterized instruction templates that users invoke explicitly — like slash commands in Slack or Claude's / menu. The model doesn't call prompts autonomously; the user picks one and fills in any parameters.

Use a prompt when:

• You want to give users a structured, repeatable way to kick off a task
• The interaction is user-initiated, not model-initiated
• You're standardizing how a common task is described to the model

export const metadata = {
  name: "review_code",
  description: "Review a code snippet for issues",
  arguments: [
    { name: "language", description: "Programming language", required: true },
    { name: "code", description: "The code to review", required: true },
  ],
};

export default function reviewCode({ language, code }: { language: string; code: string }) {
  return `Review this ${language} code for bugs, security issues, and style problems:\n\n\`\`\`${language}\n${code}\n\`\`\``;
}

The mental model

If you're unsure which primitive to use, ask:

1. Should the AI decide when to use this? → Tool
2. Is this data the app should pre-load for context? → Resource
3. Should the user explicitly choose this interaction? → Prompt

Most MCP servers only need tools. Resources and prompts are powerful but narrower — use them when you have a clear use case for the non-model-controlled interaction.

How xmcp maps to this

xmcp follows the file-based convention: each capability lives in its own directory and file. Drop a file, get a registered capability:

src/
  tools/       → tools the AI calls
  resources/   → data the app loads
  prompts/     → templates the user picks

No manual registration. xmcp discovers and registers everything at build time.

Next steps

• What Is an MCP Server? — the full primer on MCP.
• How to Build an MCP Server in TypeScript — scaffold a project with tools, resources, and prompts.
• Core concepts — the xmcp docs for each primitive.