Solo dev builds an AI control plane that lets Claude run servers safely
A developer running four personal servers solo got tired of silent backup failures, config drift, and containers dying at 3am, so they built a self-hosted to fix it. Adding a new server takes one line: running a bootstrap script with a single-use token joins it to a WireGuard mesh network and installs the agent automatically, with no SSH keys to manage. Claude, connected via MCP (a protocol that lets an AI model call external tools), can operate deploys, migrations, restarts, DNS, and backups across the whole fleet, but destructive actions never run directly.
Every action first goes through a dry run (a preview showing what would happen without actually doing it), then policy checks, then gets written to an , and comes with a rollback point. Every file the AI writes gets automatically scanned and broken into reusable patterns, language-agnostic and with no predefined rules, so the pattern library grows on its own as it sees new code. Later, when building something similar, proven patterns from past work get surfaced and reused instead of rebuilt from scratch, meaning the system gets better at working with your codebase the longer it runs.
An unexpected side benefit: real token savings, since the AI stops burning context re-reading the codebase and re-deriving the same solutions every session.
Key points
- Solo-run homelab with 4 nodes; new node enrollment is one command (auto-joins WireGuard mesh, no SSH key management)
- Claude connects via MCP to handle deploys, migrations, restarts, DNS, and backups across the fleet
- Destructive actions are blocked; every action goes through dry-run preview, policy gates, an , and a rollback point
- AI-written files are automatically scanned into a self-growing, language-agnostic library of reusable patterns
- Reusing proven patterns cuts repeated context reads, producing real token savings over time