Study guide for the CCA-F tools domain: tool descriptions, schemas, error responses, MCP primitives, transports, security and scoping.
Tool Design & MCP Integration is 18% of the Claude Certified Architect – Foundations blueprint, about 11 questions on a 60-question form. Candidates who learn concepts mainly from reading often find it tricky, because the right answers come from treating tools as interfaces you design for a model — with the same care you would give a public API.
What the domain covers
Descriptions are prompts. Claude chooses tools from their names, descriptions and parameter definitions. Good descriptions explain what the tool does, when to use it instead of a similar tool, what each parameter means with formats and examples, and what comes back. When Claude keeps picking the wrong tool, better descriptions are usually the first fix.
Fewer, task-shaped tools. Every tool definition costs context and adds a decision. Sixty narrow tools that mirror database tables are harder to use than a handful shaped around real tasks. Overlapping tools should be merged, with an explicit parameter for the mode, and outputs should use consistent, labelled units.
Inputs and outputs. Use typed parameters, enums for fixed choices and required flags only where values are truly required. Validate every input again in the handler; the schema guides the model but the handler is the boundary. Keep results bounded with filters, pagination and summaries so a single call cannot flood the context window.
Errors as part of the interface. A useful error tells the agent what happened and what to do next: not found versus system failure, retryable versus permanent, and for invalid input, what was wrong. Opaque codes and empty results lead to blind retries or wrong answers.
MCP fundamentals. The Model Context Protocol connects applications to external capabilities through servers that expose tools (actions), resources (readable data) and prompts (reusable templates). Local servers usually use the stdio transport, launched by the client as a subprocess; shared remote servers use the HTTP-based Streamable HTTP transport and need authentication and per-user authorization.
Security boundaries. Treat tool results as untrusted: fetched pages and documents can contain injected instructions. The structural defences are least privilege (a research agent does not need an email tool), read-only credentials where writes are never needed, and confirmation for actions with external side effects. Descriptions that say "read-only" are not access control.
How questions are framed
Scenarios often describe a developer-productivity setup or a support agent with tool trouble: wrong tool selection, retry loops, oversized results, risky write access. The best answer usually improves the interface (descriptions, consolidation, structured errors) or moves a guarantee into the permission layer.
Study checklist
- Rewrite two overlapping tool descriptions so the choice between them is obvious.
- Design an error format that distinguishes retryable from permanent failures.
- Add pagination and a summary mode to a tool that returns large lists.
- Build a tiny MCP server with one tool and one resource, run it over stdio, and describe what changes for a remote deployment.
- List the permissions each tool in a research agent really needs.
Domain weights and the 60-question form size come from published summaries of Anthropic's exam guide, which is available to registered partners; see our exam guide page for what is confirmed by Anthropic's public FAQ.
Free sample questions: Tools & MCP
Original questions written for practice, not real exam items. Each one appears again, with fresh option order, in the full practice exam.
Sample question 1
Scenario: Customer support resolution agent
Claude keeps choosing the wrong tool between search_orders and get_order. Both have one-line descriptions. What is the most effective fix?
- A. Expand each description: purpose, when to use it over the other, parameters, returns
- B. Merge them into one generic order tool with a free-text "query" parameter
- C. Tell Claude in each user message which of the two tools it should call
- D. Rename the tools to search_orders_v2 and get_order_v2 so they look more distinct
Show answer and explanation
Answer: A. Tool descriptions are the main signal Claude uses to choose, so they should read like documentation for a new colleague, including how to choose between similar tools. Cosmetic renames add no information, a vague merged tool moves the ambiguity into a free-text field, and naming the tool in every message defeats the purpose of tool use.
Sample question 2
Scenario: Developer productivity with Claude
An MCP server exposes 60 narrow tools (one per database table and operation). Claude often picks the wrong one and context usage is high. Which redesign is usually better?
- A. Consolidate into fewer task-shaped tools (e.g. find_customer) with clear parameters
- B. Keep all 60 tools but shorten their descriptions so the full list uses less context
- C. Split them across six MCP servers of ten tools each, all connected at once
- D. Replace them with a single raw SQL tool that can read and write any table
Show answer and explanation
Answer: A. Every tool definition costs context and adds a decision. Fewer tools shaped around real tasks improve selection and reduce tokens. Shorter descriptions make selection worse, a write-capable raw SQL tool trades selection problems for safety problems, and spreading tools over more servers still loads all 60 definitions.
Sample question 3
Scenario: Developer productivity with Claude
Your team needs an MCP server that runs locally next to each developer's Claude Code to access their local files. Which transport is the natural fit?
- A. Streamable HTTP, exposed on the office network so every machine can reach it
- B. stdio, with the client launching the server as a local subprocess
- C. Streamable HTTP on a public URL, protected by a shared API key
- D. A WebSocket server on localhost that each client polls for new results
Show answer and explanation
Answer: B. The stdio transport is designed for local servers: the client starts the server process and communicates over standard input and output, with nothing exposed on a network. HTTP-based transport is for remote, shared servers and needs proper authentication — unnecessary exposure for local file access.
Sample question 4
Scenario: Developer productivity with Claude
A central MCP server provides access to company documentation for hundreds of users. Which transport and design considerations apply? Choose two.
- A. Use the HTTP-based (Streamable HTTP) transport for remote clients
- B. Publish the endpoint publicly so that it is easy for clients to reach
- C. Use stdio, with each user running a local copy holding the admin key
- D. Skip authentication because the documentation is internal-only data
- E. Authenticate users and authorize access per user, not via one admin key
Show answer and explanation
Answer: A and E. Shared remote servers use the HTTP-based transport and must handle authentication and per-user authorization, so each person sees only what they are entitled to. Distributing an admin key to local copies, skipping auth for "internal" data, or exposing the endpoint publicly all widen the attack surface.