Config Guard

Zero-dependency security linter for MCP configurations.

Scans your .mcp.json for 20 types of security vulnerabilities before any MCP server starts. No API keys. No cloud. No LLM required.

Why?

43% of public MCP servers have command injection flaws (BlueRock TechReport 2026). Every MCP config you use is a trust boundary — and most developers never audit them.

Config Guard catches what humans miss:

• Typosquatted packages that look like real ones
• Servers with known CVEs (22 CVEs across 20 packages)
• 44 confirmed malicious packages (JFrog, Kaspersky, Lazarus APT, SANDWORM_MODE)
• Secret leakage in environment variables
• Rug-pull vectors (npx @latest auto-updates)
• Shadow servers exposing via tunnels

Install

pip install config-guard

Quick Start

# Scan your current directory's .mcp.json
config-guard

# Scan a specific project
config-guard --path /my/project

# Auto-discover all MCP configs on your system
config-guard --discover

# CI/CD integration (SARIF output for GitHub Code Scanning)
config-guard --sarif > results.sarif

# JSON output for scripting
config-guard --json

20 Security Checks

Every check is mapped to the OWASP MCP Top 10. See docs/OWASP-MAPPING.md for full CWE mappings.

#	Check	Risk	OWASP
1	Network exposure (non-localhost URLs)	HIGH	MCP-03
2	Rug pulls (npx @latest auto-update)	HIGH	MCP-07
3	Secret leakage (API keys in args/env)	HIGH	MCP-04
4	Command injection (shell=True)	CRITICAL	MCP-01
5	Path traversal (.. in arguments)	MEDIUM	MCP-05
6	Typosquat detection (Levenshtein distance)	HIGH	MCP-07
7	Dangerous permissions (--no-sandbox, sudo)	HIGH	MCP-06
8	Missing authentication on HTTP transport	MEDIUM	MCP-08
9	Sensitive path access (.ssh, .aws, .env)	HIGH	MCP-04
10	Overbroad filesystem access (/, C:\)	MEDIUM	MCP-06
11	Environment variable leaks (hardcoded secrets)	MEDIUM	MCP-04
12	Excessive server count (attack surface)	LOW	MCP-10
13	Known CVEs (22 CVEs across 20 packages)	CRITICAL	MCP-09
14	Symlink bypass (CVE-2025-53109)	HIGH	MCP-05
15	Shadow servers (ngrok, cloudflared, 0.0.0.0)	HIGH	MCP-05
16	Code execution (eval/exec patterns)	CRITICAL	MCP-01
17	Known malicious packages (44 confirmed malware)	CRITICAL	MCP-07
18	Deprecated SSE transport (no per-request auth)	MEDIUM	MCP-03
19	Shell servers (raw shell as MCP server)	CRITICAL	MCP-01
20	Unpinned packages (npx/uvx without version)	MEDIUM	MCP-04

CVE Database

Config Guard tracks known vulnerable MCP packages:

Package	CVE	Severity
mcp-remote	CVE-2025-6514	Critical (CVSS 9.6)
@modelcontextprotocol/server-git	CVE-2025-68145	Critical
mcp-server-git	CVE-2026-27735	Medium
@anthropic/mcp-server-filesystem	CVE-2025-53109	High
gemini-mcp-tool	CVE-2026-0755	Critical
mcp-vegalite-server	CVE-2026-1977	Critical
github-kanban-mcp	CVE-2026-0756	High
godot-mcp	CVE-2026-25546	High
fermat-mcp	CVE-2026-2008	Critical
@anthropic/mcp-inspector	CVE-2026-23744	Critical (CVSS 9.8)

Config Guard also detects confirmed malicious packages (e.g., postmark-mcp, @lanyer640/mcp-runcommand-server) that contain reverse shells or malware payloads.

Output Formats

Human-readable (default)

MCP Security Scan Results
========================================

[!] CRITICAL (1):
  [my-server] [MCP-09] Uses package with known CVE (CVSS-9.6): mcp-remote
    Fix: Update to latest patched version

[i] INFO (1):
  [safe-server] Server 'safe-server' passed all checks

Security Score: 85/100

SARIF (CI/CD)

config-guard --sarif > results.sarif

Upload to GitHub Code Scanning, Azure DevOps, or any SARIF-compatible tool.

JSON

config-guard --json | jq '.score'

GitHub Actions Integration

- name: MCP Config Security Scan
  run: |
    pip install config-guard
    config-guard --sarif > results.sarif

- name: Upload SARIF
  uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: results.sarif

Supported Configs

Config Guard scans these MCP configuration formats:

• Claude Code / Claude Desktop (.mcp.json, claude_desktop_config.json)
• Cursor (.cursor/mcp.json)
• VS Code (.vscode/mcp.json)
• Windsurf (.windsurf/mcp.json)

Use --discover to auto-find all configs on your system.

Scoring

Config Guard calculates a security score from 0-100:

Finding	Deduction
CRITICAL	-25 points
HIGH	-15 points
MEDIUM	-5 points
LOW	-3 points

100/100 = no findings. 0/100 = critical issues found.

Exit code is 1 if any CRITICAL or HIGH findings exist (useful for CI gates).

Comparison: Config Guard vs Agent Scan (Snyk)

Feature	Config Guard	Agent Scan (Snyk/Invariant Labs)
Approach	Deterministic regex + CVE database	LLM-based judges + deterministic rules
Dependencies	Zero (stdlib only)	Requires API calls to invariantlabs.ai
Offline	Fully offline	Requires internet (sends tool descriptions to cloud)
Privacy	No data leaves your machine	Tool names/descriptions sent to Snyk for analysis
Checks	20 static config checks	Prompt injection + tool poisoning + toxic flows
CVE database	22 CVEs across 20 packages	Dynamic (cloud-updated)
Malicious packages	44 confirmed (JFrog, Kaspersky, Lazarus, SANDWORM)	Dynamic detection
OWASP mapping	All 20 checks mapped to MCP Top 10 + CWE IDs	Not documented
SARIF output	Native SARIF v2.1.0 with CWE tags + fingerprints	Not available
CI/CD	GitHub Actions, any SARIF tool	Background mode (MDM/CrowdStrike)
Speed	<100ms (no network)	Seconds (API round-trip)
Install	pip install config-guard	uvx mcp-scan@latest
Scope	Config files only (pre-start)	Config + runtime tool descriptions
License	MIT (open source)	Proprietary (Snyk)

When to use Config Guard: CI/CD pipelines, air-gapped environments, privacy-sensitive setups, pre-commit hooks. Config Guard runs instantly with zero network access.

When to use Agent Scan: Runtime tool description analysis, prompt injection detection in live MCP servers, enterprise environments with Snyk Evo integration.

Best practice: Use both. Config Guard catches config-level issues before servers start. Agent Scan catches runtime tool poisoning after servers are running.

Zero Dependencies

Config Guard uses only Python standard library modules (json, os, re, sys, pathlib, argparse). No pip install surprises. No supply chain risk from transitive dependencies.

License

MIT License. See LICENSE for details.

Contributing

Issues and PRs welcome at github.com/KGT24k/config-guard.