Concepts

• Resources
• Prompts
• Tools

Resources

Application-controlled, but Some implementations may even allow the AI model itself to determine which resources to use

Cline Client

Parsing message from LLM, you can get AssistantMessageContent. AssistantMessageContent is Text or ToolUse.

If the ToolUse is accept_mcp_resource, it call readResource.

Then the server must implement readResouce function.

LLM knows which resources they can use by creating prompt dynamically.

Tools

Model-controlled: AI model determines which tools to use.

Anthropic SDK has the interface to receive tools. code

LLM returns message saying which tools they want to use. Client should handle it. For example, Anthropic SDK returns response.content.type with "tool_use"