I built a panel that lets your AI agent manage your own VPS over MCP
I run a handful of VPSes across Hetzner and DigitalOcean, and I got tired of SSHing into each one just to check whether something was up. So I built ServerOS, a panel for servers you already own, and gave it an MCP server so Claude (or any MCP client) can look after them too.
How it works:
- You run one command on your server. It installs a small daemon (written in Rust, open source: github.com/serveroshq/daemon) that finds what's already running, like Docker containers, systemd services and databases, and leaves it all alone. Nothing gets migrated or reinstalled.
- Each workspace gets its own MCP endpoint. You connect it in Claude or your MCP client of choice, sign in with OAuth, and pick what the agent is allowed to do.
- The agent gets 109 tools: list machines and services, read and search logs, restart things, deploy from a git repo, roll back, set up uptime monitors, manage status pages, run backups, open firewall ports, install package updates and so on.
Things I can now just ask:
- "Why is the shop API slow today?" It pulls the logs and stats and tells me what it found.
- "Deploy the latest main of my repo to web-1 and tell me when it's healthy"
- "Set up a monitor for every site on db-main and put them on a status page"
- "Which of my servers haven't been backed up this week?"
The part I worried about most was letting an AI loose on real servers, so:
- Permissions are split into read, write and commands. You can give an agent read-only access and nothing else.
- Anything destructive, like rebooting, removing a service or restoring a backup, needs the machine or service name typed back as confirmation. The agent can't just fire it off.
- Every call is logged in the workspace activity as "done through MCP", against the person who connected it.
- Raw shell access is its own permission, off unless you turn it on.
It's a paid product (from £5/month for one machine), with a 14-day free trial of Pro. I'm one of the two founders, so I'd really like honest feedback, especially from anyone already letting agents touch their infra. What would you want the agent to be able to do, and what would you never let it do?