website-checks-mcp
# UnlockLive website checks (MCP server)
Two free, read-only checks you can run from Claude, Cursor or any MCP client:
| Tool | What it checks |
|---|---|
| `check_llms_txt` | Is the site's `/llms.txt` online and in the [llmstxt.org](https://llmstxt.org) format? Are links described and working? Does `robots.txt` block GPTBot, ClaudeBot, PerplexityBot or other AI crawlers? No file yet? You get a draft built from the site's home page and sitemap. |
| `check_mcp_server_security` | Is a remote MCP server safe to connect? HTTPS and certificate, login required, OAuth discovery and PKCE, CORS, HSTS and version leaks. For open servers: the tool list, risky tools without safety hints, tool-poisoning phrases, and missing Origin checks. It never calls a tool. |
Both tools are marked read-only. They call the same public endpoints as the web tools ([llms.txt checker](https://www.unlocklive.com/tools/llms-txt-checker), [MCP security checker](https://www.unlocklive.com/tools/mcp-security-checker)), which only read public files and never log in. Only check servers you own or are allowed to test.
## Install
With [uv](https://docs.astral.sh/uv/):
```json
{
"mcpServers": {
"unlocklive-checks": {
"command": "uvx",
"args": ["unlocklive-checks-mcp"]
}
}
}
```
Add this to your Claude Desktop config (`claude_desktop_config.json`) or your client's MCP settings, then restart the client.
From source:
```bash
pip install .
unlocklive-checks-mcp
```
## Try it
- "Check the llms.txt of example.com and write me a better one."
- "Is https://mcp.example.com/mcp safe to connect? What should we fix first?"
## Limits
- The checks look at a site from the outside, like any visitor. They don't replace a security review.
- The public endpoints are rate limited (a few checks per 10 minutes per IP).
## License
MIT
---
Made by [UnlockLive IT](https://www.unlocklive.com). We build and secure MCP servers: [MCP server development](https://www.unlocklive.com/services/mcp-server-development).
TDQS
Scored across 2 tools
The two tools target entirely different artifacts: a website's llms.txt/robots.txt versus a remote MCP server's security posture. There is no plausible way for an agent to confuse one for the other, and the inputs (domain vs. HTTPS URL) are clearly disjoint.
Both follow the same check_<subject> snake_case pattern, and the subject portion is descriptive of exactly what is inspected. No mixed conventions or vague verbs.
Two tools is thin for a server branded as 'website-checks', which implies a broader catalog of checks. Each tool is substantial on its own, but the surface feels minimal rather than deliberately scoped.
Each check is self-contained and returns actionable output (score, problems with fixes, and even a generated draft llms.txt), so there are no dead ends within a single call. Gaps exist at the domain level — e.g., no general site hygiene, DNS/TLS or MCP functional/health checks — but these are minor for the apparent narrow purpose.