Claude Certified Architect Foundations (CCAR-F): The Complete 2026 Exam Guide

The Claude Certified Architect – Foundations (CCAR-F) is Anthropic's technical certification for solution architects who build production applications with Claude. It is a proctored, 60-question, 120-minute exam that costs $125. You need a scaled 720/1,000 to pass. It covers Claude Code, the Claude Agent SDK, the Claude API and MCP.

I checked every fact in this guide against Anthropic's official CCAR-F exam guide (version 1.0, effective July 2026), the Pearson VUE Anthropic page and Anthropic's Claude Partner Network announcement. Many other pages repeat numbers that conflict with each other. Here you get the official figures, what each domain actually tests, and a study plan you can start today.

Claude Certified Architect exam: key facts at a glance

What is the Claude Certified Architect certification?

The Claude Certified Architect Foundations credential (officially "Claude Certified Architect – Foundations") is the first technical certification in Anthropic's Claude Certification Program. Anthropic launched it on March 12, 2026, at the same time as the Claude Partner Network, which came with an initial $100 million investment in partner training and support.

The official guide says the exam checks that you can make sound decisions about tradeoffs when you build real-world solutions with Claude. It does not test general AI theory. You will not be asked what a transformer is. You will be asked what a given team should do with its CLAUDE.md file, why an agent picks the wrong tool, or where a refund limit should be enforced.

The exam covers four technologies:

• Claude Code: CLAUDE.md files, rules, skills, slash commands, plan mode and CI/CD use

• Claude Agent SDK: agentic loops, subagents, hooks and sessions

• Claude API: tool use, tool_choice, structured output and the Message Batches API

• Model Context Protocol (MCP): tool and resource design, server scoping and error handling

The ideal candidate is a solution architect who designs and ships production applications with Claude. If you build agents, internal tools or AI features for clients, this exam is aimed at you.

Who can take the Claude architect certification?

This catches a lot of people out. Right now, the exam is open to people at organizations in the Claude Partner Network. Pearson VUE says that certification is open to organizations in the network, and that it also counts toward an organization's partner standing.

The good news:

• Partner Network membership is free and open to any organization bringing Claude to market, such as consultancies, agencies, software vendors and systems integrators.

• The preparation courses on Anthropic Academy are free to study, even if you cannot sit the exam yet.

• If you are a freelancer or work at a company that is not a partner, ask your company to apply. You can also start preparing now so you are ready when access opens up.

Tip: if your employer is already a partner, look for the Claude Certified Architect page in the Anthropic Partner Academy. That page has the "Purchase" button and the official exam guide PDF.

Claude Certified Architect exam format and scoring

The exam is built around scenarios. Each scenario describes a realistic production setup, for example a support agent with four MCP tools, and is followed by a set of questions about it. You will see 4 of the 6 scenarios, chosen at random, so you need to prepare for all six.

• 60 items in 120 minutes, which gives you about 2 minutes per question. Candidates who have passed say the time is enough. The difficulty comes from having several answers that look plausible.

• Multiple-choice and multiple-response. Each item tells you how many answers to choose.

• Criterion-referenced scoring. You are measured against a fixed standard, not against other candidates. The cut score of 720 was set through a formal standard-setting study.

• Score report. You get pass or fail, a scaled score from 100 to 1,000, and the percentage you got right in each domain. The domain percentages are for feedback only. Only your total scaled score decides pass or fail.

What makes it hard: often three of the four options would "work." The exam wants the most reliable design. Three principles come up again and again:

1. Deterministic beats probabilistic. If a business rule must always hold, enforce it in code (hooks, prerequisite gates), not only in the prompt.

2. Context is explicit. Subagents do not inherit the coordinator's history, so you have to pass them what they need.

3. Fix the root cause. Better tool descriptions, schemas and configuration scope beat workarounds such as "add more instructions" or "use a bigger context window."

The 5 Claude Certified Architect exam domains (with weights)

Domain 1: Agentic Architecture & Orchestration (27%)

This is the biggest domain and the one most worth your study time. It has seven task statements:

• Agentic loops: continue while stop_reason is "tool_use" and stop on "end_turn". Do not parse the model's text or rely on arbitrary iteration caps to decide when to stop.

• Coordinator-subagent (hub-and-spoke) patterns: all communication goes through the coordinator, and the coordinator chooses subagents dynamically.

• Spawning subagents with the Task tool. The coordinator's allowedTools must include "Task". To run subagents in parallel, emit several Task calls in one response.

• Enforcement and handoffs: programmatic prerequisites (for example, no process_refund before a verified get_customer) and structured escalation summaries for human agents.

• Agent SDK hooks: use PostToolUse to normalize data, and tool-call interception to block actions that break policy.

• Task decomposition: fixed prompt chains for predictable reviews, dynamic decomposition for open-ended investigations.

• Sessions: --resume, fork_session, and knowing when to start fresh with a summary instead of resuming with stale tool results.

Domain 2: Tool Design & MCP Integration (18%)

• Tool descriptions are how the model chooses a tool. Rename or split tools that overlap, and include inputs, examples, edge cases and when not to use each tool.

• Structured MCP errors: set the isError flag and include an error category (transient, validation, business, permission) and whether the error is retryable. A generic "Operation failed" is the wrong answer.

• Tool distribution: give each agent only the 4–5 tools its role needs, not 18. Know the tool_choice options: "auto", "any" and forced selection of a named tool.

• MCP scoping: project-level .mcp.json for shared servers (with ${ENV_VAR} expansion for secrets) and user-level ~/.claude.json for personal ones.

• Built-in tools: Grep searches file contents, Glob matches file paths, and Edit needs unique anchor text.

Domain 3: Claude Code Configuration & Workflows (20%)

• CLAUDE.md hierarchy: user (~/.claude/CLAUDE.md), project and directory levels. The user level is not shared with teammates.

• Modular configuration: @import and topic-specific files in .claude/rules/, with path-specific rules set through YAML glob patterns.

• Slash commands and skills: project scope (.claude/commands/, .claude/skills/) versus user scope, and frontmatter such as context: fork and allowed-tools.

• Plan mode versus direct execution: use plan mode for multi-file or ambiguous changes and direct execution for simple, well-scoped fixes.

• CI/CD: running Claude Code non-interactively for code review, test generation and pull request feedback.

Domain 4: Prompt Engineering & Structured Output (20%)

• Explicit criteria that reduce false positives in reviews.

• Few-shot examples for ambiguous cases and documents in varied formats.

• Structured output through tool use and JSON schemas. Use nullable or optional fields so the model does not make up missing data.

• Validation-retry loops: retry with the specific validation error. Know which failures a retry can fix (format) and which it cannot (the information is missing from the source).

• Batch processing with the Message Batches API, and multi-pass review architectures for large code reviews.

Domain 5: Context Management & Reliability (15%)

• Keeping critical facts across long conversations by extracting structured facts and using scratchpad files.

• Escalation: when to hand off (policy gaps, an explicit customer request, no way to progress) and when to resolve autonomously.

• Error propagation in multi-agent systems: return partial results and say what was attempted.

• Human review workflows with routing based on confidence scores.

• Provenance: keep sources attached to claims, and when sources conflict, report both values instead of silently picking one.

The 6 exam scenarios

Every exam draws 4 of these 6 scenarios. Prepare for each one, because you will not know in advance which ones you get.

Claude Certified Architect exam cost, retakes and policies

The Claude Certified Architect exam costs $125 USD per attempt. The price shown at checkout reflects any discount your partner tier qualifies for.

• Retakes: you must wait 14 days after a first fail, 30 days after a second and 90 days after a third. The limit is 4 attempts per exam in a rolling 12 months, and each attempt costs the full fee.

• Rescheduling: the exam guide says you can cancel or reschedule up to 24 hours before your appointment. Pearson VUE's Anthropic page says 48 hours. Plan on 48 hours to be safe. A late change or a no-show means you lose the fee.

• ID: the name on your Pearson VUE profile must exactly match your government photo ID, or you will not be allowed to test. To correct your name, email certifications-support@anthropic.com before your exam.

• Exam rules: no notes, phones, smartwatches, headphones or second monitors. If you test online, you must stay in view of the webcam. You also accept a non-disclosure agreement, so be wary of any site selling "real exam questions."

• Renewal: the credential lasts 12 months. To renew on time, you complete a free, non-proctored assessment on the Anthropic Partner Academy. If your credential lapses, you have to retake the full exam and pay the full fee again.

• Accommodations: request them from Pearson VUE and wait for approval before you schedule.

How to become a Claude Certified Architect (step by step)

1. Confirm eligibility. Check that your organization is in the Claude Partner Network. If it is not, ask it to apply. Membership is free.

2. Read the official exam guide. Download the CCAR-F exam guide from the Anthropic Partner Academy and read every task statement. This matters more than any other step.

3. Build hands-on experience. Complete the four preparation exercises in the guide: a multi-tool agent with escalation, a team Claude Code setup, an extraction pipeline and a multi-agent research system.

4. Take the practice test. Do it early, review every explanation and focus on your weakest domains. Do not memorize the practice questions.

5. Register and schedule. Buy the exam ($125), create your Pearson VUE account through the confirmation email, and book an online or test-center slot.

6. Pass and share your badge. You need 720 or more. Add the Credly badge to LinkedIn, and set a reminder to renew for free before 12 months are up.

Claude Certified Architect study guide: a 4-week plan

This plan assumes you already build with Claude every week. If you are new to the Agent SDK or MCP, double each week. The plan is ordered by domain weight, so the biggest domain comes first.

Free resources

• Anthropic Academy: free courses on the Claude API, Claude Code, MCP and agents

• Claude documentation: tool use, the Agent SDK, Claude Code settings and the Message Batches API

• Model Context Protocol docs: tools, resources and error handling

• Our Claude Code guide and our MCP explainer

Study tip from people who have passed: if you use Claude to generate practice questions, ask for diagnose-the-failure and choose-the-most-reliable-design questions. Plain recall questions are far easier than the real exam.

Claude Certified Architect sample questions

These five questions are original practice items written for this guide, one per domain. They follow the exam's scenario style but are not real exam questions.

Question 1: Agentic Architecture

Company policy says your support agent must never issue a refund over $500 without a human's approval. In testing, the agent still issues large refunds about 3% of the time, even though the system prompt states the rule. What is the most reliable fix?

1. Repeat the rule in capital letters at the start and end of the system prompt

2. Add a hook that intercepts process_refund calls over $500 and routes them to escalate_to_human

3. Add few-shot examples of the agent refusing large refunds

4. Lower the temperature to make the agent more consistent

Question 2: Tool Design & MCP

Your agent has two tools, analyze_content and analyze_document, with almost identical descriptions. It picks the wrong one about 30% of the time. What should you do first?

1. Set tool_choice to "any"

2. Merge them and add five more general-purpose tools for flexibility

3. Rename and rewrite the tools so that each description states its purpose, inputs, outputs and when to use it instead of the other tool

4. Increase max_tokens so the model can reason more

Question 3: Claude Code Configuration

A new developer's Claude Code sessions ignore the team's testing conventions, while everyone else's follow them. The conventions are in the tech lead's ~/.claude/CLAUDE.md. What is the best fix?

1. Ask the new developer to copy the tech lead's file to their home directory

2. Move the conventions into the project-level CLAUDE.md, or into .claude/rules/, and commit them to the repo

3. Paste the conventions into every prompt

4. Turn on plan mode for the new developer

Question 4: Structured Output

Your invoice extraction pipeline sometimes returns a due_date for invoices that do not have one. What is the best fix?

1. Make due_date nullable in the JSON schema and tell the model to return null when the field is absent

2. Retry the extraction until a valid date appears

3. Add few-shot examples that all include due dates

4. Raise the temperature to get more varied output

Question 5: Context Management & Reliability

In your research system, the web-search subagent times out on one of three subtopics. How should the system respond?

1. Return "Operation failed" and stop the whole pipeline

2. Retry forever until the search succeeds

3. Have the subagent return a structured error (the failure type, the query it tried, its partial results), let the coordinator continue, and note the coverage gap in the report

4. Let the synthesis subagent fill the gap from its own knowledge

Other Claude certifications compared

The Claude Certified Architect – Foundations is one of four exams that Pearson VUE lists for the Anthropic certification program:

Is there a Claude Code certification? No, not a separate one. Claude Code is 20% of the Claude Certified Architect exam, so this is currently the best way to prove your Claude Code skills.

Is the Claude Certified Architect certification worth it?

For people who build with Claude for clients or employers: yes. Here is why.

• It is new and scarce. The exam launched in 2026, so few people hold it yet. Google searches for "Claude Certified Architect" climbed from almost nothing to thousands a month within weeks of launch.

• It counts toward your company's partner standing in the Claude Partner Network, which matters to consultancies and agencies.

• The preparation makes you better at the work. The blueprint reads like a checklist of production failures: wrong tool picks, invented fields, lost context and rules enforced only in the prompt.

• The cost is low. $125 is far below most cloud architect exams, and renewal is free.

Downsides: it is limited to Partner Network organizations, it has to be renewed every year, and it is still young, so employers recognize it less than long-established cloud certifications.

Claude Certified Architect FAQ

Is the Claude certification free?

No. The exam costs $125 USD per attempt. Many preparation courses on Anthropic Academy are free, and renewing an active credential on time is free.

Who can take the Claude Certified Architect exam?

People at organizations in the Claude Partner Network. Membership is free for any organization bringing Claude to market.

How many questions are on the CCAR-F exam?

60 multiple-choice and multiple-response questions in 120 minutes, organized into 4 scenarios drawn from a bank of 6.

What is the passing score?

A scaled score of 720 on a 100–1,000 scale.

How long is the certification valid?

12 months from the date it is awarded. Renew on time with a free, non-proctored assessment.

How long should I study?

Anthropic expects 6+ months of hands-on experience. With that background, 3–6 weeks of focused study is typical. Newcomers should allow 2–3 months.

Is there a Claude Code certification?

Not as a standalone exam. Claude Code Configuration & Workflows makes up 20% of the Claude Certified Architect – Foundations exam.

What happens if I fail?

You can retake the exam after 14, 30 and then 90 days, with up to 4 attempts in a rolling 12 months. Each attempt costs the full fee. Use your per-domain score report to decide what to study next.

Is the exam online or in person?

Both. Pearson VUE offers online proctoring (OnVUE) and test centers.

Is CCA-F the same as CCAR-F?

Yes. The official exam code is CCAR-F. Many sites shorten it to CCA-F.

Final word

The Claude Certified Architect Foundations exam rewards people who have actually shipped Claude systems and learned what breaks. Start with the official exam guide, build the four practice projects, and aim to understand why one design is more reliable than another. That understanding is what the exam tests.