MCP Tools
MCP Tools are executable capabilities that allow the AI model to perform actions or execute computations. Unlike other primitives, the model itself decides when and whether to invoke a tool, making it an agentic, model-controlled capability.
Tools are defined via a JSON schema that explicitly describes input arguments and output format (either text or structured JSON). Effective tool design requires specificity in descriptions, clear declaration of side effects (e.g., write/delete operations), and defined operational limits (e.g., pagination, rate limits) to guide the model’s decision-making. Tools should catch internal exceptions and return structured error responses (isError = true) rather than crashing the server.