MCP server

Your exam prep, inside the assistant you already use

Connect one URL and your assistant quizzes you on the real blueprint — one question at a time, withholding the answer key until you commit, then explaining why every option wins or loses.

Connect it in one minute ↓ · or try the 236 free questions in the browser first — no account either way.

What it looks like

Nothing here is a mock-up. Every frame is the assistant calling this site’s tools over MCP, and the question is one of the real ones — a governance scenario where two of the four options are defensible and only one is credited.

A terminal session: the user types 'quiz me on CCAR-P governance', the assistant calls get_practice_question, and a four-option scenario question appears with no answer marked. — 1 / 3You ask for a topic. The assistant calls `get_practice_question` and reads back the scenario. **It has not been told which option is correct**— that is the whole design, and everything else follows from it.

The user commits to option A. The assistant calls check_answer, which reports 'Not correct. Credited answer: C' and marks A as the chosen option. — 2 / 3You commit first. Only then does `check_answer` run, and only then does the key exist in the conversation at all. Choosing before you are graded is the part that does the teaching.

The assistant explains why the chosen option loses and why the credited one wins, contrasting prompt instructions with enforcement at the data layer. — 3 / 3Then the part worth paying for: not that C wins, but that a system-prompt instruction is probabilistic and stays bypassable, while authorisation enforced at the data layer means the record never enters the context in the first place. Both options were plausible. Knowing why the tempting one loses is what transfers to the exam.

Where it stops being a quiz

One question at a time is a party trick. The reason to connect a server rather than paste a question is that the assistant can read your record and decide what you should do next — and it can weigh a domain by what the blueprint says it is worth.

The user asks 'am I ready to sit the CCAR-P?'. The assistant calls my_progress and returns a per-domain breakdown, weakest first, with each domain's blueprint weight. — 1 / 3`my_progress`returns the record by blueprint domain, weakest first, each with the weight it carries on the real paper. Not one number — the per-domain breakdown is the entire point.

The assistant reasons about the score: the gap is concentrated, Integration is the biggest lever at 19 percent of the paper, and two domains are already done. — 2 / 3Now it can reason. Integration is the biggest lever because it is the largest slice *and* the worst score; Governance is worse per question but worth less; two domains are finished and should be dropped. That is a study plan, and it came from the data rather than from encouragement.

The assistant builds a twenty-question drill weighted by domain value and serves the first Integration question with a progress bar. — 3 / 3And then it runs the plan — a drill weighted by what each domain is worth, keeping score as it goes. Same tools, same refusal to show you the key.

Which URL

/api/mcpFree · no account

The free form for each exam — 236 questions. No account needed, and without one nothing is recorded.

/api/mcp/fullFull access

Every question bank for all four certifications, and it tracks your answers by blueprint domain — so it drills you on what you get wrong, weakest domain first.

Two, not one: a client only offers sign-in after a server refuses it once, and the free endpoint never refuses anybody.

Set it up · pick your client

Claude Codeterminal

claude mcp add --transport http ai-certificates https://aicertificates.study/api/mcp

That is the free endpoint — no account, works immediately. Bought full access? Use /api/mcp/full instead, then run /mcp and sign in.

claude.aibrowser

01Open Customize → Connectors.

02Click +, then Add custom connector.

03Paste the URL and click Add. Advanced settings stay empty — the server describes its own sign-in.

Free plans get one custom connector. On Team and Enterprise an Owner adds it once under Organization settings → Connectors; everyone else clicks Connect.

ChatGPTbrowser

01Turn on Developer mode under Settings → Security and login.

02Add a plugin: click +, give it a name and description.

03Choose Public endpoint and paste the URL.

OpenAI says developer mode depends on your account and workspace policy, so it is not on every plan.

Codexterminal

Codex adds hosted servers in ~/.codex/config.toml:

[mcp_servers.ai-certificates]
url = "https://aicertificates.study/api/mcp"

That is the free endpoint. Bought full access? Point url at /api/mcp/full and run codex mcp login ai-certificates to sign in. codex mcp add takes a local command rather than a URL, so it is the config file for a hosted server like this one.

Or let the assistant do it:one copied prompt covers all four clients — paste it into whatever you use and it walks you through your own screen.

Paste into an assistant and it will walk you through your own client.

Five tools

list_examsAll four Claude certifications with codes, fees, and every domain at its published weight.
get_practice_questionOne question, options shown, no answer key. The assistant cannot leak what it was not given.
check_answerGrades after you commit and explains every option, including the ones you did not pick.
my_progressYour record by blueprint domain, weakest first — what the drilling runs on. Full endpoint only.
open_studyOpens the study page itself when you'd rather answer on screen than in chat — and once it is open, WebMCP hands the agent a richer set of tools for that screen.

How it is built

There are two ways an agent reaches these tools. A connected client — Claude Desktop, Claude Code, anything speaking MCP over HTTP — talks to /api/mcp. An agent running inside the browser gets the same tools through WebMCP, the document.modelContext surface, with no connection to set up at all.

The obvious way to build the second one is to write the tools again in client code. That is a mistake, and an expensive one. Two copies of a tool definition drift within a week, and the first symptom is not a crash — it is a page-side tool cheerfully handing out an answer key the server withholds. Silent, and exactly the failure the product cannot survive.

So there is one manifest. A single function is what /api/mcp answers tools/list from, and the page receives that same manifest as a prop from a server component, then proxies every call straight back to the endpoint. A browser agent and a desktop client run one implementation: same schemas, same annotations, same entitlement checks, same refusals.

One detail that cost a rewrite: register the tools synchronously. The first version called tools/liston mount and registered whatever came back — which leaves a window, however short, where the page is up and advertising no tools at all. An agent that enumerates during that window sees a site with nothing to offer, and nothing ever tells it to look again. Taking the manifest as a prop removes the network call, so the tools exist in the first effect after paint.

WebMCP is not widely available yet. ChatGPT’s in-app browser supports it on some models and not others, Chrome has it behind an origin trial, and it is off in Enterprise and Edu workspaces. Where it is absent this code does nothing at all — which is the correct behaviour, and does mean most people arrive through the connected-client route above.

Why not paste questions into a chat?

A model that can see the answer key either hands it to you or folds when you argue. Over MCP it gets the question and not the key — you commit first, then it explains why the near-misses lose. On these exams two options are usually defensible and one is credited; the difference is the lesson.

What is recorded:the outcome only — which question, which domain, right or wrong. No conversation is stored, and without an account nothing is recorded at all. More on the privacy page.

No assistant handy?

Practice here instead — free, no signup

Pick an exam

Sign-in is OAuth 2.1 with PKCE · openapi.json · llms.txt · developers