Domain 2 · 18% of the exam · ~11 of 60 questions
Tool Design & MCP Integration
Tool interface design, structured error responses, tool distribution and tool_choice, MCP server integration, and built-in tool selection.
What this domain covers
The lightest domain by weight, and the one candidates most often underestimate. Its central claim is that a tool description is not documentation — it is the mechanism by which the model selects a tool, so a vague description is a functional bug rather than a cosmetic one. The rest follows from that: when to split a tool versus consolidate it, when an MCP resource beats an MCP tool, and what an error should return so the agent can recover locally instead of escalating.
What the exam asks here
Drawn from the in-scope topic list in the official exam guide.
- Tool interface design: writing descriptions that disambiguate, naming to reduce collisions, and knowing when to split versus consolidate
- MCP tool and resource design: resources for content catalogues, tools for actions, and why description quality drives adoption
- MCP server configuration: project versus user scope, environment variable expansion, and multiple servers in use at once
- Structured error responses: distinguishing transient, business and permission errors, and recovering locally before escalating
- Tool distribution across agents and configuring tool_choice, plus when a built-in tool beats a custom one
The 5 task statements in this domain
- 2.1Tool Interface DesignDesign effective tool interfaces with clear descriptions and boundaries
- 2.2Structured Error ResponsesImplement structured error responses for MCP tools
- 2.3Tool Distribution & Tool ChoiceDistribute tools appropriately across agents and configure tool choice
- 2.4MCP Server IntegrationIntegrate MCP servers into Claude Code and agent workflows
- 2.5Built-in ToolsSelect and apply built-in tools (Read, Write, Edit, Bash, Grep, Glob) effectively
Common questions about Domain 2
How much of the CCAR-F exam is Tool Design & MCP Integration?
18% — the lightest of the five domains, or about 11 of the 60 items. It is still worth full attention because MCP concepts reappear inside the Claude Code and agentic architecture questions.
What is the Model Context Protocol, for exam purposes?
MCP is the open protocol Claude uses to reach external systems through a server that exposes tools and resources. For the exam the important distinction is what belongs in each: resources for content the model may want to read, such as a catalogue or a document set, and tools for actions that change something or fetch something dynamically.
Why does the tool description matter so much?
Because the description is the mechanism the model uses to select a tool — not supplementary documentation. Two tools with overlapping descriptions produce wrong selections no amount of prompting fixes. Exam questions typically present an agent calling the wrong tool and expect you to fix the interface rather than the prompt.
Should a tool return an error or throw?
Return a structured error the agent can reason about, distinguishing transient failures worth retrying from business-rule rejections that need a different approach and permission errors that need escalation. An unstructured failure gives the model nothing to recover from, so it escalates everything or retries blindly.