MX3 OS
A private operating system for a one-person studio, with autonomous Claude agents under human review.
- Role
- Designer and sole engineer
- When
- 2026
- Status
- Ran in production from May to September 2026; shut down to cut hosting costs
The problem
Running several products alone means the same chores every day: triaging issues, shipping small fixes, watching deploys and costs, answering email. MX3 OS put all of it in one place and handed the repetitive parts to agents.
What I built
Autonomous agents with code-enforced permissions
Scheduled Claude Code agents worked against project repos on isolated Fly.io machines, filed suggested issues for a human to approve, and opened pull requests only after approval.
Review queue for risky actions
Any action that spends money, sends email or writes code went to a review queue first. Invoice creation through Stripe and QuickBooks was verified end to end behind that gate.
Operator console
Clients, projects, deploy ingestion from GitHub Actions, logs, status pages, time and invoices, an encrypted secrets vault, and cloud and AI cost tracking in one Angular app.
Email in, agent out
Inbound email through Amazon SES could summon an agent with the message as context, and replies went back out as branded HTML.
Outcome
- Three agents shipped daily pull requests against real project repos while it ran.
- Full stack ran on AWS (ECS Fargate, CloudFront, ElastiCache) and Fly.io, provisioned with Terraform and deployed through GitHub Actions.