Groundlens-mcp

MCP server for groundlens — LLM hallucination detection for Claude Desktop, Cursor, Windsurf, and any MCP-compatible client. No second LLM. Deterministic. Same inputs → same scores, every time.

What it does

Adds three tools to your AI assistant:

SGI (Semantic Grounding Index) measures whether the response actually used the source material or just rephrased the question. Score > 0.95 = grounded.

DGI (Directional Grounding Index) measures whether the response follows geometric patterns typical of grounded answers. Score > 0.30 = grounded.

Install

pip install groundlens-mcp

Or with uv:

uv pip install groundlens-mcp

Configure your client

Claude Desktop

Add to your claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "groundlens": {
      "command": "groundlens-mcp"
    }
  }
}

If you installed with uv and the command isn't on your PATH:

{
  "mcpServers": {
    "groundlens": {
      "command": "uv",
      "args": ["run", "groundlens-mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "groundlens": {
      "command": "groundlens-mcp"
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "groundlens": {
      "command": "groundlens-mcp"
    }
  }
}

How to use

Once configured, ask your ai assistant:

"Check if this response is hallucinated"

"Is this answer grounded in the document I provided?"

"Run a hallucination check on this ChatGPT output"

The tools return JSON with a verdict (GROUNDED or HALLUCINATION RISK), a numeric score, and a plain-language explanation.

Example output

{
  "verdict": "HALLUCINATION RISK",
  "explanation": "The response may not be based on the source material provided.",
  "method": "SGI (Semantic Grounding Index)",
  "score": 0.8721,
  "threshold": 0.95,
  "flagged": true,
  "detail": {
    "q_dist": 0.4312,
    "ctx_dist": 0.4945,
    "interpretation": "Response stayed close to the question rather than engaging with the context."
  }
}

How it works

groundlens uses embedding geometry — not a second LLM — to detect hallucinations:

• SGI computes dist(response, question) / dist(response, context). If the response moved toward the context, it's grounded. If it stayed near the question, the context was likely ignored.

• DGI projects the question→response displacement onto the mean direction of verified grounded pairs. Positive alignment = grounded pattern.

Both methods run a single embedding call. No model inference for evaluation. Deterministic.

First-call latency

The first tool call downloads and loads the sentence-transformer model (~100MB). Subsequent calls are fast. The model is loaded lazily so your MCP client doesn't slow down on startup.

Running from source

git clone https://github.com/groundlens-dev/groundlens-mcp.git
cd groundlens-mcp
pip install -e .
groundlens-mcp

Or:

python -m groundlens_mcp

Links

• groundlens library — pip install groundlens

• Documentation

• Website

• Demo

License

MIT