Skip to content

MCP client integration · mcp-sentinel-scanner

Evidence, inside
your MCP client.

Bring Sentinel's 184-rule catalogue into Claude, Cursor, or another MCP client. Supply server metadata or enumerate a live endpoint, then inspect the findings, evidence, and coverage-aware verdict.

npx mcp-sentinel-scanner

The client integration analyzes supplied inputs and lists a live server's declared surface. It does not invoke tools on that target.

Quick Setup

1

Add to Claude Desktop

Open your Claude Desktop config and add the scanner as an MCP server:

{
  "mcpServers": {
    "mcp-sentinel-scanner": {
      "command": "npx",
      "args": ["-y", "mcp-sentinel-scanner"]
    }
  }
}
2

Ask Claude to scan

Just ask in natural language. The scanner exposes three tools that Claude can call:

Scan a live endpoint“Scan the MCP server at https://api.example.com/mcp for security issues”
Analyze server metadata“Check this MCP server's tools for prompt injection risks”
List detection rules“What security rules does the scanner check for?”
3

Get actionable results

Every finding includes evidence (what triggered it) and remediation (how to fix it).

Available Tools

scan_server

Analyze server metadata without a live connection. Pass tools, descriptions, source code, and dependencies. Returns findings + verdict.

Input: tools, source code, dependenciesOutput: findings + verdict

scan_endpoint

Connect to a live MCP server endpoint. Enumerates tools via initialize + tools/list (safe, read-only), then runs all 184 detection rules.

Input: endpoint URLOutput: findings + verdict + connection info

list_rules

List all available detection rules. Filter by category or severity to explore specific rule sets.

Input: optional category/severity filterOutput: rule list with names, categories, severity

Example Output

What a scan result looks like:

{
  "server_name": "my-mcp-server",
  "verdict": "risk",
  "verdict_summary": "Risk - a critical issue (or a lethal-trifecta
    capability pattern) was observed across 3 findings.",
  "lethal_trifecta": false,
  "severity_counts": {
    "critical": 1,
    "high": 1,
    "medium": 1,
    "low": 0,
    "informational": 0
  },
  "findings_count": 3,
  "findings": [
    {
      "rule": "Prompt Injection in Tool Description",
      "severity": "critical",
      "evidence": "Tool description contains injection pattern:
        'ignore all previous instructions'",
      "remediation": "Remove instruction-like language from
        tool descriptions. Use factual, concise descriptions.",
      "rule_id": "A1"
    },
    {
      "rule": "Hardcoded Secrets",
      "severity": "high",
      "evidence": "Hardcoded API key pattern detected:
        sk-proj-abc123...",
      "remediation": "Move secrets to environment variables.
        Never hardcode API keys in source code.",
      "rule_id": "C5"
    },
    {
      "rule": "Missing Input Validation",
      "severity": "medium",
      "evidence": "Parameter 'query' has type string with
        no maxLength, pattern, or enum constraint.",
      "remediation": "Add maxLength, pattern, or enum
        constraints to string parameters.",
      "rule_id": "B1"
    }
  ],
  "rules_version": "2026.09"
}

Verdict Interpretation

A scan leads with a coverage-aware verdict, not a number. A clean result is only Safe when coverage was sufficient for that verdict - otherwise it is Insufficient coverage, never Safe.

SafeSufficient coverageCoverage was sufficient and nothing found warrants concern.
CautionReview recommendedFindings worth reviewing before you rely on this server.
RiskDo not deploySerious findings - fix before using in production.
InsufficientNot a passKey inputs were not analysed, so a clean assessment has not been established.

Detection Coverage

184 rules across 21 categories - counts derived from the live rule catalogue

A
Description AnalysisPrompt injection, unicode attacks, encoded instructions
9
B
Schema AnalysisMissing validation, dangerous defaults, unconstrained properties
7
C
Code AnalysisCommand injection, SSRF, SQL injection, hardcoded secrets
17
D
Dependency AnalysisKnown CVEs, typosquatting, malicious packages
7
E
Behavioral AnalysisMissing auth, insecure transport, declared-vs-observed divergence
7
F
Ecosystem ContextLethal trifecta, exfiltration chains, data flow risks
8
G
Adversarial AIIndirect injection, rug pull, context saturation
7
H
2026 Attack SurfaceOAuth vulnerabilities, initialize injection, multi-agent
3
I
Protocol SurfaceAnnotation deception, sampling abuse, consent fatigue
17
J
2026 Threat IntelligenceGit injection, schema poisoning, config poisoning
9
K
Compliance & GovernanceNIST, ISO 27001, EU AI Act, OWASP Agentic
20
L
Supply Chain AdvancedPost-install hooks, manifest confusion, secrets in build layers
16
M
AI Runtime ExploitationMulti-turn state injection, encoding attacks, unbounded tool chains
9
N
Protocol Edge CasesJSON-RPC batch/notification abuse, transport framing
15
O
Data Privacy AttacksTiming inference, sensitive local-data exfiltration
7
P
Infrastructure RuntimeContainer and Dockerfile misconfiguration
10
Q
Cross-Ecosystem EmergentMulti-protocol and cross-ecosystem emergent risks
7
R
MCP Apps / UI SurfaceUI HTML resource surface (ui://, SEP-1865)
1
S
Async Tasks AugmentationTerminal-state schema, task-handle caller identity
2
T
Streamable HTTP TransportStateless HTTP without auth, DNS-rebinding Origin gate
2
U
Schema 2020-12 + OAuthToken pass-through, audience validation, DCR consent reuse
4

Safety

This MCP client integration never invokes tools on target servers. It initializes the connection and uses the read-only list methods for enumeration. Its analysis is deterministic and does not use a language model. Findings still need review, and the verdict is limited by the inputs available.

Other MCP Clients

Works with any MCP-compatible client:

CursorSettings → MCP → Add Server
VS Code (Copilot).vscode/mcp.json
Windsurf~/.codeium/windsurf/mcp_config.json
Claude Codeclaude mcp add mcp-sentinel-scanner

02 / Follow the finding

A verdict opens the question.
Evidence lets you examine it.

Inspect the rule, its supplied inputs, and the evidence behind each finding. Keep completed execution separate from checks that could not run.

The same rule catalogue serves different input shapes. Coverage belongs to the actual assessment, rather than to the number of rules available.

The MCP client integration enumerates a live target; it never invokes that target's tools.

Read the evidence guide

A result is a starting point

Follow the finding to its source.

Understand what each rule examines, which inputs it needs, and how its findings relate to security controls.

Explore the rule catalogue