fin-regbase
A wiki of UK and US financial regulation that AI agents can query. Every claim cites a document and section.
Agents in fintech need regulation they can rely on. Searching raw PDFs on every call means redoing the same work each time. So I compile each regulation once into short, cited pages and check them by hand. Agents read the pages through an MCP server.
Not legal advice.
What’s in it
As of 4th of October, 2026: 204 pages, 155 UK and 49 US.
UK: Consumer Duty, consumer credit (CONC), complaints (DISP), AML and sanctions, SM&CR, SYSC, conduct of business (COBS), client money (CASS), market abuse (MAR), crypto promotions, UK GDPR.
US: CFPB rules (Dodd-Frank Title X, Regulations B, C, E, V, X, Z), Bank Secrecy Act, OFAC.
Not yet: SEC Reg BI and FINRA.
Tools
The MCP server has four: search_regulations, get_page, list_pages, get_related. Search is BM25.
Evals
A small model answers each question from the wiki. A larger model grades the answer. Pass is 0.80 or more.
| run | questions | date | result |
|---|---|---|---|
| UK set | 58 | 28th of April, 2026 | 94.8% pass |
| UK and US, full | 109 | 13th of May, 2026 | 69.7% pass |
| UK and US, search only | 109 | 13th of May, 2026 | 91 in top 3 |
The 109-question set added the US questions. Only partial runs since 13th of May, 2026.
Three UK questions fail and I left them failing on purpose. If I edit pages to pass specific questions, the eval stops meaning anything.