Use TigZig's US lending data. Five datasets, all running to Q2 2026: banks (FDIC), credit unions (NCUA), households (NY Fed Consumer Credit Panel), insurers' invested assets (Fed Z.1) and bank lending to non-bank financial firms. Loans, delinquency and charge-offs by segment, with the deepest series back to 1999.
Three of the five are on the API, the MCP server and the full database download - banks, credit unions and households. The other two come out as CSV from the tool. Worth knowing before you plan an integration around all five.
Each dataset carries its own methodology section and a validation section that reconciles it against the published source and names the anomalies. Delinquency stays on each source's own basis rather than being forced onto a common one: banks show noncurrent, 90+ and 30+ days, households show 90+ including severely derogatory plus the flow rates, and credit unions show 60+ and 30 to 59 days. Two of those three - NCUA and the NY Fed CCP - are not on the FRED API, so this is one of the few no-auth ways to pull them programmatically. All three are on the Tigzig API, free, with no key.
The same data is on four surfaces: interactive charts and custom cuts on the TREMOR web app, an open HTTP API, an MCP server for Claude / Codex / any agent, and full-database downloads (Parquet, SQLite, DuckDB, CSV). Everything is sourced from the regulators' own releases and validated before publishing.
Web tool: https://www.tigzig.com/tremor (US-Banks and US-credit-union modules). API reference: https://www.tigzig.com/apis/tremor. Methodology + validation: https://www.tigzig.com/tremor/bank-aggregates-methodology. Release notes: https://www.tigzig.com/post/us-lending-data-q1-2026-tigzig-api-jun2026.
Building something like this? How I work covers the rates, the availability and what I take on.