MCP Server Features

MCP servers expose three primary primitives that define their capabilities and how they interact with AI applications.

1. Tools (Action-Oriented)

Tools are executable functions that an LLM can actively call based on user intent.

• Mechanism: Defined via JSON Schema for input validation.
• Control: The Model decides when and how to call them.
• Examples: search_flights, send_email, query_database.
• Protocol: tools/list and tools/call.

2. Resources (Data-Oriented)

Resources provide read-only data that acts as context for the LLM.

• Mechanism: Accessed via unique URIs (e.g., file:///logs/today.txt).
• Types:
  • • Direct Resources: Fixed URIs for specific data.
    • Resource Templates: Dynamic URIs with parameters (e.g., logs://{date}/errors).
• Control: The Application/Host decides which resources to provide to the model.
• Protocol: resources/list, resources/read, and resources/subscribe.

3. Prompts (Instruction-Oriented)

Prompts are reusable instruction templates that guide the LLM's interaction with tools and resources.

• Mechanism: Parameterized templates (e.g., "Summarize these logs").
• Control: The User explicitly selects and invokes them (often via slash commands).
• Protocol: prompts/list and prompts/get.

Comparative Overview

References

• Source: 00_Raw/mcp/Understanding MCP Servers.md
• mcp-primitives
• mcp-architecture
• mcp-client-features