Anthropic created the Claude Certified Architect (CCA) credential to distinguish engineers and architects who have demonstrated real competence with Claude's architecture — not just familiarity with prompting, but a working knowledge of the API, MCP protocol, Claude Code, agentic systems, and the security and governance patterns that enterprise deployments require. The credential is proctored through Anthropic Academy on the Skilljar platform, and passing it signals to employers and clients that you understand Claude at the depth required to design and lead production deployments.
This guide covers what the exam tests, how it is structured, how to study for each domain, and the specific patterns and concepts you need to internalise before exam day. We also include sample questions with explanations. If you want guided preparation rather than self-study, our CCA Certification Prep service covers the full exam curriculum in structured workshop format.
📋 Exam Fast Facts (2026)
Format: 60 multiple-choice and multi-select questions • Duration: 120 minutes • Passing score: approximately 72% • Delivery: proctored online via Anthropic Academy / Skilljar • Retake policy: 14-day waiting period • Pre-requisites: none formally required, but 6+ months of Claude development experience is strongly recommended
Why the CCA Credential Matters in 2026
Anthropic invested $100 million in the Claude Partner Network in early 2026 because they recognise that enterprise Claude adoption requires qualified human expertise — not just API access. As Deloitte deploys Claude across 470,000 associates and Accenture trains 30,000 professionals on Claude, organisations are competing to hire and retain people who can architect these deployments correctly. The CCA is Anthropic's mechanism for validating that expertise.
From a hiring standpoint, the CCA is becoming a differentiator in the same way AWS Solutions Architect certification differentiates cloud engineers. It is not a guarantee of competence, but it is a credible signal that someone has internalised the technical depth required to work at the architecture level — token budgets, context management, MCP design, agentic safety, and enterprise security controls. If you are building a Claude consulting practice or seeking to lead Claude implementation projects, the CCA is the credential you need.
For our consulting team at ClaudeImplementation, requiring CCA for all architects is a quality signal we put in front of clients. It means every architect on your engagement has passed a rigorous, Anthropic-validated exam on the same technology they are deploying for you.
Exam Structure and Format
The CCA exam contains 60 questions to be completed in 120 minutes — two minutes per question on average. Questions are a mix of multiple-choice (select one correct answer) and multi-select (select all that apply). Multi-select questions are more demanding: partial credit is not awarded, so you must identify every correct answer and no incorrect ones to receive the point.
The exam is proctored, meaning a proctor monitors your session via webcam and screen share. You cannot use reference materials, documentation, or AI tools during the exam. This is deliberate — Anthropic wants to test internalised knowledge, not the ability to look things up. The exam is delivered through Anthropic Academy (the Skilljar platform), and you register and pay through the Academy portal.
Scoring is straightforward: each question is worth one point, and the passing threshold is approximately 72% (43 of 60 questions). Anthropic reserves the right to adjust the passing score between exam versions, so confirm the current threshold at time of registration. If you fail, you must wait 14 days before retaking. There is no limit on the number of attempts.
The 5 Exam Domains: What Each Covers
The CCA exam is organised into five domains. Anthropic has not published exact question weightings per domain, but based on the curriculum structure and exam experience, the domains carry approximately equal weight — roughly 12 questions each. Do not over-index on any single domain. A weakness in any one of the five can sink your score below passing.
Claude API & Application Architecture
API fundamentals, model selection, token management, context windows, system prompts, streaming, batching, prompt caching, and production API patterns.
~20% of examModel Context Protocol (MCP)
MCP specification, server types (local vs remote), tools/resources/prompts primitives, transport security, authentication, and integration patterns.
~20% of examClaude Code Architecture
CLAUDE.md configuration, skills, hooks, sub-agents, GitHub integration, enterprise deployment, permission models, and CI/CD pipeline integration.
~20% of examAgentic AI Systems
Agent SDK, multi-agent orchestration, tool use patterns, agentic safety controls, human-in-the-loop design, state management, and evaluation frameworks.
~20% of examEnterprise Security, Governance & Deployment
Claude Enterprise features, SSO/SCIM, data residency, audit logging, compliance requirements, Claude Cowork administration, change management, and the Claude Partner Network programme.
~20% of examDomain 1: Claude API & Application Architecture
This domain tests whether you understand Claude's API at an architectural level — not just how to make API calls, but how to design production systems around them. Key topics include model selection (Opus for complex reasoning, Sonnet for balanced performance, Haiku for high-throughput/low-cost), context window management, and the economic and performance tradeoffs between streaming and batching.
Prompt caching is heavily tested
Prompt caching is a significant cost reduction mechanism — reducing API costs by up to 90% for repeated context — and the exam tests your knowledge of when it applies, how it is implemented, and what the minimum token thresholds are for the cache to activate. Know that prompt caching requires at least 1,024 tokens in the cacheable prefix, that cached tokens cost less than uncached input tokens, and that cache lifetime is configurable. Our prompt caching guide covers this in depth.
Extended thinking
Extended thinking — Anthropic's mechanism for Claude to reason through complex problems with a visible chain of thought — is tested in the context of when to use it (complex multi-step reasoning, mathematics, code debugging) and what it costs (thinking tokens count toward your usage, and extended thinking requires Opus or Sonnet, not Haiku). Know the thinking parameter in the API and how budget tokens work.
Token and context management
Understand context window sizes for each model tier, how to calculate token counts for cost estimation, and strategies for managing long contexts (RAG, summarisation, sliding window approaches). The exam may present scenarios where you need to recommend the right context management strategy for a given use case. Our detailed guide on the Claude API for enterprise covers these patterns.
📝 Sample Domain 1 Question
A financial services firm runs an AI assistant that uses the same 50,000-token legal policy document in every conversation. Which approach best minimises API costs at scale? (A) Summarise the document to 2,000 tokens before each request (B) Use prompt caching to cache the document across requests (C) Use the Batch API to process requests asynchronously (D) Split requests across multiple Haiku calls instead of Sonnet — Correct answer: B. Prompt caching is specifically designed for this pattern and reduces cached input token costs significantly.
Domain 2: Model Context Protocol (MCP)
Domain 2 is MCP-heavy and tests knowledge of the specification at a working level. You need to understand the three primitive types (tools, resources, prompts), the difference between local stdio and remote HTTP/SSE transports, authentication patterns (OAuth 2.1, API keys), and the security implications of different MCP configurations. This is not a "what is MCP" question set — it is architecture and implementation depth.
Tools, resources, and prompts
Be precise about the distinction. Tools are functions Claude can call (side effects, actions). Resources are data Claude can read via URI patterns (no side effects). Prompts are templated interaction patterns that users or applications can invoke. The exam tests whether you can identify which primitive type to use for a given scenario — for example, a database query that returns data and has no side effects is a resource access, not a tool call.
Transport security and authentication
Know the security requirements for each transport type. Local stdio servers run as OS processes — security is about binary integrity and file system ACLs, not network security. Remote servers require TLS and authentication. OAuth 2.1 with PKCE is the pattern for user-delegated access. Service account credentials with a secrets manager are the pattern for system-level access. Our dedicated MCP security guide is required reading for this domain.
MCP vs direct tool use
The exam will test your ability to choose between MCP servers and direct tool use (function calling in the API request). MCP is for reusable, multi-surface integrations. Direct tool use is for application-specific, tightly controlled flows. Know the tradeoffs covered in our MCP vs Custom API guide.
Domain 3: Claude Code Architecture
Domain 3 covers Claude Code at the enterprise architecture level. You need to understand how CLAUDE.md files configure behaviour at different scope levels, how skills and hooks extend Claude Code's capabilities, how sub-agents manage complex multi-step tasks, and how to deploy Claude Code safely across an engineering organisation.
CLAUDE.md configuration hierarchy
CLAUDE.md files can exist at the project level, directory level, and user level, each inheriting and overriding settings from higher levels. The exam tests your understanding of this hierarchy — which settings take precedence, how to set organisation-wide defaults while allowing team-level customisation, and what should be in a CLAUDE.md for a production codebase. See our CLAUDE.md configuration guide for full coverage.
Skills, hooks, and sub-agents
Skills are reusable instruction bundles that extend Claude Code's capabilities for specific workflows. Hooks are event-driven triggers that run before or after Claude Code actions — pre-commit hooks, post-edit hooks, and so on. Sub-agents allow Claude to delegate portions of a complex task to specialised sub-processes, managing context independently. The exam tests your ability to choose the right extensibility mechanism for a given scenario.
GitHub integration and CI/CD
Claude Code's GitHub integration — PR reviews, issue triage, CI/CD pipeline integration — is tested at the configuration level. Know how to set up Claude Code as a GitHub Action, what permissions it requires, and how to scope its access to prevent unintended repository access. Our Claude Code GitHub integration guide covers the setup patterns in detail.
Domain 4: Agentic AI Systems
Domain 4 is where the exam gets demanding. Agentic architecture — designing systems where Claude acts autonomously across multiple steps, using tools, calling sub-agents, and making decisions without continuous human oversight — requires understanding both the technical patterns and the safety implications. This domain tests both.
The Claude Agent SDK
The Agent SDK provides the primitives for building production Claude agents: agent lifecycle management, tool registration, memory and context management, and multi-agent orchestration. Know the key components: the Agent class, tool definitions, memory backends (ephemeral vs persistent), and how agents communicate in a multi-agent system. Our Agent SDK guide is the primary study resource for this topic.
Agentic safety and human-in-the-loop
This is tested at an architectural level. When should an agent pause and request human confirmation? How do you design confirmation gates that are meaningful (not just rubber stamps) without creating friction that makes the agent useless? The exam uses scenario-based questions here: given a specific agentic workflow, identify the correct points for human confirmation. High-impact, irreversible actions always require confirmation. Reversible, bounded actions can proceed autonomously.
Multi-agent patterns
Orchestrator-worker patterns, parallel agent execution, agent-to-agent communication, and task decomposition strategies are all testable. Know when to use a single agent vs multiple specialised agents, and how to manage state across a multi-agent system. Our multi-agent systems guide covers these patterns with worked examples.
Want Structured CCA Exam Preparation?
Our CCA prep programme covers all 5 domains in guided workshops, with practice exams, architecture exercises, and our internal exam strategy guides. Our pass rate for prepared candidates is 94%.
See CCA Prep Programme → Book a CallDomain 5: Enterprise Security, Governance & Deployment
Domain 5 covers the enterprise deployment context — Claude Enterprise features, security controls, governance policies, and the operational requirements for running Claude at scale in a regulated environment. This domain is often underestimated by candidates who focus on the technical domains and assume the governance material is easy. It is not.
Claude Enterprise features
Know what is included in Claude Enterprise that is not in Team or Pro: SSO (SAML 2.0, OIDC), SCIM provisioning for user lifecycle management, advanced audit logging, data residency options (AWS EU/US), the ability to create system prompt templates, and the admin console capabilities. The exam tests the specific controls available to enterprise admins — what they can configure, what they cannot, and why.
Data residency and privacy
Claude Enterprise's privacy guarantees (no training on customer conversations), data residency options, and the handling of system prompts vs conversation content are all testable. Know the difference between what is protected by the Enterprise agreement and what requires additional contractual arrangements (HIPAA BAA, GDPR DPA). Understand how MCP servers sit outside the Claude API's privacy controls and require separate data governance documentation.
Claude Cowork administration
The enterprise admin controls for Claude Cowork — connector management, Dispatch permissions, usage limits, audit logging, SSO integration — are tested in this domain. Know how Cowork Enterprise differs from Cowork Pro/Max, what controls are available to organisations vs individual users, and how to design a Cowork rollout that maintains security while enabling productivity. Our Cowork deployment service covers these admin patterns in detail.
8-Week Study Plan for the CCA Exam
Based on coaching candidates through the CCA exam, the following eight-week plan balances breadth across all domains with depth in the areas that trip most candidates. Adjust based on your existing experience — if you have two years of Claude API development experience, compress Domain 1 and spend the saved time on agentic architecture (Domain 4), which is where most senior engineers underperform.
Week 1–2: API Architecture & Prompt Caching
Complete Anthropic Academy's API fundamentals course. Build a sample application using prompt caching and extended thinking. Study token economics and context management patterns. Target: full Domain 1 coverage.
Week 3: MCP Protocol Deep Dive
Read the full MCP specification. Build a local MCP server in Python (use the tutorial at Anthropic Academy). Study authentication patterns and security requirements. Read our MCP enterprise and security guides. Target: Domain 2 coverage.
Week 4: Claude Code Architecture
Set up Claude Code and build a real CLAUDE.md for an existing project. Create a skill and a hook. Study the sub-agent documentation and GitHub integration setup. Target: Domain 3 coverage.
Week 5–6: Agentic Systems (most challenging domain)
Study the Agent SDK documentation thoroughly. Build a multi-agent system with orchestrator and worker agents. Understand agentic safety patterns and human-in-the-loop design. This domain requires hands-on experience to internalise. Target: Domain 4 coverage.
Week 7: Enterprise Security & Governance
Study Claude Enterprise documentation, especially SSO, audit logging, and data residency. Read Anthropic's security and privacy documentation. Review the Claude Partner Network programme structure. Target: Domain 5 coverage.
Week 8: Practice Exams and Weak Area Review
Complete at least two full-length practice exams under timed conditions. Identify weak domains and review specific topics. Focus on multi-select questions — they are the most common failure mode. Review scenario-based questions in agentic safety and enterprise governance.
Sample Exam Questions with Explanations
These representative questions illustrate the style and depth of the actual exam. They are not taken from the exam itself — they are constructed to test the same concepts at similar depth. For a full set of 50 practice questions, see our companion article CCA Practice Questions.
Q1. An enterprise is building a Claude agent that processes customer refund requests. The agent reads the order database, checks refund eligibility, and can initiate refund transactions. Which design decision best aligns with agentic safety best practices?
Q2. Which of the following is a valid use case for MCP resources (as opposed to MCP tools)? Select all that apply.
Q3. An engineering organisation wants to deploy Claude Code to 200 developers while preventing Claude from pushing code changes directly to the main branch. What is the correct configuration approach?
Why Candidates Fail (and How to Avoid It)
Having coached dozens of candidates through the CCA, we see the same failure patterns repeatedly. Understanding them before you sit the exam is half the battle.
Underestimating the agentic domain
Senior API developers who ace Domains 1 and 2 often underperform in Domain 4 because agentic architecture requires a different mental model — you are not designing an API call, you are designing a system with autonomous decision-making, state, memory, and safety boundaries. The exam tests architecture instincts here, not memorised facts. Build an actual multi-agent system before the exam. Reading about it is not enough.
Confusing MCP primitives
Tools, resources, and prompts are conceptually simple but easy to mix up under exam pressure. Write out the distinction until it is automatic. A tool has side effects and executes actions. A resource is data Claude reads with no side effects. A prompt is a templated interaction pattern. Every scenario question about which MCP primitive to use reduces to these definitions.
Ignoring the governance domain
Candidates with deep technical backgrounds often deprioritise Domain 5 as "soft" material. The exam does not treat it that way. Questions about SSO configuration options, data residency controls, and audit log requirements are specific, detail-oriented, and require you to have actually read the Claude Enterprise documentation rather than assuming you know the answers from general cloud experience.
Multi-select question strategy
Multi-select questions (select all that apply) are partial-credit-free — you must get every option right. Under time pressure, candidates rush these and miss a correct option or include an incorrect one. Budget extra time for multi-select questions: read every option carefully, eliminate clearly wrong answers, and do not assume a two-option answer when the question might have three correct choices.
🎯 CCA Exam Preparation Checklist
- Complete all relevant Anthropic Academy courses (13 free courses on Skilljar)
- Build hands-on projects in each domain — reading alone is insufficient
- Study all 5 domains equally — do not assume your strongest domain compensates for a weak one
- Internalise the MCP primitive distinctions (tools vs resources vs prompts)
- Build at least one multi-agent system before the exam
- Read Claude Enterprise documentation thoroughly — especially SSO, audit logging, and data residency
- Practice under timed conditions — 120 minutes for 60 questions is tight with multi-select questions
- Budget additional time per multi-select question — all-or-nothing scoring makes these high-stakes
Official and Recommended Study Resources
The official study path runs through Anthropic Academy (academy.anthropic.com), which offers 13 free courses on Skilljar covering the API, MCP, Claude Code, and agentic systems. These courses are the primary source material for the exam. Completing them is necessary but not sufficient — the exam tests application and architecture judgement, not just recall of course content.
Supplementing the Academy courses with hands-on project work is essential. Our recommended approach: complete each Academy course module, then immediately build a small project that exercises the concepts from that module. For the MCP module, build an MCP server. For the agentic module, build a two-agent system with an orchestrator. For the Claude Code module, set up Claude Code on a real codebase with a full CLAUDE.md configuration.
The ClaudeImplementation blog covers each exam domain in depth. Relevant articles: Claude API for Enterprise (Domain 1), MCP Enterprise Guide (Domain 2), Claude Code for Enterprise (Domain 3), Enterprise AI Agent Architecture (Domain 4), and Claude Security & Governance (Domain 5). For full structured preparation, our CCA Certification Prep service includes guided instruction, practice exams, and access to our exam strategy materials.
CCA Exam FAQ
Is there a prerequisite for the CCA exam?
There is no formal prerequisite. Anthropic does not require a particular work history or prior certification. However, the exam is designed for architects and engineers with substantial Claude implementation experience — Anthropic Academy recommends 6 months minimum of active Claude development. Candidates without hands-on MCP, Claude Code, and agentic systems experience will find the exam extremely difficult to pass through study alone.
How much does the CCA exam cost?
Exam pricing is set by Anthropic Academy and may change. Check the Anthropic Academy portal for current pricing. As of early 2026, the exam is competitively priced relative to major cloud provider certifications (AWS, Azure, GCP). Volume pricing is available for organisations registering multiple candidates.
Is the CCA exam updated when Claude features change?
Anthropic updates the exam curriculum periodically to reflect significant product changes. The exam version you sit is tied to a specific curriculum version. When you register, confirm you are studying the materials for the current exam version. If Anthropic releases a major update (new model generation, significant MCP spec changes, new Claude Code features), expect the exam to reflect those changes within 2-3 months.
What is the passing score?
The passing threshold is approximately 72% — roughly 43 of 60 questions. Anthropic reserves the right to adjust the passing score between exam versions based on psychometric analysis. The specific threshold is confirmed when you register. Do not plan to pass at 72% — target 80% or higher in your practice exams to give yourself a comfortable buffer on exam day.
Can I use notes during the exam?
No. The exam is proctored and you may not use reference materials, documentation, or AI tools during the session. The proctor monitors via webcam and screen share. Any violation of exam policy results in immediate disqualification and a waiting period before re-registration. The exam is designed to test internalised knowledge, not the ability to look things up.
Is our CCA Certification Prep service useful if I have significant Claude experience?
Yes. Our prep programme is calibrated for experienced engineers, not beginners. We focus on the architecture and design judgement the exam tests rather than the API basics you already know. Senior developers most often fail on the agentic domain and the governance domain — our programme targets those specifically. We also provide two full-length practice exams that mirror the actual exam's difficulty and format, which is the single most effective preparation tool available.