SiteVigil

     .      o                 .               o                       .
  ______  _____  _______  ______ __      __ _____   ______   _____  _
 / ____/ |_   _||__   __||  ____|\ \    / /|_   _| / ____/  |_   _|| |
| (___     | |     | |   | |__    \ \  / /   | |  | |  ___    | |  | |
 \___ \    | |     | |   |  __|    \ \/ /    | |  | | |_  |   | |  | |
 ____) |  _| |_    | |   | |____    \  /    _| |_ | |___| |  _| |_ | |____
|_____/  |_____|   |_|   |______|    \/    |_____| \_____/  |_____||______|
   |                                                |               |
   .                                                .              \___/

One command. A security audit of any website — scored, explained, and with the fix written out for you.

pip install sitevigil
sitevigil aqles.github.io

https://aqles.github.io
Score: 75/100  Grade: C  HTTP 200

[MEDIUM  ] Missing Content-Security-Policy header
           No CSP header; the page has no second line of defence against XSS.
           Fix: Add a CSP that fits your site, starting strict, e.g. Content-Security-Policy: default-src 'self'
[LOW     ] Missing X-Content-Type-Options: nosniff
           Header absent or not set to 'nosniff'.
           Fix: Add: X-Content-Type-Options: nosniff
[MEDIUM  ] No clickjacking protection
           Neither X-Frame-Options (DENY/SAMEORIGIN) nor CSP frame-ancestors is set.
           Fix: Add: X-Frame-Options: DENY (or SAMEORIGIN), or frame-ancestors 'none' in your CSP.

SiteVigil is a small Python tool with zero dependencies — standard library only. It runs the checks that turn up again and again when real websites get audited, it works as a plain CLI for humans, and it speaks MCP so AI coding agents (Claude Code, Cursor, …) can run the same audit and fix what it finds.

Why not just use a free web checker?

Fair question — sites like SecurityHeaders or Mozilla Observatory grade a page for free, and for a one-off look they're fine. SiteVigil exists for everything around that moment:

• It hands you the patch, not just the grade. --fix generates a ready-to-commit config for your actual platform (Vercel, Netlify, Cloudflare Pages, nginx, Apache). On GitHub Pages it tells you plainly that headers are a platform limit and gives you the portable _headers file instead of a lecture.
• It lives where you work. Terminal, CI (--fail-under, --baseline), and inside your AI agent via MCP — a web page can't fail your deploy or re-check a fix.
• It checks more than headers. Exposed .env/.git (verified by content, not status codes), CORS reflection, cookie flags, mixed content, missing SRI, TRACE, TLS expiry, security.txt.
• It remembers. Save a report, come back next week, and --baseline shows exactly what regressed and what got fixed — web checkers have no memory of your last scan.

What it checks

The full catalog — 30+ checks, each with its reference — is one command away: sitevigil --checks. Checks are versioned as a ruleset (currently 2026.11); every report names the ruleset that produced it, and changes land in CHANGELOG.md, never silently. To keep a site up to date automatically, drop in the weekly GitHub Action from examples/github-action: it re-audits on a schedule and fails on score drops or regressions.

Every finding includes what was actually observed and a concrete fix — not just a warning label.

Two details that took real-world pain to learn:

• No soft-404 false positives. Many sites (SPAs, some hosts) answer 200 with their homepage for every path. SiteVigil only reports an exposed file when the response content actually matches a real .env / git file — a bare 200 is never enough.
• CVE matches are verified, not googled. NVD search happily returns CVEs that only mention your software. SiteVigil re-checks every candidate against the CVE's own vulnerable configurations and version ranges before reporting it, scores it with NVD's primary CVSS, and prints the backport caveat next to every match — a version string alone never proves a server is unpatched.
• Proxy-aware TLS. If you sit behind a corporate proxy (https_proxy), the TLS check tunnels through it with CONNECT instead of dying with WRONG_VERSION_NUMBER.

Install

pip install sitevigil
# or straight from the repo:
pip install git+https://github.com/aqles/sitevigil
# or just run it from a clone, no install needed:
python -m sitevigil example.com

Requires Python 3.10+. That's it — nothing else gets installed.

Use it

sitevigil example.com                  # human-readable report
sitevigil example.com --json           # structured output, for scripts and agents
sitevigil example.com --markdown       # a report you can paste into an issue
sitevigil example.com --html > report.html   # one self-contained HTML report
sitevigil a.com b.com c.com            # audit several sites, with a summary table
sitevigil example.com --fail-under 80  # CI mode: exit 1 if the score is below 80
sitevigil example.com --fix            # + a ready-to-commit header config for your platform
sitevigil example.com --json > monday.json          # save a baseline
sitevigil example.com --baseline monday.json        # next week: what regressed? (exit 1 on new findings)
sitevigil example.com --crawl 10                    # audit up to 10 pages (sitemap.xml, else homepage links)
sitevigil example.com --badge > badge.svg           # an SVG score badge for your README
sitevigil example.com --no-cve                      # skip the CVE lookup (faster, fully offline)
sitevigil example.com --html --output report.html   # write any format straight to a file

Scoring starts at 100 and loses points per finding (critical 25, high 15, medium 8, low 3), ending in a grade from A to F.

Let your AI agent use it

SiteVigil ships an MCP server (still stdlib-only). Local agents use stdio; hosted agents can use the HTTP transport:

sitevigil --mcp                      # stdio, for desktop/CLI agents
sitevigil --mcp-http 8900            # HTTP on localhost, for hosted agents

MCP is a shared protocol, so one server works with every MCP-compatible agent — Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Gemini CLI — no per-agent builds. Only the config file's location changes; ready-to-paste snippets for each agent live in examples/mcp. Interoperability is not just claimed: the server is tested against the official MCP Python SDK client (initialize → tool list → a real audit_site call).

Your agent gets two tools: audit_site (the full scored report) and suggest_fix (the audit plus a ready-to-commit, platform-specific header config). It can check a site before telling you it's "secure" — or audit the fix it just deployed.

Honest limits

• It audits the response of the URL you give it (plus /.env / /.git probes). It is not a vulnerability scanner and it does not fuzz, brute-force, or log in anywhere.
• Only audit sites you own or have permission to test. The probes are read-only GETs, but scanning other people's property without permission is still a bad idea.
• A high score means these basics are in place. It is not a guarantee that a site is secure.

Development

git clone https://github.com/aqles/sitevigil
cd sitevigil
PYTHONPATH=src python -m unittest discover -s tests -v

The test fixture is a local HTTP server that deliberately serves a soft-404 for unknown paths and a real-looking .env file — the audit has to catch one and ignore the other.

License

MIT — see LICENSE. Use it, fork it, ship it; just keep the copyright notice. If SiteVigil saves you from shipping a site with an exposed .env, a star is the going rate.