nullsec-mcp

Trust analysis for MCP servers and AI agents.

Nullsec analyzes MCP servers for dangerous tools, permission risks, prompt injection surfaces, and agent trust metadata. Every scan produces a Trust Score, a capability manifest, and a machine-readable NSIP manifest — the first open standard for agent trust.

Quick start

npx nullsec-mcp https://github.com/org/your-mcp-server

Output:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  Nullsec MCP Trust Score: 82/100
  Grade: B-
  Risk Level: low

  Dangerous Tools:       0
  Filesystem Access:     Yes
  Shell Execution:       No
  Wallet Access:         No
  Network Egress:        Yes
  Database Access:       No
  Credential Access:     No
  Prompt Injection Risk: low

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

What it checks

Module	Description
Tool Inventory	Discovers all registered tools, classifies capabilities (filesystem, shell, network, database, wallet, credentials)
Dangerous Capabilities	Flags tools with unguarded dangerous operations; distinguishes mitigated vs unguarded
Permission Analysis	Maps credential surface from env vars, README docs, and source code
Injection Surface	AI-powered detection of confused-deputy attacks, unsanitized outputs, and tool-chaining risks
Input Validation	Static analysis for path traversal, SQL injection, command injection, SSRF, eval
Network Egress	Identifies outbound domains and dynamic URL fetching (SSRF risk)
Trust Scoring	Deterministic 0–100 score with grade, risk level, and capability breakdown

Install

npm install -g nullsec-mcp

Usage

# Full scan with AI analysis
export ANTHROPIC_API_KEY=sk-ant-...
nullsec-mcp https://github.com/modelcontextprotocol/servers

# Skip AI analysis (faster, no API key needed)
nullsec-mcp https://github.com/org/server --no-ai

# JSON output (for CI/CD pipelines)
nullsec-mcp https://github.com/org/server --json

# Generate NSIP manifest
nullsec-mcp https://github.com/org/server --nsip

# Private repos
export GITHUB_TOKEN=ghp_...
nullsec-mcp https://github.com/private-org/server

# Custom size limit
nullsec-mcp https://github.com/org/large-server --max-size 500

NSIP Manifest

Every scan can output a draft NSIP manifest — a machine-readable trust document for MCP servers:

nullsec-mcp https://github.com/org/server --nsip

Generates nsip.json:

{
  "schema": "nsip/v0.1",
  "name": "example-mcp-server",
  "trust_score": 82,
  "grade": "B-",
  "capabilities": {
    "filesystem": true,
    "shell_execution": false,
    "wallet_transactions": false,
    "network_egress": true,
    "database": false,
    "credential_access": false
  },
  "tools": {
    "total": 5,
    "dangerous": 0,
    "names": ["read_file", "list_files", "search", "fetch_url", "summarize"]
  },
  "risk_level": "low",
  "prompt_injection_risk": "minimal",
  "external_domains": ["api.example.com"],
  "credential_surface": ["API_TOKEN"],
  "verified_by": "nullsec",
  "verified_at": "2025-01-15T10:30:00.000Z"
}

Use this to build trust registries, block dangerous servers in agent orchestrators, or show trust badges in marketplaces.

Exit codes

Code	Meaning
0	Trust Score ≥ 80 (low/minimal risk)
1	Trust Score 40–79 (medium/high risk)
2	Trust Score < 40 (critical risk)

Programmatic usage

import { fetchRepo, isMCPServer, runMCPScan, computeTrustScore, generateNSIP } from 'nullsec-mcp';

const repo = await fetchRepo('https://github.com/org/mcp-server');

if (await isMCPServer(repo.path)) {
  const result = await runMCPScan({
    repoPath: repo.path,
    emit: async (event) => console.log(event.event, event.detail),
  });

  const trust = computeTrustScore(result);
  console.log(`Score: ${trust.score}/100 (${trust.grade})`);

  const nsip = generateNSIP(result);
  console.log(JSON.stringify(nsip, null, 2));
}

await repo.cleanup();

Environment

Variable	Required	Description
ANTHROPIC_API_KEY	For AI analysis	Enables injection surface detection via Claude
GITHUB_TOKEN	For private repos	Also increases GitHub API rate limits

Development

git clone https://github.com/trynullsec/nullsec
cd nullsec
npm install
npm test
npm run dev -- https://github.com/some/mcp-server

License

MIT