offsec-mcp

Offensive-security tooling for AI agents — over the Model Context Protocol.

Give Claude, Cursor, or any MCP client the ability to run reconnaissance, pull CVE intelligence, analyze JavaScript, and scan ports — only against targets you're authorized to test.

Why

Pentesters and bug-bounty hunters live in a dozen CLI tools. AI agents are great at orchestrating multi-step work — but they can't touch those tools. offsec-mcp bridges the gap: it exposes a focused set of security capabilities as MCP tools, so you can just ask —

"Enumerate subdomains of example.com, check which are live, and tell me if any run software with a critical CVE."

— and the agent chains recon_subdomains → http_probe → cve_search for you.

Built on battle-tested logic from reconwave, cve-forge, and jsintel.

Related MCP server: wrg-mcp-server

Guardrails first

This is a dual-use tool, so authorization is enforced in code, not left to good intentions:

• Passive OSINT (CT logs, DNS, RDAP, NVD, analyzing pasted JS) runs freely — it never touches the target.

• Active tools (port_scan, http_probe, security_headers) refuse to run unless the target is in your configured authorization scope. No scope → no active probing.

• Discovered secrets are redacted in output, never returned in full.

See SECURITY.md for the responsible-use policy.

Tools

* analyze_js is passive on pasted content; fetching a URL requires scope.

Install

Run instantly with uv — no clone needed:

uvx offsec-mcp        # runs the server (stdio) — or: pip install offsec-mcp

Or install from source:

git clone https://github.com/nadirzhon/offsec-mcp
cd offsec-mcp
uv pip install -e ".[dev]"

Configure your MCP client

Claude Desktop / Claude Code

Add to your MCP config (see examples/claude_desktop_config.json):

{
  "mcpServers": {
    "offsec": {
      "command": "uvx",
      "args": ["offsec-mcp"],
      "env": {
        "OFFSEC_SCOPE": "example.com,*.staging.example.com,10.0.0.0/24"
      }
    }
  }
}

Authorization scope

Active tools only run against hosts you list. Two ways to set it:

# inline — domains, wildcards, and CIDRs
export OFFSEC_SCOPE="example.com,*.staging.example.com,10.0.0.0/24"

# or a file, one entry per line (# comments allowed)
export OFFSEC_SCOPE_FILE="./scope.txt"

For a fully-owned lab you control, OFFSEC_ALLOW_ANY=1 disables the guard. NVD_API_KEY (optional) raises NVD rate limits.

Example session

You:   Find subdomains of example.com and check the main site's security headers.

Agent: → recon_subdomains("example.com")      → 143 subdomains from CT logs
       → security_headers("https://example.com")
       → Grade C (60): missing CSP, Permissions-Policy.
         www, api, staging, dev.example.com are live.

See examples/sample_output.md for real tool output.

Development

uv pip install -e ".[dev]"
pytest            # scope guard + tool logic
ruff check .

Part of an AI × Security toolkit

• offsec-mcp — MCP server giving AI agents offensive-security tools (recon, CVE, JS analysis) · (this repo)

• specter — autonomous AI recon agent that drives those tools end-to-end

• vigil — AI security review for every pull request

• mcpscan — security scanner for MCP servers (tool poisoning, injection surfaces)

• State of MCP Security — research: 87% of scanned MCP servers expose a medium+ hardening issue

License

MIT — see LICENSE. Published for authorized security testing, research, and education only. You are responsible for having explicit permission to test any target.