groundcheck-mcp

groundcheck-mcp

The grounding check agents run before they commit to an answer.

Groundcheck is an MCP server that verifies a factual claim against live sources and returns a verdict (supported, refuted or unverified), a confidence score, and cited sources. Call it mid task, before your agent states a fact it is not sure of. It refuses to guess: conflicting evidence returns unverified, and no evidence can never produce a supported verdict.

Install

claude mcp add groundcheck -- npx -y groundcheck-mcp

Works the same way in Cursor, Cline, VS Code, or any MCP client that speaks stdio.

Tools

Tool	What it does
verify_claim	Ground one factual claim. Returns verdict, confidence, sources, sufficiency, conformal guarantee, signed receipt.
check_citations	Extract and verify every claim in a draft. Per claim report.
resolve_instrument	Resolve a ticker/ISIN/CUSIP/SEDOL/FIGI or name to canonical FIGI identity with provenance.
extract_claims	Split text into independently checkable atomic claims (rule-based, receipt-bound).
attest_delivery	Signed, offline-verifiable delivery receipt binding an x402 payment to what the paid service delivered.
attribution_badge	Markdown badge marking content as groundchecked.

Engine

This package is a thin stdio layer. The evidence engine (retrieval, stance classification, verdict) is a Python service. By default the server looks for it at http://127.0.0.1:8723 and can spawn it from a local clone; set GROUNDCHECK_ENGINE_URL=https://groundcheck.seiche.info to use the hosted engine instead. On the hosted engine verify_claim is free forever; the other verification tools are pay-per-call (x402, USDC on Base) after a small free daily trial — a 402 response arrives as a readable tool result with the offer, never an opaque error. A local engine is free and unmetered.

Source, engine code, and self hosting guide: https://github.com/beepboop2025/groundcheck

License

MIT