CCAR-P - Integration (19% of the exam) - Section 3.7

Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent).

Choosing how Claude reaches an external system: an MCP server that exposes tools and resources to any compatible client, a direct API or CLI integration, or an agent-to-agent protocol. Candidates should weigh reuse across clients, maintenance ownership and control against the simplicity of a direct call.

Model Context Protocoldirect API or CLI integrationagent-to-agent communicationreuse versus simplicity

Practice question for this objective

Free sampleIntegrationhard

A regional hospital's pathology department wants a nightly job that pulls the previous day's abnormal results from its laboratory system's documented REST API, has Claude group them by ward with a short clinical note per patient, and writes the summary to the ward managers' shared drive. The steps do not vary, no other application or Claude client will use the integration, and the department has one developer who must also support it. In a two-month prototype the model was given a generic web fetch tool and the API base address, and in 4 of 60 runs it requested the wrong date range. What should the architect recommend?

ABuild an MCP server over the laboratory API with a results tool, so the job and any later clinical assistants can share one maintained integration.
BKeep the fetch tool, add the exact date range rule to the system prompt, and log each request so wrong ranges are caught at the weekly review.
CHave the job's own code call the laboratory API for the fixed date range and pass the results to the model in its request, removing the fetch tool. Correct
DMove retrieval into a separate results agent with its own prompt, and have the summary agent request the data from it through agent-to-agent calls.

For a single fixed integration with known steps and one consumer, a direct API call in code beats exposing the system to the model through a tool or server. When the parameters of a call are known in advance, letting the model choose them only adds a way to be wrong. A direct call in the job's code is deterministic, cheap to maintain and has no reuse case that would justify an MCP server or another agent, so the model is left with the judgement task of grouping and summarising.

Why A is wrong: MCP is the right choice when several clients need the same system, which makes it tempting as future-proofing. The stem states no other client will use it and one developer must support it, so a server adds a component to deploy and maintain, and the model would still choose the date range.

Why B is wrong: A prompt rule and logging are cheap and would likely reduce the error. They are compensating controls: the model still chooses the parameters, so wrong ranges remain possible, and catching them a week later does not stop a wrong summary reaching ward managers.

Why C is correct: The retrieval is a single fixed call with known parameters, so deterministic code should make it. That removes the wrong-date-range failure at its source, needs no new service for one developer to run, and leaves the model only the summarising work it is suited to.

Why D is wrong: Splitting roles can look like good separation of concerns. It adds a second model, and therefore a second chance to pick the wrong parameters, to a step that is a fixed API call, and it doubles what a single developer has to operate.

See more CCAR-P practice questions, answers explained.

Exam traps in Integration

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Publish a shared client library with the records calls and the advisee rule built in, and require each of the four teams to import it into its own integration.
Why it is wrong: A shared library is tempting because it removes the duplicated access logic for code the teams own. It fails the stated constraint: two of the clients are off-the-shelf applications whose code cannot be changed, so they cannot import a library, and each team would still maintain its own tool declarations around it.
Write a direct client against the formulary database for the oncology assistant, avoiding the protocol hop and keeping each check inside two seconds.
Why it is wrong: Going straight to the database is tempting when there is a latency budget. It duplicates an integration the pharmacy team already maintains, so the weekly interaction-rule fixes shipped in the server would not reach oncology, and nothing in the stem shows the existing server misses the budget.
The model tier in use is too small to tell supplier content apart from staff instructions, so a larger tier would have refused it.
Why it is wrong: This is tempting because resistance to prompt injection does vary between model tiers. It is wrong because the model had not changed and no tier reliably refuses every injected instruction; while the write tools remain, any successful injection still has a path to the stock records.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.