Reference Source — verified reference data lookup & MCP server

Look up a regulatory threshold, a certification status, a jurisdictional rule or a version-compatibility fact, and get back the value plus the URL it came from and a verbatim quote from that page. Every record. No exceptions.

181 datasets, 73,109 records (2026-08-19), all published at referencesource.org.

What it does

Four operations:

verify_quote is the one that is hard to get elsewhere. Give it a value and it reports whether it matches the stored value, whether it appears in the source quote, and returns the source URL either way — so a number you got from somewhere else can be traced to a page that states it, or shown not to be there.

Two ways to use it

As an integration — run the Actor with an input, read the dataset. One row per record, each row carrying source, source_quote, last_verified and stale_after. Works anywhere Apify does: Make, n8n, Gumloop, the API, a schedule.

{
  "tool": "search_records",
  "dataset_slug": "conforming-loan-limits",
  "query": "Autauga"
}

{
  "tool": "verify_quote",
  "dataset_slug": "conforming-loan-limits",
  "record_id": "01001",
  "claimed_value": "$806,500"
}

→ differs, naming the value the record actually holds and quoting the FHFA row it was read from.

As an MCP server — this Actor runs in Standby mode, so it is a streamable-HTTP MCP endpoint your agent can call directly:

https://<your-username>--referencesource-lookup.apify.actor/mcp

All four tools are read-only (readOnlyHint: true), touch nothing outside the published catalogue, and never fetch an arbitrary URL.

What's in it

Regulatory thresholds and effective dates, certification and validation registers, end-of-support and deprecation dates, licensing and eligibility rules, state-by-state legal variation, compatibility and fitment tables. Call list_datasets for the current list — it changes as datasets are added and re-verified.

Pricing

Pay per event: one charge per tool call, plus one per record actually delivered. A search that matches nothing costs the call and nothing more, and a call that fails is not charged at all.

Provenance

Every value is published with the page it was read from and a verbatim quote from that page, and the extraction is checked by asserting the value appears inside its own quote. Datasets carry a last_verified and a stale_after date; anything found wrong is fixed or withdrawn. If the source does not state it, there is no row — this data contains no estimates, no interpolation and no model-generated values.

Method: https://referencesource.org/method/ Connect: https://referencesource.org/connect/