LLMVerify

What LLMVerify does

The LLMVerify MCP server brings selected LLM safety checks into MCP-compatible agents and IDEs. It accepts model output or user-provided text and returns structured results for content verification, hallucination-risk assessment, prompt-injection detection, PII detection, and PII redaction. A capability-discovery tool reports what the server supports.

The server is intended for local triage rather than definitive fact checking or approval. Its checks can help identify risk signals before text is shown to a user, passed to another system, or written to a log. Responses include bounded outputs and PII-filtered results, along with notChecked and audit-related semantics where applicable.

How it works

The LLMVerify MCP server runs through the package’s built-in MCP command:

npx llmverify mcp

It communicates with the MCP client over standard input and output. The underlying checks are deterministic and pattern-based, so identical input evaluated under identical rules produces the same result. On the free tier, processing stays local: the server makes no model calls, sends no telemetry, and requires no API key.

The MCP interface exposes six tools:

• verify_llm_content for general output verification
• assess_hallucination_risk for hallucination-related signals
• check_prompt_injection for suspicious or malicious instructions
• check_pii for detecting personally identifiable information
• redact_pii for masking detected PII
• get_llmverify_capabilities for discovering supported checks

These checks align at a baseline mapping level with frameworks including the OWASP LLM Top 10, NIST AI RMF, EU AI Act, ISO 42001, and CSM6. This mapping is not certification.

Setup and configuration

Install the npm package in an environment with Node.js 20 or newer. Although the broader package supports Node.js 18 and above, the MCP SDK requires Node.js 20 or later. Add the following entry to an MCP client configuration:

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

No configuration file or API key is required for the local free-tier MCP server. The MCP SDK and validation dependencies are loaded for the MCP command without adding startup requirements to unrelated package commands.

Limitations and notes

LLMVerify is a 🔒 Security tool for risk reduction, not a truth oracle. Hallucination analysis uses patterns and cannot prove that a claim is false. Reliable fact checking still requires a source document or independent review.

Prompt-injection detection is also pattern-based, so new or obfuscated attacks may evade it. PII detection covers common formats such as emails, US phone numbers, SSNs, credit cards, and common API keys, but may miss encoded, hidden, image-based, or unusual variants.

The free tier has a local limit of 500 verification calls per day. Paid, ML-enhanced features require an explicit API key, while the free tier does not make network requests. Treat returned risk levels as triage signals, and review important claims or decisions separately.