Most SEO audit skills are built for Claude Code. If you work in DeepSeek Harness (DSH), that means leaving your environment — or making do with a partial check. seo-audit closes that gap: a complete, weighted SEO audit that runs natively in DSH, entirely on your machine.
DeepSeek Harness is plugin-first, but a serious SEO audit needs both deterministic measurement and judgment — crawlability, indexability, schema, E-E-A-T, local/GBP, GEO/AI-Overview readiness, Core Web Vitals. seo-audit brings all of it into the harness you already use, with the audit logic anchored in Google's primary-source guidance.
| Layer | What it is | How it runs |
|---|---|---|
| Measurement | 5 Python specialists + 53 curated scripts | ./seo <command> |
| Judgment | 24 sub-skills + 18 specialists | Agent executes via subagent / workflow |
git clone https://github.com/Haniubub/seo-toolkit.git
cd seo-toolkit
./setup.sh./setup.sh installs the pinned Python deps into a workspace-local pylibs/ and Playwright Chromium into browsers/ — nothing touches your global Python or system.
./seo doctor # sanity check
./seo audit https://example.com # full weighted audit./seo audit detects the business type, then fans out to only the relevant sub-agents in parallel (never all 18), and merges the results into one weighted score. Every recommendation carries the four claude-seo fields:
Observation → Dependency → Failure signal → Early indicator.
Unlike SEO skills tied to Claude Code, seo-audit is DSH-native: it runs in the harness you already use, no Claude Code required, no API key for the core audit, and the measurement layer is pure local Python. Everything stays on your machine.
| Claude Code SEO skills | seo-audit | |
|---|---|---|
| Environment | Claude Code | DeepSeek Harness (DSH) |
| Claude Code required | Yes | No |
| Core audit API key | Varies | None |
| Runs locally | — | Yes |
A weighted SEO health score (Technical 22% · Content 23% · On-Page 20% · Schema 10% · Performance 10% · AI-Readiness 10% · Images 5%), a prioritised action plan sorted by dependency, and a falsifiability check per recommendation — so you know whether each fix actually worked.