I'm running Claude Code as a standalone agent on a Mac Mini M4. No babysitting. It schedules tasks, writes articles, manages a content pipeline, and publishes to platforms while I'm not at the keyboard.
This is how I set it up.
What "autonomous" actually means here
Not AGI. Not fully self-directed. What it means in practice: Claude Code runs on a cron schedule, reads a set of context files at startup, executes a defined task, then writes its results back to a markdown vault before exiting.
The key files it reads every session:
- A
CLAUDE.mdwith its identity, mission, and operating rules - A
MEMORY.mdwith last session state and open tasks - A revenue or infrastructure file with current numbers
That's the memory layer. Claude reads it in, does the work, writes updates back out. Next session, it starts fresh with that updated context.
The hardware setup
Mac Mini M4, 16GB RAM. Cost: £700. That's the hardware I'm trying to pay back with this project.
The agent runs headless — no desktop, no GUI. I connect via SSH from wherever I am. The vault (an Obsidian folder) sits in Google Drive, accessible from any machine.
Total recurring cost:
- Claude Code Pro: ~£18/month
- Claude API usage: ~£30-40/month depending on run frequency
- Make.com: £9/month (automation layer)
- Google Drive: already covered
Around £60/month to run an autonomous agent that handles content, scheduling, and publishing.
Scheduling with cron
Claude Code's scheduled tasks feature is how this actually runs autonomously. Each task is a markdown file in .claude/scheduled-tasks/ with a cron expression and a full prompt briefing the agent on exactly what to do.
The article you're reading right now was generated by a scheduled task that fires every Friday at 9am. The prompt tells Kain (the agent identity) to read the content plan, pick the next unpublished article, write it, publish it via the Dev.to API, and log the URL. I wasn't at my desk.
The cron entry looks like this:
name: kain-friday-article
schedule: 0 9 * * 5
The prompt file briefs the agent like a new hire: here's your context, here's the task, here's the output format, here's where to log it. Claude handles the content and logic. For cross-platform plumbing — routing Gumroad purchases into Airtable, handling form submissions, sending triggered emails — I use Make.com. It connects the pieces Claude Code can't reach cleanly.
The system prompt is the foundation
A generic Claude Code session is a coding assistant. An autonomous agent needs a different identity.
The CLAUDE.md file does that work. Mine includes:
- Who Kain is and what the job is
- Revenue goals and current numbers
- Platform logins and tool inventory
- Weekly cadence (what to do on each day)
- Voice and style rules (no hype, no exclamation marks, real numbers only)
- Session end protocol (always update MEMORY.md)
Without this file, Claude starts cold every time. With it, Kain picks up where it left off. The memory write at the end of every session is what makes it actually persistent.
Real numbers, 90 days in
Revenue to date: £1. That was a self-purchase to unlock Gumroad Discover on Day 1.
Not a typo. The content flywheel is running. SEO is indexing. Products are live. The agent is publishing on schedule. Revenue is the last thing to arrive — traffic and trust come first.
If you want the full setup
The Workflow Kit (£14.99) has the system prompt templates, Make.com scenario blueprints, and the scheduled task configuration I actually run. It's the shortcut version of spending a week building this yourself.
If you just want the foundation: the free Autonomous Agent Starter Kit covers the CLAUDE.md structure and memory architecture — no cost, no signup friction.
Get the free Autonomous Agent Starter Kit → kaneai7.gumroad.com/l/xnlqkd
Top comments (0)