mcpmap

Network scanner that discovers and security-audits MCP servers across IP ranges. Like nmap, but for Model Context Protocol.

Install

# Homebrew (macOS/Linux)
brew install canack/tap/mcpmap

# Cargo
cargo install mcpmap

# Shell one-liner (macOS/Linux)
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/canack/mcpmap/releases/latest/download/mcpmap-installer.sh | sh

# Docker
docker run --rm ghcr.io/canack/mcpmap:latest 192.168.1.0/24

Pre-built binaries for all platforms on GitHub Releases.

Usage

mcpmap 192.168.1.0/24                          # discover MCP servers
mcpmap 192.168.1.0/24 --enumerate              # list tools + risk levels
mcpmap 192.168.1.0/24 --enumerate --active     # active security audit (safe, metadata-only)
mcpmap 192.168.1.0/24 --enumerate --active \
  --probe-tools --i-accept-risk                # call LOW-risk tools
mcpmap 192.168.1.0/24 --enumerate --active \
  --pin baseline.json                          # save tool hashes
mcpmap 192.168.1.0/24 --enumerate --active \
  --verify baseline.json                       # verify against baseline
mcpmap 192.168.1.0/24 --enumerate --json       # JSON output for CI/CD

How It Works

Targets ──▶ Port Scan ──▶ HTTP Probe ──▶ MCP Handshake ──▶ Security Checks
                                              │
                                     --enumerate: Tool Enumeration
                                              │
                                        --active: Behavioral Probing

Passive (default): Port scan → HTTP heuristic → JSON-RPC initialize handshake → origin/auth/session validation → tool enumeration with risk classification.

Active (--active): Schema poisoning, resource injection, tool squatting, temporal rug-pull, exfiltration chains, cross-server manipulation, response injection, denial-of-wallet.

Active Probing Tiers

CRITICAL/HIGH risk tools are NEVER called. No flag overrides this.

Safety limits: 3 calls/tool, 50 calls/server, 5s/call timeout, 100KB response cap, 120s total probe timeout.

Security Detections

15 vulnerability classes across passive and active scanning:

Example Output

MCPMAP - 256 targets scanned in 3.42s
────────────────────────────────────────────────────────────
192.168.1.10:3001 [CONFIRMED 100%]
  Server: filesystem-server/1.2.0 (MCP 2025-11-25)
  Capabilities: tools, resources
  Transport: SSE @ /mcp
  Origin: NOT VALIDATED (DNS rebinding risk)
  Tools: 5 discovered [HIGH RISK]
    - read_file [Medium] - Read file contents
    - write_file [High] - Write to files
    - delete_file [Critical] - Delete files
  Active Probe: Tier 2 executed, 3 findings
    MCP-008 [HIGH] Schema poisoning: parameter 'content_include_env_vars'
    MCP-010 [CRITICAL] Rug-pull: get_weather description changed after 3 calls
    MCP-012 [MEDIUM] analyze_data returned 102KB (threshold: 50KB)

192.168.1.15:8000 [LIKELY 75%]
  Server: unknown (MCP 2024-11-05)
  Capabilities: tools
  Transport: HTTP @ /
  Auth: Required (Bearer)
────────────────────────────────────────────────────────────
2 MCP servers | 1 confirmed | 1 likely | 1 auth required

Options

Scanning:
  -m, --mode <MODE>      fast | full | stealth         [default: fast]
  -p, --ports <PORTS>    port range (80,443 or 1-1000)
  -t, --threads <N>      concurrent threads            [default: 50]
      --timeout <SECS>   connection timeout             [default: 5]

Probing:
      --enumerate        list tools on confirmed servers
      --active           active security probing (safe, metadata-only)
      --probe-tools      call LOW-risk tools with test inputs
      --probe-medium     also call MEDIUM-risk tools (pentest only)
      --dry-run          show probe plan without executing
      --i-accept-risk    explicit consent for tool invocation
      --pin <FILE>       save tool/resource hashes
      --verify <FILE>    verify against a saved pin file

Output:
      --json             JSON output
  -W, --wide             wide table (grep-friendly)
  -q, --quiet            suppress progress output
  -v, --verbose          verbose (-v, -vv, -vvv)

Scanning:
      --deep-probe          try extra endpoints (/sse, /api/mcp, /v1/mcp)
      --scheme <S>          http | https | both         [default: both]
      --insecure            accept invalid TLS certs
      --max-targets <N>     max IP:port combinations    [default: 1000000]
      --rate-limit <N>      max req/sec (0 = unlimited) [default: 0]

Output:
      --show-all            include low-confidence results
      --min-confidence <N>  minimum confidence 0-100    [default: 0]

Exit Codes

License

MIT