Skip to content
Home › Domains

Tool Design & MCP Integration

18% of exam20 concepts

Domain 2 of the Claude Certified Architect – Foundations exam (18% of scored content). Covers how to define custom tools, steer tool calls, connect and scope MCP servers, handle tool errors, and pick the right built-in tool. The tested skills are concrete: exact field placement, regexes, beta headers, transport rules, and error categories.

Quiz this domain

Concepts

  • Tool Definition Schema — required `name`/`description`/`input_schema` plus optional `strict`/`defer_loading`/`input_examples`.
  • Tool Name Regex — `^[a-zA-Z0-9_-]{1,64}$`: letters/digits/`_`/`-`, no spaces, ≤64 chars.
  • Strict Tool Use — `strict: true` on the tool (not `tool_choice`) forces schema-valid inputs; pair with `any`.
  • Input Examples — schema-validated example inputs (20–200 tokens); not for server tools.
  • Effective Tool Descriptions — the #1 performance lever: 3–4 sentences, `action` params, stable IDs.
  • Tool Choice — `auto`/`any`/`tool`/`none`, `disable_parallel_tool_use`, extended-thinking incompatibility, cache busting.
  • MCP Connector — beta header `mcp-client-2025-11-20`, HTTP-only (no STDIO), not on Bedrock/Vertex, not ZDR.
  • MCPToolset Allowlist and Denylist — `default_config` + `configs.enabled` to allow/deny; unknown tools = warning only.
  • Defer Loading and Tool Search — `defer_loading: true` + server-side Tool search for large tool catalogs.
  • Structured Tool Errors — `isError`/`errorCategory` (transient|validation|permission)/`isRetryable`; generic "operation failed" is always wrong.
  • Subagent Error Handling — resolve locally, know when retry won't help, propagate structured context.
  • MCP Config Scopes — `.mcp.json` (project, shared) vs `~/.claude.json` (user).
  • Built-in Tools — Grep (contents) vs Glob (names) vs Read (full file) — pick over a large Read.
  • Tool Namespacing — service (`asana_search`) vs resource (`asana_projects_search`) prefixes; prefix-vs-suffix has non-trivial eval impact; cuts descriptions loaded into context.
  • Poka-Yoke Tool Design — shape the agent-computer interface so mistakes are hard; "describe to a new hire", unambiguous params (`user_id`), consolidation, actionable errors.
  • Agentic Tool Evaluation Loop — realistic multi-tool eval tasks + verifiers → agentic `while`-loops with reasoning/feedback → metrics (calls/tokens/errors) → Claude Code refines tools.
  • Token-Efficient Tool Responses — pagination/truncation/filtering with sensible defaults, `response_format` enum (~65% cut), semantic/0-indexed IDs over UUIDs.
  • MCP Primitives — server: Tools/Resources/Prompts; client: Sampling/Elicitation/Logging; over JSON-RPC 2.0.
  • MCP Architecture — host/client/server roles, one client per server, data vs transport layer, transports STDIO / Streamable HTTP (SSE).
  • Token-Efficient Tool Use — beta `token-efficient-tools-2025-02-19`, built-in on Claude 4+ (legacy 3.7-era savings figures).