The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Seo Performance MCP listing page.
Know which blog posts to refresh, expand, merge, or kill - without guessing.
A MCP server that turns your scattered SEO and analytics data into one clear verdict per URL. Plug it into Claude, Cursor, or any MCP-aware client and ask: "Which three posts should I update this week?" - and get an answer backed by hard numbers.
seo-performance-mcp unifies post-publish signals from every channel you already pay for:
It then runs a deterministic rule engine over those signals and emits a verdict per URL:
refresh/expand/merge/kill/double_down/hold
with reason codes, evidence, and a 0-1 confidence score. Reporting only - the server never mutates your posts.
Most content teams have analytics in five tabs and a gut feeling. That's how good posts rot quietly, mediocre posts get over-promoted, and the obvious "rewrite this one" is invisible until traffic has already cratered.
This MCP closes the loop:
src/verdict/rules.ts.After one cohort run you have:
In a Claude, Claude Code, or Cursor MCP config:
Every env var is optional. Adapters that lack their env config skip their slice of the snapshot; the server still boots. The verdict engine works on whatever slices are present.
Point it at any site, no CMS plugin required. The post-discovery layer resolves in priority order:
POSTS_LIST - JSON array of {url, title?, published_at?, tags?, word_count?}. Use this when you already have a content index and want exact control.GHOST_ADMIN_API_URL and GHOST_ADMIN_API_KEY are set, Ghost is used as a richer metadata source. Optional.og:title, article:published_time, and JSON-LD datePublished are read live from the URL.POSTS_SITEMAP_URL to your sitemap (or sitemap index) and the server enumerates posts from <loc> + <lastmod>.Most users only need POSTS_SITEMAP_URL. WordPress, Hugo, Astro, Next.js, Webflow, Framer, Wix, Squarespace, Notion-as-a-site, Substack-mirror sites all expose a sitemap by default.
To add a brand-new platform: nothing to build - just point POSTS_SITEMAP_URL at it.
| Tool | What it returns |
|---|---|
posts_list | Posts with {url, title, age_days, tags} from sitemap, Ghost, or your POSTS_LIST. |
posts_snapshot | Per-URL unified rollup for a 30/60/90-day window: GSC + Matomo + GA4 + Clarity + citations + meta. |
posts_decay_curve | Weekly GSC clicks/impressions/position buckets + a decay/plateau/growth trend label. |
posts_verdict | Verdict (refresh/expand/merge/kill/double_down/hold) + reason codes + 0-1 confidence. |
posts_refresh_brief | Markdown brief for a human or downstream LLM editor: numbers, top queries, suggested actions. |
cohort_report | Cohort verdict table sorted by priority + confidence. "Which three posts should I refresh this week?" |
posts_cite_loss | LLM citations that dropped off for a given URL. Needs CITATION_INTELLIGENCE_URL. |
gsc_quick_wins | (page, query) pairs at positions 5-15 with low CTR - fastest title-rewrite wins. |
Run any of the tools on a cron from CI and post the output to a GitHub Issue, Discussion, or PR. The action is published on the GitHub Marketplace.
Outputs:
| Output | Description |
|---|---|
result | Tool output as a multi-line string (markdown or JSON, per format). |
result-file | Path of the file the tool output was written to. Hand to peter-evans/create-issue-from-file etc. |
rows | For cohort_report with format: json only: number of rows returned. |
A complete weekly-audit workflow that opens a GitHub Issue with the cohort report is in examples/weekly-cohort-report.yml.
The package also ships a seo-perf-cli bin so you can run a single tool without an MCP client:
Same env vars as the MCP server. --format markdown is supported for cohort_report and posts_refresh_brief; other tools fall back to fenced JSON.
Three thin routing files ship in the repo so the LLM in your client knows when to reach for these tools:
skills/seo-performance/SKILL.md - tool-routing skill. Drop into ~/.claude/skills/seo-performance/ (or .claude/skills/ per project) to auto-load in Claude Code. Routes a single question to the right tool.skills/weekly-audit/SKILL.md - one-shot weekly audit playbook. Composes gsc_quick_wins + cohort_report + posts_cite_loss into a deduped, cross-signal ranked digest with proposed edits per URL. Drop in alongside the routing skill.cursor/rules/seo-performance.mdc - copy to .cursor/rules/seo-performance.mdc in any Cursor workspace.All optional. The MCP server works without them; they just shorten the "which tool do I call" round-trip.
The server exposes three prompts that bundle the playbook. Any MCP client (Claude Desktop, Claude Code, Cursor, Continue) can list and invoke them:
| Prompt | What it runs |
|---|---|
audit_cohort | cohort_report on posts >=90d, then posts_refresh_brief per refresh/expand/merge row. The weekly audit. |
find_quick_wins | gsc_quick_wins (positions 5-15) + per-URL posts_snapshot, then proposes verbatim-query meta_title rewrites. |
citation_loss_sweep | posts_cite_loss per URL, refresh_brief for any with losses, targeted H1/lead phrasing recommendations. |
Deterministic, rule-based, traceable. Reason codes:
ctr_below_position_expectedposition_driftdecay_30d_over_30pct / decay_60d_over_50pctstagnant_no_clicksthin_content_low_dwellrising_impressions_low_ctr / rising_clicks_continue_investmentcitation_loss / citation_growthduplicate_or_cannibalizinghigh_bounce_low_scrollfresh_post_too_youngThe mapping (reasons → verdict) and every threshold lives in src/verdict/rules.ts. Edit it, pin it in tests, ship your own rule book.
MIT