Domain 2 · 18% of the exam
Tool Design & MCP Integration
Design effective tool schemas, implement MCP servers and clients, distribute tools across agents, and integrate external services into Claude-powered applications.
2.1 Tool Interface Design
The tool description — not the tool name — is the primary signal Claude uses to pick a tool, so when the model misroutes among similar tools, fix the descriptions before adding routing layers or few-shot examples.
2.2 Structured Error Responses
MCP tools signal failure with `isError: true`; returning structured metadata — error category, retryable flag, attempted query, and partial results — lets the agent recover intelligently, while a generic "Operation failed" destroys any chance of recovery.
2.3 Tool Distribution & Tool Choice
Give each agent only the tools its role needs (least privilege) — a large tool set degrades selection reliability — and for a frequent-but-simple cross-role need, hand out one narrow scoped tool rather than the whole toolbox.
2.4 MCP Server Integration
Put team-shared MCP servers in a committed `.mcp.json` with `${ENV}` expansion to keep secrets out of the repo, keep personal servers in `~/.claude.json`, reuse community servers for standard integrations, and expose content catalogs as resources.
2.5 Built-in Tools
Reach for Grep to search code content, Glob to match file paths, Edit for a precise change anchored on unique text, and Read + Write as the fallback when Edit can't find a unique anchor.