SEO Agent Kit

A self-improving SEO agent for Claude Code, from Creator OS (the marketing system behind KevBuildsApps). It writes blog posts that rank on Google and get cited by ChatGPT, Claude and Perplexity, reads PostHog to see which posts bring signups, and writes more like the winners. Fully open source.

What you need

Three connections:

1. Creator OS API connected to a WordPress site. This is where the agent publishes. Sign up at creatoros.ca, connect WordPress (self-hosted or WordPress.com), copy the API key.
2. PostHog API. A personal key (phx_) with Query: Read, and PostHog installed on your site.
3. Google Search Console API. A service account added as an Owner of your property.

Plus the basics any agent needs: a Postgres database (Supabase, Neon, Railway, local) and a model key (Claude, or any OpenAI-compatible API). Then follow Setup below.

What's inside

• The SEO agent (.claude/skills/seo-engine, wordpress-blog, ai-search-files). See The SEO agent below.

• The PostHog analytics kit the agent learns from:

• A twice-daily snapshot job that pulls from PostHog (HogQL Query API) into Postgres:

  • visitors, pageviews and sessions per day
  • your signup / onboarding / activation funnel events
  • where each session came from (referrer or UTM)

• A live "today" top-up that keeps today's numbers current between snapshots.

• Traffic-source grouping of raw referrers into families: YouTube, Google, Threads, Instagram, X, LinkedIn, TikTok, ChatGPT, Claude, Perplexity, Gemini, Copilot, other search, internal and direct.

• Dashboard cards (React + Tailwind):

  • Web chart: visitors with signups and subscribers layered on top, same bar width, so the height shows conversion
  • Web funnel
  • Traffic sources, with Today / Yesterday / 7 days / 30 days tabs

Every day is an America/New_York day, never UTC. That stops evening visits from landing on tomorrow's date. Change America/New_York everywhere if you want another zone.

Why snapshots?

Charts read from your own Postgres, not from PostHog on every page load. That makes pages fast and costs no PostHog query credits. You also keep history even if you change PostHog projects.

Setup (5 minutes)

1. Install PostHog on your site. Also capture your funnel events. Rename them in the queries if yours differ.

   copy

   posthog.capture("user_signed_up")
   posthog.capture("onboarding_completed")
   posthog.capture("api_key_copied")   // your "activated" moment
   

2. Create a personal API key in PostHog (Settings > Personal API keys) with Query: Read.

3. Configure and run:

   copy

   cp .env.example .env     # fill DATABASE_URL, POSTHOG_API_KEY, POSTHOG_PROJECT_ID
   npm install
   npm run backfill         # last 90 days
   npm run snapshot         # then schedule this at 9 AM + 9 PM
   

   You can schedule it with any cron: Railway cron, GitHub Actions, crontab, or Vercel cron hitting a route that runs it.

   The tables are created on first run. sql/schema.sql has them if you'd rather create them yourself.

Using the dashboard pieces (Next.js App Router)

Copy lib/ and components/ into your app. You need pg, recharts and Tailwind.

// app/dashboard/page.tsx  (server component)
import { getWebVisitorsDaily } from "@/lib/posthog-client";
import { getWebSourcePeriods } from "@/lib/posthog-sources";
import { WebTrafficChart } from "@/components/WebTrafficChart";
import { WebFunnel } from "@/components/WebFunnel";
import { TrafficSources } from "@/components/TrafficSources";

export default async function Page() {
  const days = await getWebVisitorsDaily("default", process.env.POSTHOG_PROJECT_ID!, process.env.POSTHOG_HOST ?? "https://us.posthog.com", 30);
  const periods = await getWebSourcePeriods("default");
  const sum = (k: keyof (typeof days)[number]) => days.reduce((a, d) => a + Number(d[k]), 0);

  return (
    <>
      <WebTrafficChart
        variant="web"
        data={days.map((d) => ({ date: d.date, count: d.signups }))}
        traffic={days.map((d) => ({ date: d.date, count: d.visitors }))}
        signups={days.map((d) => ({ date: d.date, count: d.signups }))}
        subscribers={[] /* paid subs per day from Stripe / RevenueCat, if you have them */}
      />
      <WebFunnel steps={[
        { label: "Visitors", value: sum("visitors"), color: "#64748b" },
        { label: "Signed up", value: sum("signups"), color: "#8b5cf6" },
        { label: "Onboarded", value: sum("onboarded"), color: "#06b6d4" },
        { label: "Activated", value: sum("apiKeys"), color: "#10b981" },
      ]} />
      <TrafficSources periods={periods} />
    </>
  );
}

The components use a few utility classes from the original design system: card, card-title, num and eyebrow. Define them in your global CSS, or swap them for your own.

Tracked links (bonus pattern)

To see which post drove a visit, send links through UTM-tagged URLs, e.g. https://yoursite.com/?utm_source=threads&utm_campaign=post-123. UTM sources win over the referrer in the source grouping. ChatGPT, for example, sends utm_source=chatgpt.com with no referrer.

The SEO agent (blog engine)

The same PostHog data drives a self-improving SEO loop for a WordPress blog. It writes posts that rank, reads PostHog to see which posts bring signups, and writes more like the winners. It is three Claude Code skills in .claude/skills/:

 FEEDS                 PIPELINE (publish, 1 post per run)              LEARN (measure, nightly)
 spin-offs ─┐
 GSC gaps  ─┤  scout   brief -> write -> fact gate -> publish -> IndexNow   PostHog entry visitors,
 news RSS  ─┼──────>   (2 rewrites max, else rejected with a reason)        signups, AI referrals
 releases  ─┤  queue                                                      + Search Console clicks
 seeds     ─┘    ^                                                                  |
                 └──────── cluster weights (signups per post, 0.5x to 3x) <─────────┘

Setup

1. Connect your WordPress site in Creator OS and put that workspace's API key in .env as CREATOROS_API_KEY (see .claude/skills/wordpress-blog/SKILL.md).
2. cp seo.config.example.json seo.config.json and fill in your site, brand, signup path and signup event.
3. In .claude/skills/seo-engine/: cp facts.example.json facts.json and cp seeds.example.json seeds.json. facts.json is the product truth the writer may use. Keep it accurate; it is what stops invented features.
4. Set a model: ANTHROPIC_API_KEY, or LLM_BASE_URL + LLM_API_KEY + LLM_MODEL for any OpenAI-compatible API. Optional: POSTHOG_AI_KEY logs every generation to PostHog LLM analytics.
5. Optional: Search Console service account (GSC_SA_JSON_B64) and INDEXNOW_KEY.

npm run seo:scout -- --dry-run     # what would it queue?
npm run seo:publish -- --dry-run   # write + fact-check one post to a temp file, publish nothing
npm run seo:scout && npm run seo:publish && npm run seo:measure

Schedule (ET): scout 7:30, publish hourly 8:00 to 17:00 (capped by daily_max), measure 22:00. SEO_PUBLISH_AS_DRAFT=1 lands posts as drafts if you want to review first.

Full details, guard rails and debugging queries: .claude/skills/seo-engine/SKILL.md.

Use it with Claude Code

Everything in .claude/skills/ is a Claude Code skill: posthog-analytics, seo-engine, wordpress-blog and ai-search-files. Copy the folders into your project's .claude/skills/ and Claude will know how to set up, run, debug and extend each piece.

Files

Built by Kevin (KevBuildsApps on YouTube). Learn to build systems like this in Kev's No Code Academy.

MIT licensed.