This isn't an opinion, is fully verifiable: ask your AI the same regulatory question twice (once with your document corpus connected, once without) and compare what it cites. I ran it. The gap is not stylistic. It is the gap between a citation and a paraphrase.
The experiment: same question, same model, two answers
The question: which identifiers must a Digital Product Passport contain, and where does the data carrier physically go?
Without access to my corpus, working from memory, the answer was: a unique product identifier, operator and facility identifiers, and a carrier placed on the product, its packaging, or accompanying documentation. Directionally correct. Zero pinpoint references. Nothing an inspector could be shown.
With the corpus connected, the same model returned the consolidated ESPR text : CELEX 2024/1781, and told me where: in page 21 for the requirement that the carrier be physically present on the product, its packaging or documentation; in page 80 for the complete list of identifiers, letter by letter, EORI included, with the obligation to conform to the ISO/IEC 15459 standard. Then it added CWA 18186 (and his page) on the standards governing the weblink (61406, and 18975).
That ISO series and that page number I would never have produced from memory — or worse, I would have produced them almost right. Almost right is how compliance documents fail. A paraphrase is not a citation.
Italy uses AI six times more than it builds for it
I stopped guessing about this and counted. On 27 July 2026 the official Model Context Protocol registry contained 18.677 servers. I enumerated all of them, then resolved every GitHub owner namespace (7.763 owners, 7.617 profiles still live), and attributed nationality from public profile location ( .it WHOIS, and VAT) on the publisher’s own site.
Servers attributable to Italian authors or Italian organisations: 48 confirmed, across 40 publishers. That is 0.26% of the world registry. Way lower than I ever imagined.
Now the other side. Anthropic publishes country-level usage in its Economic Index (re. 15 Sep 2025, Anthropic/EconomicIndex - Hugging Face): 1.49% of global Claude usage, 13th country out of 201, with a per-capita usage index of 1.47, means that Italians use it 47% more than population alone would predict.
1.5% of the usage. 0.26% of the building. A ratio of 5.7 to 1. Italy consumes this technology intensively and builds almost nothing that other people’s AI can plug into. “That is the whole article in two numbers.”
Two honest qualifications, because a number without its perimeter is propaganda. 1st: 48 is a floor, not a total: only 40.6% of GitHub profiles declare a location, so the true figure is plausibly 90–100, still under 0.6% of the registry. 2nd: I ran this census against a second, independent one with different channels. It found 3 Italian publishers I had missed and it caught 1 error. It came out of the count.
Two censuses, built by different methods, reconciled item by item on the primary source rather than by majority vote.
A connector is not a plugin. It is a socket.
One paragraph, no jargon. An MCP server is a standard socket between a body of knowledge (your documents, your database, your email, your internal tools) and any AI model that supports the protocol. You build the socket once. Whoever plugs in gets your data, structured, with your rules about what may be read and what may not. It is USB-C for knowledge: the model changes, the socket stays.
Public MCP server vs Private MCP server
Reecopedia (that have IP https://ia.reeco.eco/mcp ) is Reeco’s public MCP server, published in the official registry as eco.reeco/reeecopedia . It exposes the EU textile sustainability corpus: ESPR, Digital Product Passport, CSRD, CBAM, EU ETS, CWA 18291, JRC preparatory studies, and national transposition records. It answers only from the returned passages, and it returns file and page with every answer. When the corpus does not contain the answer, it says so. That behaviour is the product.
Alongside the public one, the suggestion to my readers is : I run a private MCP hub. It connects my email, calendar, internal tools and local models. It is not in any registry, and it never will be.
Of these private servers there is no public census because they leave no public trace. Almost nobody in the corporate landscape realises yet what this enables. Italian material on MCP stops at basic “hello world” weather tutorials. Nobody writes about what it means to run one inside a company what it automates, how to secure it, and what it costs to maintain.
Why this matters commercially: you stop hoping to be cited
Every company in Europe is currently paying someone to make its content more visible to AI systems. The honest description of that work is: hoping. You publish, you optimise, you wait to see whether a model repeats you, and a competitor with a bigger budget can optimise you out tomorrow.
A connector inverts the relationship. When a client’s AI is plugged into my corpus, my sources are not one candidate among the model’s recollections they are the answer, with file and page attached. Nobody can outbid me out of my own socket.
You do not hope the AI cites you. You are the source, by contract.
You do not need a script to check this. You need two minutes.
If you want to test this yourself right now: open your AI client’s connector settings, add https://ia.reeco.eco/mcp , and ask it where the ESPR places the data carrier. Then switch it off and ask again.
For the technical readers: the census is reproducible on the official MCP registry (go to registry.modelcontextprotocol.io) and the country usage share on the Anthropic Economic Index dataset, huggingface.co/datasets/Anthropic/EconomicIndex. If you want to re-run the queries and compare numbers, write me and I will give you the exact calls.
Guides to build basic MCP servers are already everywhere online. But if you are thinking to build a real, secure socket for your company’s proprietary knowledge and you need a hand on how to structure it, just let me know.
Stefano Cipriani is founder of Reeco®, an Expert Member of CIRPASS-2 (EWG1, EWG3), and a JRC Registered Stakeholder.
Disclosure: language drafting of this document was assisted by a large language model. All the rest (concepts, data, analysis and conclusions) are the author’s own, verified against primary sources.