recon-mcp

English | 繁體中文

An MCP server that gives AI coding agents — Claude Code, Codex, Cline, and any MCP client — safe, structured network and security reconnaissance tools.

Most MCP servers wrap CRUD APIs. recon-mcp instead exposes the kind of read-only recon an engineer reaches for when investigating an asset, and returns clean JSON — with a graded verdict — so the agent can reason over results instead of parsing console output.

⚠️ Authorized use only. These tools are for security testing of assets you own or have explicit written permission to assess, for CTF practice, and for education. Do not point them at third-party infrastructure without authorization. You are responsible for how you use this software.

Tools

Example

Just ask your agent: "run a security recon report on example.com." It calls recon_report once and gets a graded overview it can act on:

{
  "domain": "example.com",
  "overall_grade": "F",
  "summary": "Overall posture F: email A, TLS B, headers F; 13 actionable issue(s).",
  "components": {
    "email":   { "grade": "A", "issues": [] },
    "tls":     { "grade": "B", "issues": [] },
    "headers": { "grade": "F", "issues": [
      { "severity": "high", "label": "Missing Content-Security-Policy", "detail": "CSP not set; cannot restrict resource load sources" }
    ] }
  }
}

Need more detail on one area? The agent can call dns_recon, subdomain_enum, subdomain_takeover, tls_check, http_headers_audit, cookie_audit, cors_check, tech_detect, http_methods_audit, well_known_audit, ip_info, or port_scan directly.

Install

Requires Python ≥ 3.10. Runs on Linux, macOS, and Windows (tested in CI).

Recommended — no clone, via uv:

uvx recon-kit-mcp

Or from source (for development):

git clone https://github.com/nan786521/recon-mcp
cd recon-mcp
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate
pip install -e .

Use with Claude Code

Add the server (stdio transport). With uvx you don't need an absolute path:

claude mcp add recon -- uvx recon-kit-mcp

Or add it manually to any MCP client config:

{
  "mcpServers": {
    "recon": {
      "command": "uvx",
      "args": ["recon-kit-mcp"]
    }
  }
}

(From a source checkout, point the command at /absolute/path/to/.venv/bin/recon-kit-mcp instead.)

Then just ask: "run a security recon report on example.com" — or target one area, e.g. "check the email security of example.com."

The server also ships a security_recon prompt: pick it from your client's prompt menu and pass a domain for a guided, severity-sorted audit.

Tool reference

recon_report(domain, timeout?) -> dict

Runs DNS/email, TLS, HTTP-header, web-stack (tech_detect), and apex subdomain-takeover checks together and returns overall_grade (as weak as the weakest component, capped at F if a live takeover is found), a one-line summary, components (email / tls / headers, each with its grade and actionable issues), a tech section (detected technologies + any version disclosure), and a takeover section when the apex is at risk. Uses a fast single-handshake TLS check for speed — call tls_check for the full cipher/vulnerability analysis. The best starting point; use the tools below for raw detail.

dns_recon(domain, checks?, timeout?) -> dict

• records — A, AAAA, MX, NS, TXT, SOA, CNAME, CAA records
• whois — parsed registration fields + raw WHOIS text
• email — SPF, DMARC, and DKIM posture, plus advisory MTA-STS, TLS-RPT, BIMI, and DNSSEC signals, and a graded assessment (letter grade A–F, a summary, and per-check findings with severity and a recommended fix). The advisory signals surface as findings but don't move the core SPF/DKIM/DMARC grade.

checks is any subset of ["records", "whois", "email"]; omit it to run all.

subdomain_enum(domain, wordlist?, source="dns", timeout?) -> dict

Discovers subdomains from two complementary sources:

• source="dns" (default) — resolves candidate labels via DNS. wordlist is comma-separated labels ("www,api,dev"); omit it for a built-in common list. Capped at 512 candidates per call. Returns resolved ips.
• source="ct" — queries public Certificate Transparency logs (crt.sh) for every name ever certified for the domain. Fully passive; finds real hosts no wordlist would guess.
• source="both" — runs both and merges, recording which source(s) saw each host.

Returns sources, found_count, and found (each with subdomain, the sources that saw it, and ips when resolved).

subdomain_takeover(hosts, timeout?) -> dict

Checks subdomains for a dangling-CNAME takeover — a subdomain that CNAMEs to a third-party service (GitHub Pages, S3, Heroku, Azure, Fastly, Shopify, …) whose resource was deleted or never claimed, letting anyone who registers that resource serve content on the victim's subdomain. For each host it resolves the CNAME, recognizes known takeover-prone services, fetches the page, and flags the provider's "unclaimed resource" fingerprint and/or a CNAME target that no longer resolves. hosts is one hostname or a comma-separated list (capped at 100). Read-only — DNS lookups plus one HTTP GET per host. Pair it with subdomain_enum: enumerate first, then check the interesting hosts.

Returns checked, vulnerable_count, and results (each with host, cname, service, status, vulnerable, severity, and detail). status is one of not_applicable, not_vulnerable, potential, dangling_cname, or vulnerable.

tls_check(host, port=443, timeout?) -> dict

Returns grade, certificate (validity / expiry / key algorithm), protocols (flags legacy SSLv3 / TLS 1.0 / 1.1), cipher info, forward_secrecy, hsts, vulnerabilities (each with a vulnerable flag), and a findings list.

http_headers_audit(host, port?, use_ssl=True, timeout?) -> dict

Returns grade, score, the observed security headers, and a findings list with a recommendation per header. Defaults to HTTPS (port 443).

cookie_audit(host, port?, use_ssl=True, timeout?) -> dict

Follows the redirect chain from the host (capped at 10 hops, flagging any HTTPS→HTTP downgrade) and audits every Set-Cookie seen for the Secure, HttpOnly, and SameSite flags. Returns redirect_chain, final_url, cookies (flags only — values are never returned), cookie_grade, cookie_score, and a findings list.

cors_check(host, port?, use_ssl=True, timeout?) -> dict

Sends one GET with an untrusted Origin and inspects the Access-Control-Allow-Origin / -Allow-Credentials response. Reflecting an arbitrary Origin with credentials is high severity (any site can read authenticated responses); a wildcard or trusted null origin are lesser issues. Returns acao, allows_credentials, reflects_origin, wildcard, severity, and findings.

tech_detect(host, port?, use_ssl=True, timeout?) -> dict

Fingerprints the technology stack behind a website from one HTTP GET. It matches response headers, set cookies, the HTML body, and the <meta name="generator"> tag against a signature table to identify the web server, reverse proxy / CDN, WAF, programming language, web framework, CMS, JavaScript framework, and analytics. Where a version is exposed it is captured and flagged (info) — a precise version eases known-CVE lookup. Read-only.

Returns status, technology_count, technologies (each with name, category, version when known, and evidence), and a findings list noting any version disclosure.

http_methods_audit(host, port?, use_ssl=True, path="/", timeout?) -> dict