HomeExam InfoDomain 02
02

Tool Design & MCP Integration

How to design tool surfaces, schemas, and MCP servers so Claude selects the right action, passes correct arguments, and recovers gracefully from errors.

Section 01

Study Guide

Long-form notes and references.

This study guide provides a comprehensive synthesis of Domain 2 for the Claude Certified Architect — Foundations exam. The core philosophy of this domain is that a Tool serves as a critical interface between the probabilistic world of the model and the deterministic world of code. High-quality agent performance is determined by how well the tool's name, description, schema, output, and security permissions are engineered.

North Star

A tool is the contract between probabilistic reasoning and deterministic infrastructure — name, description, schema, output, and permissions are all behavioral surface area for the model.

Glossary of Key Terms

Allow-list
A security model where only specifically permitted actions or entities are allowed; more secure than a deny-list.
Async Polling Pattern
A design where a tool returns a job_id for long-running tasks, and the agent polls for completion via a status-check tool.
Capability-Based Security
A security philosophy where access is restricted at the infrastructure / capability level rather than through prompts.
Data Minimization
The practice of returning only the specific fields and data points necessary for the model to complete its current task.
Dynamic Provisioning
Selecting and presenting a subset of available tools to the model based on the specific context of the request.
Idempotency Key
A unique identifier ensuring an operation (like a transaction) is executed only once, even if the request is retried.
is_error: true
An official MCP flag used to signify that a tool execution has failed in a way the model can structurally recognize.
MCP (Model Context Protocol)
Anthropic's protocol for defining tools and resources as external services for AI models.
POLA (Principle of Least Authority)
A security principle stating a tool should have only the minimum permissions necessary to perform its function.
Schema Constraints
Parameters within a JSON schema (like maxLength or enum) that enforce the shape and type of data the model can provide.
Tool Shadowing
Confusion caused when the model cannot distinguish between tools due to overlapping names or vague descriptions.
Top-K Ranking
A technique of returning only the most relevant K results to maintain context efficiency.
Typed Envelope
A standardized structure for tool outputs that includes status flags, content types, and error categories.