@tierverify/analyzer-mcp

v1.0.0

Published

3 months ago

TIERverify Analyzer MCP Server — governance scoring for Claude

0High
0Medium
0Low

nfosec

mcp model-context-protocol tierverify governance ai-safety trust-score hallucination-detection compliance

TIERverify Analyzer for Claude

Governance scoring for AI-generated content — directly in your Claude conversation.

Paste any AI output and get a trust score, hallucination flags, fabrication detection, and governance gap analysis. Powered by the MIKE (Managed Inference Kernel Engine) scoring pipeline — a patent-pending, multi-dimensional governance engine that evaluates AI responses for accuracy, completeness, citation integrity, prescriptive language, and regulatory compliance.

No copy-paste to a browser. No switching tabs. Ask Claude to check it, and the results appear inline.

What You Get

tierverify_quick_check

Fast trust check. Paste AI-generated text, get a trust score (0-100), a plain-language headline, and a list of issues found. Takes a few seconds.

Example prompt: "Check if this AI response is trustworthy"

Example output:

## TIERverify Trust Check

Trust Score: 62/100 (Moderate Trust)

This response has governance concerns that should be reviewed.

### Issues Found (3)
- Potential issue: GDPR Article 33 requires notification within 72 hours,
  not 48. This is a factual error that could lead to non-compliant
  incident response timelines.
- Potential issue: SOC 2 TSC CC7.1 requires testing but does not mandate
  a specific frequency.
- Potential issue: ISO 27001 Section 14.2.8 does not exist in the current
  revision (2022).

Citations: 5/8 claims supported

tierverify_analyze

Full governance analysis. Evaluates AI output against regulatory frameworks (NIST 800-53, ISO 27001, SOC 2, HIPAA, PCI DSS, GDPR, FedRAMP, CMMC, and more). Returns five dimension scores, hallucination flags with severity, fabrication indicators, unsupported claims, and governance gaps.

Example prompt: "Run a governance analysis on this against NIST 800-53"

Parameters: | Parameter | Required | Description | |-----------|----------|-------------| | content | Yes | The AI-generated text to analyze | | framework | No | Regulatory framework (default: none). Options: nist_800_53_r5, nist_800_171_r3, iso_27001, soc2, hipaa, pci_dss, gdpr, fedramp, cmmc, nist_csf, nist_ai_rmf, eu_ai_act, oecd_ai | | role | No | User role for context (default: general_user). Options: auditor, control_owner, implementor, developer, analyst, risk_manager, business_user | | prompt | No | The original prompt that produced the output (enables deeper contextual analysis) |

tierverify_image_forensics

Image authenticity analysis. Upload an image and get verdicts for hidden data (steganography), tampering/manipulation, and AI generation detection — with confidence scores.

Example prompt: "Check if this image has been manipulated or AI-generated"

Setup

Prerequisites

Node.js 18 or higher (check with node --version)
A TIERverify account — sign up at analyzer.tierverify.com
An API key — generate one in your account settings at analyzer.tierverify.com/settings
An active subscription — API calls use your subscription unit balance

Install for Claude Desktop

Open your Claude Desktop config file:
- Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
- Windows: %APPDATA%\Claude\claude_desktop_config.json
Add this to the mcpServers section:

{
  "mcpServers": {
    "tierverify-analyzer": {
      "command": "npx",
      "args": ["-y", "@tierverify/analyzer-mcp"],
      "env": {
        "TIERVERIFY_API_KEY": "your_api_key_here"
      }
    }
  }
}

Restart Claude Desktop.
Start a new conversation and try: "Check this for governance compliance: [paste any AI-generated text]"

Install for Claude Code

Add a .mcp.json file in your project root (or wherever you launch Claude Code from):

{
  "mcpServers": {
    "tierverify-analyzer": {
      "command": "npx",
      "args": ["-y", "@tierverify/analyzer-mcp"],
      "env": {
        "TIERVERIFY_API_KEY": "your_api_key_here"
      }
    }
  }
}

Then restart Claude Code. The /mike slash commands become available automatically.

Tiered Disclosure — Badge First, Detail on Demand

TIERverify is designed around a two-tier model. You don't read a full report every time — you see a trust badge, and drill in only when something's flagged.

Tier 1 — Trust Badge (always visible): Every scored response gets a one-line badge. The procurement accountant sees "90/100, 0 flags" and moves on.

Tier 2 — Full Report (on demand): If there are flags, type mike detail for the full breakdown — dimension scores, hallucination flags, unsupported claims, governance gaps. No re-analysis; it pulls from the results already in context.

Usage Models

| Model | How | Best For | |-------|-----|----------| | Post-hoc | Have a conversation, then type mike to check the last response | Default — most users start here | | Combo | mike nist [question] — answer + badge in one shot | Power users who want both | | Detail | mike detail after any scored response | Investigating flagged issues | | Always-on | Configure via memory/project instructions to auto-score every response | Enterprise / compliance teams |

MIKE Command Vocabulary

Once the MCP server is installed, you can invoke MIKE with short, natural commands directly in your Claude conversation. No prompt engineering required — just type the command and your question.

Quick Commands

| Command | What It Does | |---------|--------------| | mike | Quick trust check on Claude's last response | | mike full | Full governance analysis (no framework) | | mike nist | Full analysis against NIST 800-53 Rev 5 | | mike hipaa | Full analysis against HIPAA | | mike soc2 | Full analysis against SOC 2 | | mike image | Image forensics on an attached image | | mike check [text] | Quick trust check on arbitrary pasted text | | mike detail | Show full breakdown from the most recent analysis |

Example: Inline Scoring

Ask a question with a framework command and get a scored answer in one shot:

You:   mike soc2 what are the accounting control requirements for procurement?

Claude: [answers the question, then appends:]

        MIKE Trust Score: 90/100 (High Trust) | 0 flags
        Type "mike detail" for full report

Need the full breakdown? Just type:

You:   mike detail

Claude: ## TIERverify Governance Analysis

        Overall Trust Score: 90/100 (High Trust)

        | Dimension    | Score | Assessment |
        |--------------|-------|------------|
        | Citations    | 85    | Good       |
        | Accuracy     | 95    | Excellent  |
        | Completeness | 80    | Good       |
        | Governance   | 100   | Excellent  |
        | Objectivity  | 100   | Excellent  |

        No hallucination flags.
        No unsupported claims.

Persistent Memory Setup (Claude Desktop / claude.ai)

Claude's memory persists across conversations. Paste the following block into your Claude memory settings once, and the command vocabulary works in every future conversation — no re-configuration needed.

Recommended memory entries:

MIKE command vocabulary: "mike" = tierverify_quick_check on Claude's last
response. "mike full" = tierverify_analyze, no framework. "mike nist" =
tierverify_analyze, framework=nist_800_53_r5. "mike hipaa" = tierverify_analyze,
framework=hipaa. "mike soc2" = tierverify_analyze, framework=soc2. "mike image" =
tierverify_image_forensics on attached image. "mike check [text]" =
tierverify_quick_check on provided text. "mike detail" = surface full dimension
breakdown, flags, and unsupported claims from the most recent MIKE analysis.

After every MIKE tool call, always surface the trust score, verdict, and top
flags directly in response text. Never leave results buried in the collapsed
tool call block.

To set this up: go to Settings → Memory in Claude, add the entries above, and you're done. Every new conversation will recognize the mike commands automatically.

Slash Commands (Claude Code)

If you use Claude Code (CLI), install the MCP server via .mcp.json (see Setup above), then use these built-in slash commands:

| Command | Description | |---------|-------------| | /mike | Quick trust check on Claude's last response — badge only | | /mike-check [text] | Check arbitrary pasted text — badge only | | /mike-detail | Full breakdown from last analysis (no re-run) | | /mike-full | Full governance analysis with complete report | | /mike-answer [question] | Answer the question, then score the answer — auto-detects frameworks | | /mike-answer [framework] [question] | Answer the question, then score against a specific framework |

Framework shortcodes for /mike-answer: nist, nist-171, iso, soc2, hipaa, pci-dss, gdpr, fedramp, cmmc, nist-csf, nist-ai, eu-ai, oecd-ai

Which Command Do I Use?

| I want to... | Command | |--------------|---------| | Score text someone else gave me | /mike-check [paste the text] | | Score what Claude just said | /mike | | Ask a question and score the answer | /mike-answer [question] | | Ask a question against a specific framework | /mike-answer pci-dss [question] | | Get a full governance report on Claude's last response | /mike-full | | See the details behind a trust badge | /mike-detail |

Common mistake: Don't pass a question to /mike-check. It scores text as-is — so your question gets flagged as unsupported claims. If you want Claude to answer first and then score, use /mike-answer.

How It Works

You: "Check this AI response for trustworthiness"
 |
Claude recognizes TIERverify tool is available
 |
Calls tierverify_quick_check with your text
 |
Your API key authenticates the request
 |
MIKE scoring engine analyzes the content:
  - Citation verification
  - Hallucination detection
  - Fabrication analysis
  - Completeness evaluation
  - Prescriptive language check
  - Governance gap identification
 |
Trust score + findings returned to Claude
 |
You see the results inline in your conversation

The MCP server runs locally on your machine. Your content is sent to the TIERverify Analyzer API for scoring and is not retained after analysis — content is discarded immediately after the trust score is computed. See our Privacy Policy for details.

Pricing

API calls consume units from your TIERverify subscription:

| Analysis Type | Unit Cost | |---------------|-----------| | Quick Check | 3 units minimum (1 unit per 3,000 characters) | | Full Analyze | 3 units minimum (1 unit per 3,000 characters) | | Image Forensics | 3 units per image |

View plans and pricing at analyzer.tierverify.com/pricing.

Environment Variables

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | TIERVERIFY_API_KEY | Yes | — | Your API key from analyzer.tierverify.com/settings | | TIERVERIFY_API_URL | No | https://analyzer.tierverify.com | API endpoint (only change for self-hosted or testing) |

Supported Frameworks

TIERverify evaluates AI output against major governance and compliance frameworks:

NIST SP 800-53 Rev. 5 — Federal information security controls
NIST SP 800-171 Rev. 3 — CUI protection for defense contractors
NIST Cybersecurity Framework 2.0 — Cross-sector risk management
NIST AI Risk Management Framework — AI-specific governance
ISO/IEC 27001:2022 — International information security
SOC 2 (TSC) — Service organization trust criteria
HIPAA — Healthcare data protection
PCI DSS v4.0 — Payment card industry security
GDPR — EU data protection regulation
FedRAMP — Federal cloud authorization
CMMC — Cybersecurity maturity for defense supply chain
EU AI Act — European AI regulation
OECD AI Principles — International AI governance

FAQ

Do I need a TIERverify account? Yes. Sign up at analyzer.tierverify.com and generate an API key in your settings.

Is my content stored? No. Content is analyzed and immediately discarded. Only the trust score and metadata are retained for your session history. See our Privacy Policy.

What AI models does this work with? TIERverify scores the output, not the model. It works on text from any AI — ChatGPT, Claude, Gemini, Llama, Mistral, or any other source.

Can I use this in CI/CD pipelines? Yes. The same API key works for any integration. Send a POST to analyzer.tierverify.com/api/analyze with your Bearer token.

What happens if I run out of units? The tool returns a clear error message with a link to purchase more units or upgrade your plan.

Development

npm install
npm run build
npm run mock      # Start mock API server for testing (no real API calls)
npm run inspect   # Open MCP Inspector for interactive testing

See mock-server/HOW-TO-TEST.md for a step-by-step testing guide.

Support

Issues: github.com/TIERverify/analyzer-mcp/issues
Email: [email protected]
Website: tierverify.com