I built HostAI in 48 hours — Cloudflare Worker + Claude AI + Stripe (full stack breakdown)

I've been building a series of AI micro-tools under the name BaseAI. The latest: HostAI, a free listing optimizer for Airbnb, Booking.com and Vrbo. Here's exactly how it's built, including the parts that surprised me.

The stack (zero monthly cost)

• Cloudflare Workers — API, AI calls, rate limiting, Telegram cron
• Cloudflare KV — rate limiting (5 free req/day per IP) + Pro token storage
• Cloudflare Pages — static frontend
• Anthropic Claude — Haiku for quick optimizations, Sonnet for full audits
• Brevo — transactional email (300/day free tier)
• Stripe Payment Links — no checkout page needed

Total fixed monthly cost: $0. It runs on free tiers until traffic justifies upgrading.

Rate limiting without a database

The key design decision: no user accounts, no database, no friction. Three free optimizations per day per IP, verified via KV:

async function checkRateLimit(env, ip, type) {
  const limits = { optimize: 3, 'mini-audit': 1 };
  const today = new Date().toISOString().split('T')[0];
  const key = `rate:${type}:${ip}:${today}`;
  const count = parseInt((await env.HA_RATE.get(key)) || '0');
  if (count >= limits[type]) return { allowed: false };
  await env.HA_RATE.put(key, String(count + 1), { expirationTtl: 86400 });
  return { allowed: true };
}

KV keys auto-expire after 24h. No cleanup needed. No cron job. No database.

The Pro activation flow (no webhook needed)

Stripe webhooks are overkill for this use case. The approach: Stripe redirects to the success URL with ?session_id=cs_live_xxx. The frontend POSTs that to /activate-pro, the worker validates the cs_live_ prefix and generates a permanent token:

if (!sessionId?.startsWith('cs_live_')) return json({ error: 'Invalid session' }, 400);
const existing = await env.HA_TOKENS.get('session:' + sessionId);
if (existing) return json({ token: existing, reused: true });
const token = 'ha_pro_' + crypto.randomUUID().replace(/-/g, '');
await env.HA_TOKENS.put('session:' + sessionId, token);
await env.HA_TOKENS.put('token:' + token, JSON.stringify({ email, activated: new Date().toISOString() }));

Idempotent — replaying the same session_id returns the same token. The user gets it by email via Brevo.

Async audit delivery via ctx.waitUntil

The full HostScan audit (10 sections, Sonnet model) takes 15-30 seconds. Too long for a synchronous response. Solution: respond immediately, run the audit in the background:

ctx.waitUntil((async () => {
  const audit = await runFullAudit(env, body);
  const html = buildAuditEmail(audit, body.title);
  await sendEmail(env, body.email, `HostScan: ${body.title}`, html);
})());
return json({ success: true, message: 'Check your email in ~15 minutes.' });

ctx.waitUntil keeps the Worker alive after the response is sent. The user gets instant confirmation, audit arrives by email.

Telegram daily cron (passive audience building)

# wrangler.toml
[triggers]
crons = ["0 8 * * *"]

export default {
  fetch: handleRequest,
  async scheduled(event, env, ctx) {
    ctx.waitUntil(postDailyTip(env));
  }
};

Every morning at 08:00 UTC, one actionable Airbnb hosting tip goes out to the Telegram channel. No manual work. The channel grows while I sleep.

What surprised me

Platform-specific prompting matters more than I expected. Airbnb titles max out at 50 characters. Booking.com weights location keywords heavily. Vrbo leans toward family/group language. Giving Claude explicit platform context in the prompt increased output quality noticeably — generic prompts produce generic results.

The audit model selection is important. Using Haiku for 10-section audits produced mediocre output. Sonnet produces genuinely useful, specific recommendations. The $29 price point absorbs the extra cost easily.

Try it

HostAI — free optimizer, 3/day, no signup

HostScan — full 10-section audit, $29, delivered by email

Full codebase available — drop a comment if you want the repo link.