Opensource hosting platform built for AI agents to operate
retained text
Disclosure: this is my project. It's free and open source, and I'm not selling anything. Project: https://agentic.hosting Code: https://github.com/dennisonbertram/agentic-hosting Quick context if you missed it: on Oct 21 Railway disclosed that someone may have accessed customers' environment variables, some of them decrypted, and paused deploys for about a day. Their post says it was part of a series of attacks on platform companies. I was a customer and the cleanup was painful The annoying part was the timing: you were told to rotate, but rotated keys couldn't go live until deploys came back. I spent dinner in the living room rotating. A couple of posts from that night that sum it up: “we can invalidate and rotate but can't deploy the new key values. this breaks all services with no remedy or timeframe." @[jimmy0x52](https://x.com/jimmy0x52/status/2113105079960027366) "openrouter has revoked my keys, but i can't rotate and redeploy my environment cause builds are paused??!!" @[roguesherlock](https://x.com/roguesherlock/status/2113093275834102105) Not here to bash them. The thing that stuck with me is that a lot of us keep our secrets in the same few places, so one break-in hits everyone. Most of you already know self-hosting works. The usual pushback I hear is that it needs someone to babysit the box. That's what I've been testing: can an AI agent do the babysitting? **What I built** agentic-hosting, a deploy server meant to be driven by an agent over an HTTP API instead of a UI. \- Single Go binary, SQLite for state. I run it on a Hetzner CX33 \- Give it a git URL, it builds with Nixpacks and returns a URL \- Postgres and Redis per app \- Containers run under gVisor, one Docker network per tenant \- Env vars encrypted at rest (AES-256-GCM), master key stays on the host \- \`ah backup\` writes WAL-consistent, gzipped snapshots of the state DB (keeps the last 10) \- Self-healer checks every 30s and restarts unhealthy containers **Honest tradeoffs** \- Your agent can build out redundancy and backups for you, like a second box to fail over to and scheduled volume backups. \- \`ah backup\` covers the control-plane DB only. Postgres/Redis volumes are not included yet, so you need your own volume backups. \- Backups only work with the same master key. Lose the key and the encrypted env vars are gone. \- Firewall rules are only checked when \`ah\` starts. If you reload ufw/fail2ban, restart \`ah\`. \- No polished UI. If you want a dashboard, Coolify or Dokploy are better fits right now. \- Single node. No HA. \- Self-hosting isn't automatically safer. An unpatched VPS is worse than a good platform. The point is that your agent keeps it patched and watched, and you're not one of thousands of tenants behind the same front door. Quickstart (about 5 minutes) is at the top of the README: https://github.com/dennisonbertram/agentic-hosting **What I'd like feedback on** 1. If you let an agent touch your servers, where do you draw the line? Read-only, or can it run \`apt upgrade\`? 2. Env-var encryption at rest vs. just full-disk encryption: worth it? 3. Anything in the quickstart that breaks on your setup.
document fetched 2026-10-24 10:32 · canonical https://reddit.com/r/selfhosted/comments/1xddsys/opensource_hosting_platform_built_for_ai_agents