Purpose
Qiniso is a deterministic fact-checking layer for AI agents. Agents often confidently output structured facts such as IBANs, phone numbers, domains, VAT numbers, crypto addresses, and more, but these values may be subtly and silently wrong. Qiniso compares the structured facts produced by agents against checksums and curated authoritative data, thereby catching erroneous values instead of continuing to trust them.
It is the deterministic complement in the guardrail stack: safety guardrails check whether output is safe, structural guardrails check whether output is well-formed, and hallucination guardrails then have another LLM judge whether the output is faithful to the prompt. None of these actually check whether a given structured fact matches the real world, because that requires computation or curated data, not another model's opinion. This is exactly where Qiniso is positioned.
On arbitrary identifiers, frontier LLMs have about a 91% probability of validating incorrectly in a cold-start, silent setting. Qiniso is 0%.
Main capabilities
- Validation scope: IBAN, VAT, VIN, barcodes, national and tax IDs, crypto addresses, domains, phone numbers, dates and holidays, and more.
- 56 tools in total, covering 8 domains.
- Each tool encapsulates an authoritative method: published checksum standards, audited libraries (libphonenumber-js, jsvat, date-holidays, @noble/hashes), or curated reference data (IANA root zone, UK VAT historical data).
How to use
Connect to Claude
In Claude, go to Settings → Connectors → Add custom connector, then paste it in. No login and no key are required. The endpoint is stateless, does not read user data, and does not require any secrets.
Use as a library
Each check is also a typed function, with no MCP required:
- It is a TypeScript monorepo. Each domain is a typed library under packages/*; the qiniso umbrella package aggregates them and exposes an MCP server through three transports: stdio (local / npx), Streamable HTTP (self-hosted), and Cloudflare Worker (managed edge endpoint). The same core also underpins the importable libraries.
Suitable scenarios
Suitable forscenarios that need to make structured fact fields generated by AI agents more reliable and verifiable, especially in domains sensitive to identifier correctness such as identity, finance, communications, and taxation.
Limitations
- It is not a real-time data provider. It validates the facts you provide; it does not return the current time, weather, or live exchange rates.
- It is not a credential store. It never asks for secrets or API keys. (To verify JWT signatures, use the library in your own runtime so that secrets never leave your machine.)
- It is not a registration status check. It validates the checksum of a VAT number, not whether that number is in valid registration status (VIES); it confirms that a domain's TLD actually exists, not that the domain has been registered.
