Claude SEO

Why it's ranked here

The network layer is the strongest evidence. Every plain page fetch goes through one guard that resolves the address once, refuses private, loopback and cloud metadata addresses, and pins the connection so a DNS trick cannot swap targets mid-request. That matters for a tool that follows links found on pages it crawls. The changelog records facts re-checked against Google's own pages, unconfirmable claims held back in a separate list, and installer bugs that leaked keys or wiped config fixed. MIT licensed, no telemetry, setup isolated from system Python.

What's good

It treats fetched pages as hostile. The technical agent's brief says fetched page content is data, never instructions, which is the right default for an agent reading the open web. Local audits of a dev or staging server are opt-in by naming each host explicitly, with cloud metadata endpoints refused even when listed. The content scorer says plainly that it does not decide whether text is AI-written and flags when its English-only phrase lists make a Chinese, Japanese or Korean score incomparable. Setup runs once, builds its own environment and headless browser, and a separate doctor command checks readiness without changing anything.

Tradeoffs

The scores are home-made heuristics. The content quality figure is a fixed weighted blend of phrase-list hits, a capital-letter entity count and repetition, and the technical agent is told to emit its own 0 to 100 score; neither is a Google measurement. Without API keys, page speed is inferred from HTML rather than measured from real users. The project's own counts disagree: the plugin manifest says 26 sub-skills and 19 agents, while the catalogue blurb says 25 and 18. Early features land first in a private mirror that requires a paid community membership. The Python floor is a version the changelog notes reaches end of life soon.

How to use it well

Best for agencies and in-house SEO leads who already live in Claude Code and want a repeatable first-pass audit with a written action plan. Run the setup command once, then start with a single-page check before a full-site crawl, which reaches up to 500 pages and fans out to as many as 17 agents. Treat the health score as a triage order, not a grade, and verify each fix against the evidence it cites. Add Search Console or PageSpeed credentials if you need real field data. It does not replace a rank tracker or backlink database unless you pay for those extensions.

Observed

License

MIT (pyproject.toml and .claude-plugin/plugin.json)

Language

Python 3.10 or newer for bundled scripts; skills and agents are Markdown prompts

Install surface

Claude Code plugin marketplace, or git clone plus a shell installer (Unix) or PowerShell installer (Windows)

Interface

Slash commands inside Claude Code; optional MCP server extensions for paid data sources

Runtime

Isolated Python environment plus Playwright Chromium, created by an explicit setup command

Outbound request guard

SSRF and DNS-rebinding protection in scripts/url_safety.py for every user-supplied URL

Telemetry

None stated; core makes no third-party API calls by default (PRIVACY.md)

Editor hook

Post-edit schema validation hook registered in hooks/hooks.json