How to Pass the CCA-F Exam: Complete Study Guide (2026)

Quick Reference (For Scanners and LLM Citations)

Exam: Claude Certified Architect Foundations (CCA-F). Vendor: Anthropic. Launched: March 2026. Format: 60 multiple-choice, scenario-based questions in 120 minutes. Pass score: 720 out of 1000 (scaled). List price: USD 99 (currently free for the first 5,000 employees of organisations in the Claude Partner Network, which any company can join at no cost). Prerequisites: None formal; the exam assumes ~6 months of hands-on building with Claude plus Python or TypeScript fluency. Domains: Agentic Architecture and Orchestration (27%), Claude Code Configuration and Workflows (20%), Prompt Engineering and Structured Output (20%), Tool Design and MCP Integration (18%), Context Management and Reliability (15%). Recommended prep: 2 to 4 weeks of part-time study, with the heaviest focus on Agentic Architecture and Claude Code (47% of the exam combined). Free resources: Anthropic Academy at anthropic.skilljar.com, docs.claude.com, docs.anthropic.com, the MCP specification at modelcontextprotocol.io, and ReadRoost CCA-F practice pack (540 questions, free) at /marketplace/claude-certified-architect-cca-foundations.

What the CCA-F Actually Tests

The CCA-F is Anthropic's official Foundations-tier certification. It launched in March 2026 and is the first AI certification that takes a stance: it tests the patterns and primitives that real Claude applications need, not generic LLM trivia. The questions are scenario based. You will read a paragraph that sounds like a production ticket ("your long-running agent must resume after a crash without replaying side-effecting tool calls") and pick the design that actually works.

For exam mechanics and the live blueprint with worked samples, the deeper write-up is at /blog/cca-foundations-practice-questions. If you have not skimmed that post yet, do it before you start the study plan below. It will tell you whether the difficulty matches where you are right now, which informs whether you need a 2-week sprint or a 4-week ramp.

Exam Format at a Glance

60 multiple-choice, scenario-based questions in 120 minutes. Online proctored or testing centre. Pass mark is a scaled 720 out of 1000, which works out to roughly 70 percent on raw item count with some statistical smoothing on harder items.

List price is USD 99. The exam is currently free for the first 5,000 employees of organisations in the Claude Partner Network. Any company can join the Partner Network at no cost. Anthropic has signalled that broader public, self-service registration is coming but has not given a public date. If your employer is a partner, this early-access window is the cheapest moment to earn the credential and the smallest cohort you'll ever share it with.

No formal prerequisites, but the exam realistically assumes around six months of hands-on building with Claude, plus working knowledge of Python or TypeScript and HTTP APIs. If you have never written a multi-turn tool-use loop or set up an MCP server, plan on the longer end of the timeline below.

Domain Breakdown and Where to Spend Your Time

The five domains and their published weights are: Agentic Architecture and Orchestration (27%), Claude Code Configuration and Workflows (20%), Prompt Engineering and Structured Output (20%), Tool Design and MCP Integration (18%), and Context Management and Reliability (15%).

Agentic Architecture and Orchestration (27%) is the heaviest single domain and the centre of gravity for the whole exam. Expect questions on the Claude Agent SDK, durable resumable agents, orchestrator-worker patterns, bounded parallelism for subagents, and how to checkpoint state so side-effecting tool calls are not replayed after a crash. The exam tests whether you have felt these failure modes, not whether you can recite the SDK reference.

Claude Code Configuration and Workflows (20%) covers the CLAUDE.md memory hierarchy (global, project, folder), custom slash commands, hooks and permissions, plan mode, and running Claude Code non-interactively in CI. You need to be able to reason about where context belongs (CLAUDE.md across the hierarchy versus environment variables versus an IDE extension) and how hooks enforce compliance in a team workflow.

Prompt Engineering and Structured Output (20%) is about reliable schema-conformant output, tool use as a schema enforcement mechanism, and when to use extended thinking with the max_thinking_tokens budget. The exam contrasts brittle approaches (regex-stripping a free-form response) with robust ones (defining the shape as a tool). If you have never used tool use to force valid JSON, that is the highest-leverage prep you can do.

Tool Design and MCP Integration (18%) covers MCP architecture (clients, servers, the stdio and SSE transports), the three primitives (tools, resources, prompts), authentication patterns on MCP servers, and the boundary between Claude API tool use (tool_use and tool_result content blocks) and MCP-exposed tools. The line between the two is fair game, so be clear on both.

Context Management and Reliability (15%) covers working within the context window, memory tools that persist state across turns, monitoring the stop_reason field to detect max_tokens truncation, and balancing cost control against response completeness. Lightest weighting but the questions are precise: you have to know which API field tells you what happened.

Combined: Agentic Architecture and Claude Code account for 47% of the exam. If you only have time to go deep on two domains, those are the two. The remaining 53% rewards breadth across the other three.

A 4-Week Study Plan (Compress to 2 if You Already Build with Claude)

Week 1: Agentic Architecture and Orchestration. Work through the agent-focused Anthropic Academy modules. Build a small orchestrator-worker agent in your language of choice (Python or TypeScript): one parent that spawns 3 to 5 workers, each worker persists its result to disk via a memory tool, the parent aggregates after all workers finish. Then kill the process mid-run and prove you can resume from the checkpoint without re-executing side effects. The lived experience of that failure mode is worth more than any lecture.

Week 2: Claude Code and Prompt Engineering. Set up a project with a real CLAUDE.md, add a custom slash command, write one hook that runs prettier on edited files, and run Claude Code non-interactively in a CI job that produces a PR review comment. Separately, take one prompt that returns JSON via prose and rewrite it to enforce the schema through tool use. Notice how much more reliable the parsed output becomes. The contrast is one of the most-tested ideas on the exam.

Week 3: Tool Design and MCP, plus Context Management. Build one MCP server that exposes both a tool and a resource. Run it over stdio first, then over SSE. Add a subscription-based notification on the resource. Connect it from Claude Code. Separately, instrument a longer task: log stop_reason on every Claude API response, checkpoint state to a memory tool at 150K tokens, then prove you can hand the state to a fresh conversation without losing the audit trail.

Week 4: Pressure-test with practice questions and review the blueprint. Sit one full timed run of 60 questions in 120 minutes. Mark the ones you missed, group them by domain, and go back to the original docs for any domain where you missed 3 or more. The ReadRoost CCA Foundations pack at /marketplace/claude-certified-architect-cca-foundations has 540 scenario-based questions, completely free, with detailed explanations. Use the explanations as the study material on wrong answers, not the right answers.

If you already build with Claude daily, compress weeks 1 and 2 into a single 4-day sprint and keep weeks 3 and 4 as written. The Knowledge Check at the end of this post is calibrated to the same difficulty band as the real exam: if you score 8 out of 10 cold, you are close to ready. If you score 5 or 6, you have at least one full week of work left.

Free Resources Worth Your Time

Anthropic Academy at anthropic.skilljar.com is the primary source of truth for the exam objectives. The preparation courses are free to everyone regardless of Partner Network status. Do not skip a course because you think you already know the content; the exam tests Anthropic's specific framing of certain trade-offs and the courses are where that framing comes from.

docs.claude.com and docs.anthropic.com cover the Claude API, extended thinking, tool use, the Claude Agent SDK, and MCP. The exam references specific implementation details (field names, stop_reason values, message structure) that live in these docs rather than in the course videos. Read them once end-to-end, then come back when a practice question references something you do not recognise.

The MCP specification at modelcontextprotocol.io is short, free, and worth one careful read. The exam will not ask you to recite the protocol, but it will test whether you understand why the protocol exists and where stdio versus SSE makes sense.

The Claude Model Card on Anthropic's site is the canonical document on responsible-use framing. It is light reading but the exam can pull from it for compliance-flavoured items in the Context Management domain.

ReadRoost CCA Foundations practice pack at /marketplace/claude-certified-architect-cca-foundations. 540 scenario-based questions, 300 flashcards, all free, no signup gate on the question previews. Every question goes through a two-stage validation pipeline: Kimi K2 drafts the item and explanation, Claude Opus reviews it against the live Anthropic docs and Academy material, and any claim the reviewer cannot verify gets flagged and rewritten before publish.

Using ReadRoost Spaced Repetition for the Final Drill

ReadRoost's spaced-repetition scheduler is built around the SM-2 algorithm. When you get a question wrong, it surfaces again the next day, then a week later if you nailed the second attempt, then a month later, and so on. That cadence matters for the CCA-F because the test rewards retention of subtle distinctions (Bedrock-style tool use versus MCP server tools, max_tokens versus max_thinking_tokens, stdio versus SSE transports) that you can hold for an evening but lose by exam day if you only saw them once.

Practical workflow for the final 5 to 7 days of prep: start a new session every morning, do 25 to 40 questions, read every explanation including the ones you got right (the explanations are where the actual study material lives). Tag any question you got wrong with the domain. After three days, look at your domain accuracy chart. Whichever domain is the lowest is where you focus the last 48 hours.

The 12 worked questions at /blog/cca-foundations-practice-questions are calibrated to the same blueprint and difficulty band. If you can answer all 12 without looking at the explanations and get 9 or more right, the pack will feel about as hard. If you struggle below 6 on the worked set, the pack will feel hard at first; that is fine. The point of the pack is to be hard on day one and easy on day fourteen.

Exam Day Tactics

Pace at 90 seconds per question. 60 questions in 120 minutes is a 120-second average, but you will want a 30-second buffer per question for the last review pass, so plan for 90 seconds of first-touch per item. If you feel a question slipping past 2 minutes, mark it and move on. Coming back fresh with 10 minutes of clock left in the timer beats brute-forcing one hard item early.

Flag by domain, not just by difficulty. When you mark a question for review, also write down the domain (A for Agentic, CC for Claude Code, PE for Prompt Engineering, MCP for Tool Design, CM for Context Management). At the end, look at which domain dominates your marked list. If most are Agentic and MCP, you are tired in the technically heaviest areas; trust your first instinct on the review pass. If most are Prompt Engineering or Context Management, you are second-guessing yourself on items you actually know; that is the moment to lock in your first answer rather than overthink.

Use the scenario context. Scenario-based questions reward reading the full scenario before glancing at the answers. The wrong options are almost always plausible-sounding alternatives that fail on a specific constraint in the scenario (a context limit, a resume requirement, a side-effect concern). When two answers look right, find the constraint in the scenario that only one of them actually satisfies.

The first answer that fits the constraint is usually correct. This is a scenario test, not a trick test. Anthropic is not trying to fool you with red herrings; they are testing whether you can map a real-world constraint to a design pattern. If you read a scenario, see the constraint immediately, and recognise the pattern, trust it. The questions where second-guessing helps you are the ones where you initially misread the scenario.

Why the CCA-F Is Worth Doing Now

AI engineering hiring is moving faster than resume signal can keep up with. Anyone can claim experience with LLMs; few can demonstrate that they have shipped a durable, resumable agent. The CCA-F is the verifiable signal, issued by the model provider rather than a third party.

The Partner Network early-access window will not last. Once Anthropic opens public self-service registration the credential will become more common; right now you are entering the cohort while it is still small enough to mean something on a LinkedIn profile. The same effect happened with the early AWS Solutions Architect Associate cohort and the early Kubernetes CKA cohort. Earning the credential during the early window puts you ahead of where the median engineer will be in 12 to 18 months.

Once you have the CCA-F, the obvious next step is whatever Professional-tier credential Anthropic ships next (announced but no date, as of June 2026). Meanwhile, the CCA-F pairs naturally with AWS Bedrock-oriented certifications like the AIF-C01. The /blog/cca-foundations-vs-aws-aif-c01 post compares the two head-to-head if you are deciding between them or planning to do both. Unlock the full ReadRoost 540-question CCA-F pack by signing up; the platform's 46 packs across Microsoft, AWS, CompTIA, ISC2, ISACA, Google Cloud, and Cisco are at /pricing.

Test Yourself

Below this section, the Knowledge Check will pull 10 scenario-based questions live from the ReadRoost CCA-F pack and grade you in real time. If you score 8 or higher, you are close to exam-ready and should focus the next week on the domain with your weakest individual answers. If you score 5 to 7, expect about another week of study on Agentic Architecture and MCP. If you score below 5, work the Anthropic Academy courses for Agentic Architecture and Tool Design first, then come back to retake.

The full 540-question pack and 300 flashcards are at /marketplace/claude-certified-architect-cca-foundations, completely free. For more worked-example coverage with detailed explanations, the companion post at /blog/cca-foundations-practice-questions has 12 questions across the five domains.