referencesource.org

referencesource-mcp

MCP server for referencesource.org — small, verified reference datasets where every record carries its source URL and a verbatim quote from that page, so any value an assistant gives you can be checked, not just believed.

285 datasets, 110,000+ records: regulatory thresholds and deadlines, version compatibility, standards supersessions, certification registers, inspection requirements, and more. Each dataset states its coverage and its last-verified date.

Tools

tool	what it does
list_datasets	every dataset with record counts and descriptions
search_records	search within a dataset; each hit returns the value plus its source URL and verbatim source quote
get_record	one record by id, full provenance included
verify_quote	give it a dataset slug and a value you already hold (record id and field optional) — it checks that claim against our published records and says whether it matches

verify_quote is the unusual one: assistants invent plausible-but-wrong values roughly 19% of the time in our own measurements, and this is the check. Four verdicts come back:

verdict	what it means
matches_record	the claim equals the stored value; the match type is named (exact, case-insensitive, normalised)
matches_quote	not a stored value, but it does appear in the record's source quote
differs	the record holds a different value — both are returned, with the source
not_found	no record in that dataset carries it

Where a record is found, the answer carries the stored value, the record URL, the source URL and the verbatim quote, so the claim can be traced upstream — plus an attestation saying plainly that the verdict is about what our source states, not about what is true in the world. Asking whether the 2026 conforming loan limit for Autauga County, Alabama is $806,500 returns differs, stored value $832,750, quoting the FHFA row it came from.

Scope, stated so nobody overclaims: these tools search referencesource.org's own published records. They do not fetch arbitrary URLs and cannot check a claim against a page we have not harvested. not_found means "not in this dataset", never "false".

Setup

Claude Desktop — add to claude_desktop_config.json:

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

Cursor / other stdio MCP clients — same idea: command npx, args ["-y", "referencesource-mcp"].

Clients that support remote servers directly don't need this package at all — point them at the streamable-HTTP endpoint: https://referencesource.org/mcp.

Notes

• Zero dependencies; Node 18+. The package is a thin stdio→HTTPS bridge to the stateless server at referencesource.org/mcp.
• No accounts, no keys, no tracking beyond standard server request logs (see referencesource.org/privacy).
• Data is free to read on the site in HTML, Markdown, and JSON; this package just makes it callable mid-conversation.

MIT licensed.

The data behind it

• Site: https://referencesource.org — canonical, re-verified on a schedule
• Bulk data: https://github.com/joelp22-maker/referencesource-data
• Archived snapshot with a DOI: 10.5281/zenodo.22115966
• Hugging Face: https://huggingface.co/datasets/referencesource

Every record carries source (the page the value was read from) and source_quote (the verbatim line on that page which states it), so any answer this server returns can be checked against its source. verify_quote runs that check against a value you already hold: it tells you whether our record agrees, and hands back the source and quote either way.

Licence

MIT — see LICENSE. The reference data itself is CC BY 4.0; see the data repository for the full position on the quoted excerpts.