ComplianceSME
Two filters, one standard behind both. What you give each one, what comes back, and what it will not pretend to know.
Every scan hands you a reading. A reading is the answer already written out for you in plain words, before any model sees it, and your Claude is told to show it to you exactly as it arrives and then answer your questions about it. Nothing in it is a model's own summary of your work.
A reading comes in two halves, in this order. The writing half applies the ComplianceSME writing standard to everything your work says, and names every machine fingerprint in it by line. The law half gives each duty a verdict, with the wording of the article it turns on underneath, and for a failure the evidence found and the ComplianceSME toolkit that closes the gap.
A verdict is one of four marks.
PASS means the evidence is on the page.
FAIL means the duty applies to you and the thing it requires is absent.
UNKNOWN means the check cannot see the answer, because the answer needs a rendered browser, a fact about your company, or something the served page does not carry. An unknown is never counted as a pass, and the reason it came back unknown is in the record.
PLANNED appears on a plan reading and is not a verdict. Nothing has been built for a check to read, so there is no pass and no fail. PLANNED says the work will carry this duty, and says what the duty requires.
Each account gets one free scan a day on each connector, counted on its own, so a scan on one is never charged against the other. The free reading gives you the writing result and how many duties passed, failed and came back unknown, by name. A key on your ComplianceSME account lifts the limit and opens the full report, with every fingerprint by line, the wording behind each duty, the evidence found on your page and the toolkit that closes each gap.
Fetching a reading you have already paid for does not spend a scan.
Anyone can try a ComplianceSME system in their own Claude account. Used for compliance, a system runs in a dedicated Claude account, because that is what fidelity testing showed: the systems are engineered for Claude, and frontier agentic models are not trustworthy tools for compliance.
Each reading says what it did not look at, and says why. Three things sit in that part of it.
Some duties cannot be settled by reading words at all. They turn on a record you keep, a process you follow, a notice you send or a communication nobody reading your page will ever see. Those are counted and named, so that a clean page is never read as a clean instrument.
Some instruments are held with no article quoted for them yet. They are named and left unchecked, because a verdict without the article's own words is the one thing this product will not print.
And some checks go quiet for a reason that is about the material rather than the law. A file sitting on your disk has no response headers, so the two cookie duties come back unknown on a build and say so; running the page check on the deployed address answers them.
For software built with AI, before it ships. It reads the law the build must meet and the standard of work a customer can trust, in one run, and it covers the words the build says as well as the duties it carries: interface copy, documentation, generated content, error messages, the pages a customer reads.
It offers five tools.
check_build takes the files a build produced, each one a path and its contents, and reads every page, error page and documentation file against the writing standard, then gives the HTML files their duty verdicts. It answers the two questions one page on its own cannot: whether the privacy notice can be reached from every page, and whether the page it points at is in the build at all. Nothing is fetched and nothing leaves the account. Use it for a site that is not deployed yet.
check_plan takes the plan as text, a specification, a readme or a design note, and reads the features it describes: payments, signing a person in, tracking, other people's content, AI features, output a machine generated, selling to consumers or to businesses, trading across borders, personal data, cookies, marketing by electronic mail, subscriptions, reviews, connected hardware, software placed on the market, and the surfaces a person reads and operates. Back comes every duty those features will carry, each one quoted and marked PLANNED, with the toolkit that covers it. A feature the plan does not describe carries no duty, and the reading says so rather than implying the work carries nothing.
check_page takes one live address and returns the writing result, then verdicts on the privacy notice, cookies and trackers set before consent, the accessibility markers the European Accessibility Act requires, trader information, AI disclosure, and the route for reporting a vulnerability.
check_text takes a body of copy and returns every machine fingerprint in it, located by line, with a count per list and a house style score taken per thousand words, so a long document is not scored as though it were short. Quoted material, instrument names and cited titles are reported as passed rather than scored. A body of copy owes no web duty, so the duty half does not run on it.
get_report takes the reference printed at the end of an earlier reading and hands that reading back, or one part of it: every fingerprint by page and line, every duty in the instrument's own words with the evidence and the covering toolkit, every passage the standard passed and why, or the whole thing at once.
Address on this site: /developer/mcp
The same standard, turned on what you publish. It reads a draft before it goes out, or a page you have already put out, and holds the claims in it to three rules: state only what your material supports, promise only what you control, and match a claim's confidence to its evidence.
It offers four tools.
check_text takes a draft and returns every machine fingerprint in it, located by line, with a count per list and a house style score per thousand words. It reads the shape of the writing as well as the words: the rhythm of the sentences, the proportions of the paragraphs, the spread of the vocabulary and the caveats nobody asked for. Tell it the kind of draft it has, and a news piece and a media script are read as what they are: in a script, the spoken lines and the captions are scored and the scene headings and camera notes are read and not scored, because nobody watching ever reads one. The claims rules come back with it, and the two of them that need the material behind a claim come back unknown rather than as a guess.
check_page takes a page or post you have published and returns the writing result and the claims result, then the duties ComplianceSME holds for what a page says, what it claims and what it collects: the transparency and reach of your privacy information under the GDPR, consent before trackers under the ePrivacy Directive and PECR, the trader information a buyer is owed, the total price up front, the reviews you publish, the information and the way out of a subscription under the DMCC Act, environmental claims under the Green Transition Directive, consent and a route to stop before marketing by electronic mail, and the disclosure the AI Act requires when a machine wrote it or drew it. Every failure is quoted to the duty in the instrument's own words and names the toolkit that closes the gap.
get_report works as it does on the other connector: the reference printed at the end of a reading brings that reading back, whole or in one part.
writing_script returns the writing standard itself, as plain text, for you to read and apply as you choose. It is the same standard check_text and check_page read a draft or a page against. It needs no key and uses no scan.
Address on this site: /content/mcp
Signing in is by emailed link and there are no passwords. What a tool call sends, what is written down afterwards and how to disconnect are set out in the privacy notice.