veridigit

Verified validation of structured identifiers for AI agents — checksums, not guesses.

LLMs accept malformed IBANs, mistype card check digits, and invent ISBN and VIN check digits — confidently and silently. veridigit gives an agent the real checksum algorithms instead, returning a structured result with the parsed parts and clear reasons when something is wrong.

Benchmark

On 32 randomly generated, non-memorised identifiers, a frontier model with no tool got the check digit wrong 91% of the time (IBAN 100%, VIN 100%, ISBN-13 88%, card/Luhn 75%) — versus 0% for veridigit. Reproducible harness in the repo.

Tools

• validate_iban — ISO 7064 mod-97 checksum + country-specific length (75+ countries)
• validate_card — Luhn checksum, brand detection from the BIN, length rules
• validate_isbn — ISBN-13: 978/979 prefix + mod-10 weighted check digit
• validate_vin — ISO 3779 alphabet (no I/O/Q) + position-9 transliteration check digit

Install

{ "mcpServers": { "veridigit": { "command": "npx", "args": ["-y", "veridigit"] } } }

Also importable as a typed TypeScript library: npm install veridigit. Free, no API key. Apache-2.0.

Scope

Validates the structure of an identifier (format + checksum). Does not confirm a real, active, or owned account/card/book/vehicle, and makes no network calls.

Links: GitHub · npm · Benchmark