MCP review

Server Primitives: Tools, Resources, and Prompts

These three primitives define how different parts of an application interact with the AI model and the user.

**1. Tools**

• Control: Model controlled (Cloud is responsible for deciding when to run them).
• Purpose: To provide the model with external capabilities (e. g. , executing JavaScript, performing calculations).
• Implementation: Implement tools within an MCP server or consume them via an MCP client.
• Function: Generally serve the Model.

**2. Resources**

• Control: Application controlled (Code running inside the app decides when to execute them).
• Purpose: To provide data or content to the application, such as fetching data for the UI or augmenting a prompt with context (e. g. , documents from Google Drive).
• Implementation: Use resources to fetch and process data.
• Function: Generally serve the Application.

**3. Prompts**

• Control: User controlled (The user initiates the action).
• Purpose: To implement predefined workflows or specific actions that a user triggers (e. g. , clicking a UI button, using a slash command).
• Implementation: Put predefined workflows into an MCP server.
• Function: Generally serve the User.

Summary of Control Flow

• Tools → Model
• Resources → Application
• Prompts → User

Takeaways

• Tools are model-controlled and provide the AI model with external capabilities, such as executing code or performing calculations.
• Resources are application-controlled and provide data or content to the application, such as fetching UI data or augmenting prompts with context.
• Prompts are user-controlled and implement predefined workflows or specific actions triggered by the user.
• The control flow dictates that Tools serve the Model, Resources serve the Application, and Prompts serve the User.