Skip to main content
Glama
mcp-guard
by zent7x

mcp-guard

CI npm npm downloads license Go mcp-guard MCP server

Security and network MCP server for Claude Code and Cursor. Written in Go.


Why mcp-guard

Most MCP servers are API wrappers. Claude could look up the same data itself if it had web access.

mcp-guard is different. 12 of its 22 tools require a physical machine to function. Wi-Fi scanning needs a radio chip. Bluetooth scanning needs an adapter. ARP discovery sends Layer 2 frames that never leave your local network — no cloud service receives them. File watching subscribes to kernel events on your machine's filesystem. USB enumeration reads your physical ports.

Claude runs in a data center. It has none of these things. These tools only work because the binary is running on your computer.


Related MCP server: Tengu

Tools

Local hardware (impossible for any remote service)

Tool

What it does

wifi_scan

Scan nearby Wi-Fi networks via your wireless hardware — SSID, BSSID, signal, channel, security

bluetooth_scan

Enumerate nearby Bluetooth devices via your adapter — name, address, type, pairing status

usb_devices

List all USB devices connected to this machine — vendor, product ID, speed, manufacturer

arp_scan

Layer 2 LAN discovery — finds every device on your network including ones that block ICMP/TCP, with MAC addresses and vendor IDs

ping_sweep

ICMP host discovery across a CIDR range from your machine's network stack

traceroute

Hop-by-hop network path from this machine via ICMP TTL probes

file_watch

Kernel FS event stream (FSEvents on macOS, inotify on Linux) — real-time create/write/delete/rename

sys_info

Local hardware: CPU model, RAM, disk, uptime, all network interfaces with IPs and MACs

open_files

Every file, socket, and pipe held open by processes on this machine via lsof

proc_list

Running processes with CPU% and memory — filter by name

net_connections

Live TCP connections on this machine (netstat -an)

scan_secrets

Walk local files for hardcoded credentials — 20+ patterns (AWS, GitHub, OpenAI, Stripe, Slack, DB URLs, private keys)

hash_files

SHA-256 every file in a directory — integrity baseline before/after deploys

Forensics and supply chain

Tool

What it does

persistence_scan

Checks this machine for malware persistence: LaunchAgents/Daemons (macOS), systemd units (Linux), cron jobs, shell profile injections. Flags curl-pipe-to-bash, base64 payloads, binaries in /tmp

supply_chain_audit

Audits node_modules for supply chain attacks: lifecycle scripts that download/execute remote code, typosquatting (Levenshtein distance 1 against 50 popular packages), eval() of runtime data in source

Network utilities

Tool

What it does

port_scan

Concurrent TCP scanner — 200 goroutines, service name lookup

banner_grab

Raw TCP banner from any protocol — SSH, FTP, SMTP, Redis, MySQL

ssl_inspect

Full TLS certificate chain — key size, algorithm, expiry, SANs

dns_enum

A, AAAA, MX, NS, TXT, CNAME — detects missing SPF/DMARC

audit_headers

HTTP security header audit scored 0–100 with letter grade

check_cves

npm dependency CVE check via OSV — no API key

jwt_decode

Local JWT decode — algorithm, expiry, security warnings


Setup

Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "mcp-guard": {
      "command": "npx",
      "args": ["-y", "@zent7x/mcp-guard"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "mcp-guard": {
      "command": "npx",
      "args": ["-y", "@zent7x/mcp-guard"]
    }
  }
}

Global install

npm install -g @zent7x/mcp-guard

Then use "command": "mcp-guard" instead of npx in the config above.


Example output

> wifi_scan

Wi-Fi networks (8 found)

SSID                              BSSID               SIGNAL    CHANNEL  SECURITY
──────────────────────────────────────────────────────────────────────────────────────────
HomeNetwork                       a4:c3:f0:11:22:33   -42 dBm   6        WPA2 Personal
OfficeWifi                        b8:27:eb:44:55:66   -67 dBm   11       WPA2 Enterprise
> arp_scan 192.168.1.0/24

LAN devices in 192.168.1.0/24 (6 found)

IP                  MAC                  HOSTNAME              VENDOR
────────────────────────────────────────────────────────────────────────────────
192.168.1.1         a4:c3:f0:ab:cd:ef    router.local          Apple
192.168.1.42        b8:27:eb:12:34:56    raspberrypi.local     Raspberry Pi
192.168.1.100       00:0c:29:78:90:ab                          VMware
> bluetooth_scan

Bluetooth devices (4 found)

NAME                            ADDRESS               TYPE                  STATUS
─────────────────────────────────────────────────────────────────────────────────────────
AirPods Pro                     a1:b2:c3:d4:e5:f6                           paired
MX Keys                         11:22:33:44:55:66     Keyboard              paired
Sony WH-1000XM5                 aa:bb:cc:dd:ee:ff     Headphones            not paired

How it works

The npm package downloads a pre-compiled Go binary for your platform on first run. The binary speaks the MCP stdio protocol — your editor spawns it on startup and the tools appear automatically.

Platform binaries: darwin-arm64, darwin-amd64, linux-amd64, linux-arm64, windows-amd64.


Build from source

git clone https://github.com/zent7x/mcp-guard
cd mcp-guard
go build -o mcp-guard .

Requires Go 1.21+.


License

MIT


Built by Adeeb Bashir (zentex) · zent7x.com

Available Tools

23 tools
arp_scanA
Destructive

Layer 2 LAN host discovery — finds ALL devices on local network including those that block ICMP/TCP. Returns IP, MAC address, hostname, and vendor. Only possible from a machine on the same network segment.

Inputs: cidr (required).

audit_headersB
Destructive

Audit HTTP security headers of a URL. Checks HSTS, CSP, X-Frame-Options, Referrer-Policy, Permissions-Policy. Returns score/100 and grade.

Inputs: url (required).

bluetooth_scanA
Destructive

Scan for nearby Bluetooth devices using local Bluetooth hardware. Returns device name, address, type, and pairing status. Requires a physical Bluetooth adapter — architecturally impossible for any cloud service.

Inputs: no named parameters.

check_cvesB
Destructive

Check npm dependencies in a package.json against the OSV vulnerability database. No API key needed.

Inputs: path (required).

critical_file_monitorA
Destructive

Monitor critical system files for unauthorized changes. Three actions: 'scan' shows current state + permission issues; 'baseline' saves SHA-256 hashes of all critical files to disk; 'check' compares current state against the saved baseline and reports every modified, added, or removed file. Covers SSH keys, shell profiles, /etc/hosts, sudoers, LaunchAgents (macOS), systemd/passwd/shadow (Linux). Detects world-writable files and overly permissive SSH keys.

Inputs: action (required), baseline_file, extra_paths.

dns_enumA
Destructive

Enumerate all DNS records for a domain: A, AAAA, MX, NS, TXT, CNAME. Detects missing SPF/DMARC records.

Inputs: domain (required).

file_watchB
Destructive

Watch a file or directory for changes using kernel-level FS events (FSEvents on macOS, inotify on Linux). Captures creates, writes, deletes, renames in real time. A background MCP process can do this — Claude in a chat window never could.

Inputs: path (required), seconds.

hash_filesA
Destructive

Compute SHA-256 hashes of all files in a directory. Use to create integrity baselines, detect tampering, or verify files haven't changed. Reads local disk — not possible remotely.

Inputs: path (required).

jwt_decodeA
Destructive

Decode and analyze a JWT token locally without sending it anywhere. Shows header, payload, expiry, algorithm, and security warnings (e.g. 'none' algorithm, expired token, weak signing).

Inputs: token (required).

net_connectionsA
Destructive

List active network connections on this machine (like netstat). Shows local/remote address and connection state.

Inputs: no named parameters.

open_filesB
Destructive

List files, sockets, and pipes currently open by processes on this machine. Filter by process name or PID. Uses lsof — shows exactly what a process is reading, writing, or listening on.

Inputs: filter.

persistence_scanA
Destructive

Scan this machine for malware persistence mechanisms: LaunchAgents/LaunchDaemons (macOS), systemd units (Linux), cron jobs, and shell profile injections. Flags high-risk patterns like curl-pipe-to-bash, base64-encoded payloads, and binaries executing from /tmp. Essential first step when investigating a potentially compromised machine.

Inputs: no named parameters.

ping_sweepB
Destructive

Send ICMP pings to all hosts in a CIDR range and return live hosts. Works on the local network — finds hosts even if they have no open TCP ports. Claude cannot send ICMP packets.

Inputs: cidr (required).

port_scanA
Destructive

TCP port scan a host. Returns open ports with service guesses. Uses 200 concurrent goroutines — fast. Claude cannot do this natively.

Inputs: host (required), end, start, timeout_ms.

proc_listB
Destructive

List running processes on this machine. Optionally filter by name. Shows PID, CPU%, memory usage, and command.

Inputs: filter.

scan_secretsB
Destructive

Scan a file or directory for hardcoded secrets: AWS keys, GitHub tokens, API keys, private key blocks, DB URLs, and more. 20+ patterns.

Inputs: path (required).

ssl_inspectA
Destructive

Inspect the full TLS certificate chain of a host. Returns expiry countdown, issuer chain, SANs, key size, and weak-config warnings.

Inputs: host (required), port.

supply_chain_auditA
Destructive

Audit a Node.js project's dependencies for supply chain attack indicators. Checks all packages in node_modules for: dangerous lifecycle scripts (postinstall that curl-pipe-to-bash, eval, base64 decode), typosquatting against 50+ popular package names (Levenshtein distance 1), and eval() of runtime data in source files. Reads local filesystem — no remote service can inspect your node_modules.

Inputs: path (required).

sys_infoA
Destructive

Get detailed local system information: CPU model, RAM, disk space, uptime, network interfaces with IPs and MACs. Reads from local hardware — not available to any remote service.

Inputs: no named parameters.

tracerouteA
Destructive

Trace the network path to a host hop-by-hop using ICMP TTL probes. Shows every router between this machine and the target. Requires raw packet sending — Claude cannot do this.

Inputs: host (required), max_hops.

usb_devicesA
Destructive

List all USB devices currently connected to this machine. Returns device name, vendor, product ID, speed, and manufacturer. Reads the local USB bus — no remote service can enumerate your physical ports.

Inputs: no named parameters.

wifi_scanA
Destructive

Scan nearby Wi-Fi networks using local wireless hardware. Returns SSID, BSSID, signal strength, channel, and security type. Requires physical Wi-Fi hardware — impossible for Claude to do remotely.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 23 tools

    Disambiguation5/5

    Each tool targets a specific, well-defined task with clear descriptions. For example, arp_scan, ping_sweep, port_scan, bluetooth_scan, and wifi_scan all perform different types of scans with distinct outputs. Overlap is minimal and justified.

    Naming Consistency4/5

    Most tools follow a two-word underscore pattern, but the order varies: some are verb_noun (audit_headers, check_cves) while others are noun_verb (arp_scan, dns_enum). This inconsistency is minor and does not significantly hinder readability.

    Tool Count5/5

    With 23 tools, the server covers a broad spectrum of security diagnostics (network scanning, system monitoring, vulnerability checks, etc.) without being overwhelming. Each tool contributes a unique capability, and the count is well-suited for the server's purpose.

    Completeness4/5

    The tool set is highly comprehensive for local security auditing, covering network discovery, system integrity, dependency vulnerabilities, and secret scanning. Minor gaps exist (e.g., no dedicated log analysis or vulnerability scanning beyond port/SSL), but core workflows are well-covered.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Integrates 15+ static application security testing tools (Semgrep, Bandit, TruffleHog, etc.) with Claude Code AI, enabling automated vulnerability scanning and security analysis through natural language commands. Supports cross-platform operation with remote execution on dedicated security VMs.
      7
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Turns Claude into a penetration testing copilot with 80 security tools, safety controls, and automatic reporting.
      80
      65
      MIT
    • F
      license
      Not graded
      quality
      C
      maintenance
      Securely feeds summarized expert security rules into your coding assistance Claude Code, Cursor, etc — zero config, no API key.
      -