Unchore Defend MCP server

Unchore Defend

Security help for the person protecting the system. Two tools, no install, works in Claude Code, Claude Desktop, any MCP client, or a plain terminal.

English · 한국어

July 2026: an AI agent broke out of its test sandbox and got into Hugging Face's production systems. Hugging Face counted about 17,600 attacker actions across 11 nodes over roughly 4.5 days. When the responders asked commercial frontier models to help read the logs, the requests were "blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker." They finished the forensics on an open-weight model (GLM-5.2).

Sources: disclosure · technical timeline

Defend makes that switch for you: it asks Claude, then GPT, then GLM, then DeepSeek, and moves on only when one refuses.

$ node skills/site-check/scripts/site-check.mjs example.com
https://example.com/ → F (37/100) · HTTP 200

✓ Served over HTTPS
✗ HSTS (browsers stay on HTTPS) (−15) — fix: Add `Strict-Transport-Security: max-age=31536000; includeSubDomains`.
✗ Content-Security-Policy set (−15) — fix: Add a Content-Security-Policy; start with `default-src 'self'` …
…

Install

Claude Code (skills + MCP tools):

/plugin marketplace add smilemino/unchore-defend
/plugin install unchore-defend@unchore

Then just ask: "check the security headers of mysite.com and fix what you can" or "read access.log — were we attacked?"

Only want the site check? /plugin install unchore-site-check@unchore installs just that (no key, no AI calls).

The site check needs no key. For defend, Claude Code asks for an OpenRouter or Unchore key when you enable the plugin (later: /plugin configure unchore-defend@unchore); it is kept in your system's secure storage.

Claude Desktop: download unchore-defend.mcpb from Releases and open it.

Any MCP client (Cursor, VS Code, Windsurf, …) — clone this repo, then add:

{
  "mcpServers": {
    "unchore-defend": {
      "command": "node",
      "args": ["/path/to/unchore-defend/mcp/server.mjs"],
      "env": { "OPENROUTER_API_KEY": "sk-or-…" }
    }
  }
}

Terminal only: Node 22+, nothing to install.

node skills/site-check/scripts/site-check.mjs mysite.com
node skills/unchore-defend/scripts/defend.mjs "Who got in, and what did they take?" --file access.log
tail -n 500 access.log | node skills/unchore-defend/scripts/defend.mjs "Is this an attack?"

Site check

Nine checks, weighted: HTTPS (25), HSTS (15), Content-Security-Policy (15), clickjacking protection (10), nosniff (10), Referrer-Policy (8), Permissions-Policy (5), cookie flags (7), server version hidden (5). A ≥ 90 · B ≥ 75 · C ≥ 60 · D ≥ 40.

It sends one ordinary GET to the site and nothing anywhere else. Private, local and cloud-metadata addresses are refused, every redirect is checked again, and it stops after 3 redirects, 8 seconds or 200 KB.

Put the grade in your README — the output gives you the line:

-green)

The badge says what was measured: response headers. It is not a full security audit.

Shortened. The full README is on GitHub.