Website Audit MCP
# website-audit-mcp
An MCP server (stdio) that fetches a web page over HTTP and scores it in five categories: performance, SEO, design, functionality and technology. It uses plain HTTP requests and cheerio. It needs no browser, no database and no API key.
## What it checks
| Category | Checks (weight in overall score) |
|---|---|
| performance (25) | response time, HTML size, compression header, number of external scripts, redirects, HTTP status |
| seo (25) | title, meta description, `<h1>` count, `lang` attribute, canonical link, image alt text, noindex |
| design (20) | viewport meta tag, favicon link, deprecated presentational elements, charset, inline styles, table count |
| functionality (15) | links present, empty or `javascript:` links, unlabeled form inputs, HTTP status |
| technology (15) | HTTPS, TLS validity and expiry, security headers, server version disclosure, CMS and library detection |
Scores are 0-100. The overall score is the weighted average. Only the HTML document is analyzed; linked assets are not downloaded and JavaScript is not executed, so pages rendered client-side are scored on their initial HTML.
## Input limits
- Only `http:` and `https:` URLs. URLs with embedded credentials are refused.
- Hosts that resolve to loopback, private, link-local or otherwise reserved addresses are refused. Every redirect target is checked again, and the connection uses the validated address.
- 10 s timeout, 2 MB body cap (the body is truncated and reported as `truncated`), at most 5 redirects.
- `compare_websites` takes 2 to 10 URLs.
## Install
```sh
git clone <this repository>
cd website-audit-mcp
npm install
npm run build
```
Requires Node.js 18 or newer.
## Configure
Claude Code:
```sh
claude mcp add website-audit -- node /absolute/path/to/website-audit-mcp/dist/index.js
```
Claude Desktop (`claude_desktop_config.json`):
```json
{
"mcpServers": {
"website-audit": {
"command": "node",
"args": ["/absolute/path/to/website-audit-mcp/dist/index.js"]
}
}
}
```
Setting `WEBSITE_AUDIT_ALLOW_PRIVATE=1` disables the private-host check. It exists for the test suite and should not be set otherwise.
## Tools
### `audit_website`
Input:
```json
{ "url": "string (http or https URL)" }
```
Output (JSON text): `url`, `finalUrl`, `status`, `scores` (`performance`, `seo`, `design`, `functionality`, `technology`, `overall`), `findings[]` (`category`, `severity` of `info|warning|error`, `message`), `detected` (`cms`, `frameworks`, `server`), `metrics` (`responseTimeMs`, `htmlBytes`, `truncated`, `redirects`).
### `compare_websites`
Input:
```json
{ "urls": ["string", "string"] }
```
Output (JSON text): `ranked[]` (`rank`, `url`, `overall`, `scores`) sorted by overall score, highest first, and `errors[]` (`url`, `error`) for URLs that were refused or failed.
Invalid input returns a tool result with `isError: true` and a message.
## Example
`audit_website` with `{"url": "https://example.com"}` (findings shortened):
```json
{
"url": "https://example.com",
"finalUrl": "https://example.com/",
"status": 200,
"scores": { "performance": 100, "seo": 60, "design": 100, "functionality": 75, "technology": 84, "overall": 84 },
"findings": [
{ "category": "performance", "severity": "info", "message": "Response time 67 ms" },
{ "category": "seo", "severity": "error", "message": "Missing meta description" },
{ "category": "seo", "severity": "error", "message": "No <h1>" },
{ "category": "functionality", "severity": "warning", "message": "No links on the page" },
{ "category": "technology", "severity": "info", "message": "TLS valid (TLSv1.3), 55 days until expiry" },
{ "category": "technology", "severity": "warning", "message": "No Strict-Transport-Security header" }
],
"detected": { "cms": null, "frameworks": [], "server": "cloudflare" },
"metrics": { "responseTimeMs": 67, "htmlBytes": 577, "truncated": false, "redirects": 0 }
}
```
## Development
```sh
npm test # builds, then runs unit tests and the stdio end-to-end test
npx tsc --noEmit # type check
```
The tests use a local HTTP server as a fixture. The end-to-end test starts the built server over stdio and talks to it with the SDK client.
## Limitations
- The scoring rules are fixed heuristics, not measurements from a browser. They do not replace Lighthouse.
- Only the single URL given is fetched. No other pages of the site are crawled.
## License
MIT, see `LICENSE`.
TDQS
Scored across 2 tools
The two tools differ by cardinality: audit_website targets a single URL, compare_websites targets 2-10 URLs and returns a ranking. The boundary is mostly clear, though compare_websites is essentially a batch wrapper and an agent could plausibly reach for it when auditing one site plus a competitor.
Both names follow the same verb_noun snake_case pattern (audit_website, compare_websites). The verb directly signals the action and the noun the resource, so the convention is fully predictable.
Two tools is thin for a server claiming five audit dimensions (performance, SEO, design, functionality, technology). The surface is cohesive and each tool earns its place, but there is no tooling for targeted checks, configuration, or report retrieval, putting it at the borderline.
Single-site auditing and multi-site ranking are covered, so the core workflow has no dead ends. However, there is no way to request a specific dimension, retrieve or export a report, or re-examine prior results, which are notable gaps for an audit product.