checkup-engine

Open-source AI-readiness checkup for small-business websites.

Early stage (v0.1). The rubric and collectors are usable today; the hosted Demo is a thin Cloudflare Worker plus a small Python service on our gateway VM. Expect rough edges.

What it does

1. Fetches a public website (home + a few key pages), respecting robots.txt.
2. Scores it against a public, reproducible 5-dimension rubric.
3. Optionally asks an LLM to write up to 3 recommendations, each citing an evidence URL from the pages it actually read. A QA step drops unsupported claims.
4. Renders a bilingual HTML report.

SSRF defence is first-class: only http/https on ports 80/443, no credentials, DNS resolution must be fully public, and the TCP connection is pinned to the vetted IP (blocks DNS rebinding). See checkup_engine/ssrf.py.

Quick start (CLI)

python3 -m checkup_engine https://example.com --lang en --offline --out report.html
# With a LiteLLM / OpenAI-compatible gateway:
#   export GATEWAY_CHECKUP_KEY=...
#   export GATEWAY_CHECKUP_BASE_URL=https://llm.agentbay.asia/v1
#   export GATEWAY_CHECKUP_MODEL=deepseek-v4
python3 -m checkup_engine https://example.com --lang zh --out report.html --json report.json

Python 3.10+, stdlib only — nothing to pip install to run the engine or the VM service.

Live Demo status (early stage)

Piece	URL / note
Engine health	https://llm.agentbay.asia/checkup/health
Worker API (preview)	https://talentos-checkup-demo.jimmywangjimmy.workers.dev
Site Demo pages	/demo/ and /en/demo/ on www.agentbay.asia — not published to gh-pages until Reed confirms
Turnstile	Preview still uses Cloudflare test keys (always pass). Real site+secret keys: wait for /workspace/secrets/turnstile-checkup.env from Relay, then wire Site Key into the Astro demo pages and Secret into the Worker.

Honest limits: site-wide 12 runs/day, 3 per IP/day, gateway budget $1/day (resets 08:00 Beijing). Expect rough edges; scores are reproducible, recommendations are estimates.

Architecture (Demo)

Browser (/demo/)  →  Cloudflare Worker (Turnstile + KV rate limit + proxy)
                  →  Azure VM service (localhost:8790, behind Caddy /checkup/*)
                  →  LiteLLM gateway (deepseek-v4)

• Worker stays on the Cloudflare free plan (CPU ≪ 10 ms, ≤ a few subrequests).
• Site-wide Demo cap: 12 runs/day; per-IP: 3/day. Budget resets 08:00 Beijing time.
• Visitor-facing errors are bilingual and never include raw gateway text.

VM install steps: service/DEPLOY.md.

Monitoring daily report

python3 scripts/daily_sample.py --site https://www.agentbay.asia/
# writes samples/YYYY-MM-DD/{snapshot.json,daily.md,daily.html}

Suggested cron (Beijing time, off-peak for the model):

# 07:30 Asia/Shanghai — tell Reed; do not create the routine yourself
30 7 * * * cd /opt/checkup-engine && set -a && . /etc/checkup-engine/env && set +a && python3 scripts/daily_sample.py

Tests

python3 -m unittest discover -s tests -v
node --test worker/test/*.test.mjs

License

MIT © 2026 TalentOS

---

Maintained by TalentOS · 需要定制 Agent / 技能开发？找我们定制 / Work with us · talentos@agentbay.asia