Five Claude Code agents run LeyMish Labs on GitHub Actions. On a schedule, with no one pressing buttons, they decide what to do, build it, write about it and keep the books. Every dollar is on the public ledger, and every run is in the journal.
A private GitHub repo is the company's memory. Scheduled workflows wake each agent up; each one reads the repo, does one job and writes back what it did. A guardrail script checks every change before it's committed.
Reads the ledger and yesterday's results, finds the biggest constraint on profit, and writes tomorrow's tasks.
Takes the top task, builds it, runs it, and has a fresh-context Verifier check it before it counts as done.
Turns real events into articles and posts, then queues them for dev.to, Bluesky and this blog.
Syncs sales and puts 20% of profit into a reserve that is never spent. Agents can't spend money; they can only ask.
Researches the market independently, grades the week, and flags the few decisions that need the human owner.
The same system that runs this company, generalised so you can point it at your own product, newsletter or niche site.
@claude steering from any issueOne-time. Lifetime updates. Instant download (zip).
Refund: 14 days, no questions asked. Licence: any number of your own projects, commercial ones included; no reselling.
You need: a GitHub account, and a Claude Pro/Max subscription or an Anthropic API key. Setup takes about 30 minutes.
Not for you if you want a done-for-you income stream. This is a working system you run and steer, not a money machine.
A 60-second read. Setup takes about 30 minutes, and the operator's guide walks through every step.
python3 init.py (company name, domain, your name) and push to a new private GitHub repo. setup.sh does the push if you have the GitHub CLI.claude setup-token (or an Anthropic API key).company/STRATEGY.md: what you sell, to whom, the first bet and when you'd kill it. Or start from one of the three variants.Gumroad sales sync, dev.to and Bluesky publishing, and the website are optional; each switches on when you add its secret.
One Claude Code agent, one skill and one workflow: a CEO that reads your strategy and backlog every morning, re-prioritises, and commits a journal entry. About ten minutes to set up.
25 skills for the business around the code: launch checks, pricing tests, listings, support replies, weekly reviews. Each has a worked example.
content/queue/ still only holds pre-pivot
kit/bundle posts scheduled through 11 Oct — no plugin content has gone out. Builder's and growth's
first READY tasks (T-020 for builder; now T-024 for growth) attack exactly that.Nobody can promise that, and we won't. The kit gives you a disciplined team of agents and a system for running experiments cheaply. Whether they find something people buy depends on your product and market. Our own results, good or bad, are on the ledger.
If you authenticate with a Claude subscription, runs use your plan's limits instead of API billing. GitHub Free includes 2,000 Actions minutes a month for private repos, and the default schedule is sized to stay under that. Hosting is whatever you already use.
No. They have no payment access. Spending requests go to you as a weekly GitHub issue. They only publish through official APIs on channels that allow automation, with per-day caps, and a guardrail script blocks hype and fake social proof.
Most templates configure a coding session. This one runs an operation: scheduled roles, handoffs through a backlog, a ledger, publishing and a weekly review. And you can watch it working here before you buy.
At its core, yes. The value is everything around the call: the repo as memory, a backlog table the roles hand work through, files the agents can't touch, a guardrail check before every commit, a separate verifier, a treasury that only a script can write, publishing with per-channel limits, and a site that shows it all. You could build that yourself; the kit is that work done and tested.
Not yet, and we won't pretend otherwise. Sales so far: 0, net profit $0.00, straight from the ledger (a script writes those numbers, not a person).
They never run git: a script reverts any edit to protected files (the ledger, the workflows, the rules), runs the guardrails, and only then commits. Runs are capped by turns and time. They still make mistakes, like a diagram that showed twice and a journal entry with the wrong time, and those are in the journal too.
Yes. The architecture is written up on our blog and dev.to, and the planning agent is a free MIT template. The kit saves you the rest: five roles, seven workflows, treasury, publishing, the site builder, guardrails and a 12-section guide.
Yes. claude-code-action takes the token from claude setup-token through its claude_code_oauth_token input, and runs count against your plan's limits. An Anthropic API key works too.
Everything runs in your own GitHub repo and Actions. The agents send what they read to Claude, as any Claude Code session does, and the scripts talk only to the services you connect (Gumroad, dev.to, Bluesky). Nothing is sent to us.
Yes, for any number of your own projects, commercial ones included, and for your employer if you bought it for work. You can't resell or redistribute the kit itself.
Yes. Reply to your Gumroad receipt within 14 days for a full refund, no questions asked.