Skip to content
Domain 2 · ~20%

Tool Design & MCP Integration

Design effective tools and integrate with Model Context Protocol (MCP) servers. Covers tool description best practices, structured error responses, tool distribution, MCP configuration, and Claude's built-in tools.

d2.1

Tool Description Best Practices

Write clear, effective tool descriptions that help Claude select and use tools correctly. Include input formats, examples, edge cases, and boundary conditions.

What makes a good tool description:

Include input format specifications with examples in the tool description

Specify edge cases and boundary conditions so the model handles them correctly

Clear parameter descriptions with expected types, ranges, and constraints

Tool descriptions act as documentation for the model — more detail is better

Anti-Patterns to Avoid

Vague tool descriptions that leave ambiguity about when or how to use the tool

Missing edge case documentation leading to unexpected tool behavior

d2.2

Structured Error Responses

Design error responses that give the agent enough information to recover or escalate. Use structured fields like isError, errorCategory, and isRetryable.

Error response design patterns:

isError flag: explicitly signals tool failure to the agent

errorCategory: classifies errors (e.g., 'validation', 'auth', 'not_found', 'rate_limit')

isRetryable: tells the agent whether retrying the same call might succeed

Structured error context: include what was attempted and what failed

Anti-Patterns to Avoid

Generic error messages like 'Operation failed' that hide useful context

Silently suppressing errors by returning empty results as success

Not distinguishing between access failures and genuinely empty results

d2.3

Tool Distribution & Selection

Distribute tools across agents effectively. Understand the impact of tool count on selection quality and how to scope tool access.

Tool distribution strategies:

4-5 tools per agent is optimal; too many tools (e.g., 18) degrades selection quality

Scoped tool access: each agent only gets tools relevant to its task

tool_choice options: 'auto' (model decides), 'any' (must use a tool), or forced specific tool

Tool grouping: organize related tools and assign them to specialized agents

Anti-Patterns to Avoid

Giving an agent 18+ tools when only 4-5 are relevant to its task

Not using tool_choice to constrain tool selection when the task is clear

d2.4

MCP Server Configuration

Configure Model Context Protocol servers for project-level and user-level tool integration. Understand .mcp.json vs ~/.claude.json configuration.

MCP configuration essentials:

.mcp.json: project-level MCP server configuration (shared with team)

~/.claude.json: user-level MCP configuration (personal tools)

Environment variable expansion in MCP config for secrets management

MCP servers extend Claude's capabilities with custom tools and data sources

Anti-Patterns to Avoid

Hardcoding secrets in .mcp.json instead of using environment variable expansion

Mixing project-level and user-level configs without understanding precedence

d2.5

Built-in Tools

Understand Claude's built-in tools: Read, Write, Edit, Bash, Grep, and Glob. Know when to use each tool for different tasks.

Built-in tools and their use cases:

Read: read file contents for understanding code or data

Write: create new files from scratch

Edit: modify existing files with targeted changes

Bash: execute shell commands for building, testing, and system operations

Grep: search for patterns across files in a codebase

Glob: find files matching patterns for discovery and navigation

Anti-Patterns to Avoid

Using Write when Edit would be more precise for modifying existing files

Using Bash for file operations when Read/Write/Edit are available

Exam Tips for Domain 2

1.

Keep tools per agent to 4-5 for optimal selection quality

2.

Structured error responses are critical — always include isError, errorCategory, isRetryable

3.

Know the difference between .mcp.json (project) and ~/.claude.json (user)

4.

Built-in tools: know when to use Grep vs Glob vs Read

Related Exam Scenarios

Test Your Knowledge of Tool Design & MCP

Practice with scenario-based questions covering this domain.

Practice Questions