Connect your AI assistant to verified reference data
The MCP server gives your AI assistant direct access to every dataset on this site: look up a record, search across a dataset, and verify a claimed value against the source. Every record returned carries its source URL and a verbatim quote from that page, so neither you nor the assistant has to trust the answer on faith.
What you can do
| Tool | What it answers |
|---|---|
list_datasets | What reference datasets are available? Returns every dataset with its record count and last-verified date. |
search_records | Look up a specific record by state name, chemical, certificate number, product name or any keyword. |
get_record | Fetch one record by id with full provenance -- source URL and verbatim quote. |
verify_quote | Is this value actually right? Check a claimed fact against what the published record says. |
Option 1: the npm package
For Claude Desktop, add this to claude_desktop_config.json under "mcpServers": server name "referencesource", command "npx", args ["-y", "referencesource-mcp"].
For Claude Code CLI: claude mcp add referencesource -- npx -y referencesource-mcp
For Cursor or other stdio MCP clients: same command and args as above.
The package is a zero-dependency stdio bridge (Node 18+) that forwards requests to the remote endpoint. Source: referencesource-mcp on npm.
Option 2: the remote endpoint
Clients that support remote MCP servers can connect directly to https://referencesource.org/mcp -- no package needed.
The endpoint speaks MCP streamable HTTP, stateless. No API key, no account, no session. Every request is independent.
What it covers
110+ datasets and 68,000+ records across regulatory thresholds, version compatibility tables, certification registers, inspection requirements, licensing rules and more. Each dataset states its coverage, licence and last-verified date. The data is the same as the site -- the MCP server reads the same published JSON.
No tracking beyond request logs
No accounts, no keys, no cookies. The site logs requests for operational monitoring (what was requested, when, the user-agent string). The privacy page has the full picture.