A 150-line Python script plus 5 markdown templates that lets a Claude Code agent run continuously for 36-plus hours, outputting code commits or article drafts every hour -- for 20 USD/month (Pro subscription). No API billing. No VSCode open. I built it in 14 days. Now I teach you in 30 minutes.
Root cause: Claude Code is an "IDE assistant" architecture -- not an "agent runtime." Close VSCode and it stops. Its CronCreate tasks die with the window. Memory evaporates when the session ends.
Most people hit one of these walls:
The fix: Layer an orchestration system outside Claude Code. This is what Anthropic will not give you -- and it is the difference between "I tried agents" and "agents shipped my product."
All verified on GitHub @jiejuefuyou. Privacy Manifest passed. Zero analytics SDKs. Zero network calls from apps.
All prices in CNY. Pay via Alipay / Wise / PayPal. No contract. Monthly retainer available after onboarding.
Email me with your setup (Windows/macOS) plus use case. I send you a time or WeChat handle.
Screen-share walkthrough. Calibrate coordinates, set up Task Scheduler/launchd, run your first test fire. You will see it produce output within the session.
Your agent runs overnight. You see hourly commits in your GitHub. I review loop.log and send tweaks if needed.
Your agent handles backlog autonomously. You review and approve decisions. Time-to-market collapses. Pro clients get weekly log reviews for 4 weeks.
Starter gets you set up even without coding experience -- as long as you can install Python venv and edit a markdown file. If you cannot articulate "what I want the agent to do" in writing, go Professional.
API plus workflow platforms are for teams (1000-plus USD/month). This is for solo indies (20 USD/month Pro subscription). Run both in parallel: API for high-priority tasks, desktop loop for low-priority overnight work.
Desktop-loop simulates your own keyboard and mouse into your own IDE -- not a violation of Anthropic TOS. I have run 36-plus hours with zero warnings. Rules: do not run 24/7 at full speed, do not share the account, do not market it as "unlimited automation."
Both work. Windows Task Scheduler has friendlier GUI; macOS launchd is cleaner but requires a one-time Accessibility permission. Pick whichever machine can stay on 24/7.
Starter is non-refundable (99 CNY = 30 min of my time). Professional and Enterprise: if after 2 weeks it is genuinely not working, partial refund considered case-by-case.
Starter/Professional: no. Enterprise includes a commercial license for internal deployment, but you cannot offer "Claude Code 24/7 consulting" as a service to others.
6 questions to answer before the call -- tells you if $99 Starter, $499 Professional, or DIY is right for your setup. Sent within 24 hrs, no spam.
The full toolkit is MIT open source at github.com/jiejuefuyou/autoapp-toolkit -- read it, fork it, build it yourself.
What you are paying for is the 14 days of debugging I already did so you do not have to.