Grounding Enforcer (MCP)

A standalone MCP server that validates model output against retrieved sources. It flags any claim, statistic, attribution, quote, or URL that cannot be traced back to a real source.

This repository contains the full implementation, including all grounding checks, violation logic, and output-cleaning modes. It runs independently and does not depend on the full suite.

Usage

Add this server to your MCP client configuration (Claude Desktop, Cursor, Windsurf, or any MCP‑compatible environment):

{ "mcpServers": { "grounding_enforcer": { "command": "python3", "args": ["/path/to/mcp-grounding-enforcer/server.py"] } } }

Related MCP server: truth-anchor-agent

Modes

• strict — removes sentences containing ungrounded spans

• lenient — rewrites ungrounded spans with hedging

• all (default) — removes fabricated retrieval claims only

Relationship to the full suite

This suppressor is also included in the consolidated mcp-hallucination-suite, which bundles all four suppressors and provides a meta-orchestrator: https://github.com/steveswain14/mcp-hallucination-suite

Related repositories

• mcp-prompt-suppressor

• mcp-json-suppressor

• mcp-tool-response-suppressor