A product page is a set of promises. Ours says your AI can pull the mandatory fields of a message, compare two messages, list the valid values of a code set, show what changed between versions, and find the right message for a business concept. Those promises are only worth something if they are still true today, on today’s server, against today’s specification.
So we check. Regularly, and whenever the server or the specification changes, we hand every claim on our site to an AI and ask it to prove each one against the live server.
How the check works
The AI does what your AI does. It has no inside access and no shortcuts: it reads our server the same way an assistant in Claude Code or Claude Desktop does, searching, following links and reading the details of what it finds. For each claim it asks one question: did the server return what a person would need to do this? Names, XML tags, minimum and maximum occurrences, types, definitions, code values.
Each claim gets a result. Anything short of a clean pass is written up with the exact gap. Anything that was fine last time and is worse now is flagged as a regression. The results are dated, so the next run knows what has been checked recently, and the whole report is kept.
The latest run
This run covered 23 claims: the 14 capabilities on the MCP server page and the 9 scenarios from our use-case posts. It ran against our pre-release environment, the build that is next in line for production, on the current September 4, 2026 specification snapshot (3,340 message definitions in 36 business areas).
| What we checked | Result |
|---|---|
| Ask anything: mandatory fields of a message version, comparing two messages, the valid values of a code set | All pass |
| Generate code, samples and test data: classes from components, XML samples, typed interfaces | All pass |
| Version migration: what changed between two versions, which fields are no longer valid | All pass |
| Navigation: business areas, the messages in an area, finding messages by business concept | All pass |
| Cross-referencing other standards, and validating against the base specification | Pass, with stated limits (see below) |
| The 9 scenarios from our use-case posts: invoices, remittance, ATM cassette codes, digital signatures, QA test plans and more | All pass |
Nothing got worse. And two things got noticeably better since our last full check in March:
- Mandatory fields are now plain to see. Every field in a message carries its minimum and maximum occurrences, so your AI can say what is required without piecing it together.
- Code set values arrive with their meanings. Narrow a search to code sets and your AI gets the values, their short codes and their definitions in one step.
What we said more carefully
A claim you can verify beats one that sounds good. A couple of our statements now say precisely where the specification stops and your AI’s own knowledge starts:
- Mapping from legacy SWIFT MT formats: the server supplies the ISO 20022 side in full; the MT side comes from your AI.
- Validation: the server holds the base ISO 20022 rules. The extra rules of market profiles such as CBPR+, SEPA and FedNow are not part of the specification we serve, and the page says so.
The bar keeps moving up
We do not run this to collect a clean report. We run it because the specification changes, the server changes, and a promise that was true in March can quietly stop being true in November. Each run either confirms the promise or gives us a specific item to fix, and the two improvements above came out of exactly that. Next time we will check more of what the page promises, not less.
If you rely on the MCP server, this is the work behind the sentence “your AI doesn’t guess; it looks it up.” Try it and hold us to it.