vps-forge
llms.txt40 starsChanged 3 days agoMITView on GitHub
An open Agent Skill (SKILL.md) that lets AI coding agents set up, harden and audit Debian 12/13 and Ubuntu 22.04/24.04/26.04 servers without locking the user out. On RHEL-family servers, Kubernetes nodes and hypervisor hosts it stops and explains why instead. It works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Antigravity, OpenCode and other agents that load skills. Install it with npx skills add mountain-beaver/vps-forge -g. MIT-licensed, no telemetry. Every SSH, firewall, sudo or PAM change runs inside a…
In this repository3 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .claude/CLAUDE.mdCLAUDE.md
- vps-forgeSkill
The file
Files
# vps-forge > An open Agent Skill (SKILL.md) that lets AI coding agents set up, harden and audit Debian 12/13 and Ubuntu 22.04/24.04/26.04 servers without locking the user out. On RHEL-family servers, Kubernetes nodes and hypervisor hosts it stops and explains why instead. It works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Antigravity, OpenCode and other agents that load skills. Install it with `npx skills add mountain-beaver/vps-forge -g`. MIT-licensed, no telemetry. Every SSH, firewall, sudo or PAM change runs inside a dead-man switch: snapshot, change, test from a new connection, and roll back automatically after 5 minutes unless the test login worked. It can't roll back a firewall in the provider's panel, so it changes that additively: new rule first, old rule last. The finished server has key-only SSH with no ssh-audit failures, a default-deny firewall, fail2ban that ignores the admin's IP, unattended security updates, restic backups with a tested restore, phone alerts (ntfy, Telegram or a webhook) and a SERVER.md. On a host run by Coolify, Dokploy, CapRover, Kamal or Dokku it keeps what the panel needs: root key login, the panel's proxy and its own SSH. In 120 benchmark runs on real VMs, Claude Opus 5.5 passed 79–92% of an outside-in checklist without the skill and 99–100% with it (about 70% of that gap comes from three checks that match the skill's own defaults), with no lockouts in either arm; a run with the skill cost about 3× as much (median $1.87 vs $0.62 at API list prices). Only that model has the full benchmark; a four-run spot check of GLM-5.3 showed the same pattern (67–79 without the skill, 100 with it). The scenarios were known while the skill was revised. ## Docs - [README](https://github.com/mountain-beaver/vps-forge/blob/main/README.md): what it does, install for each agent, examples, safety model, FAQ - [Quick start](https://github.com/mountain-beaver/vps-forge#quick-start): install commands for every supported agent - [SKILL.md](https://github.com/mountain-beaver/vps-forge/blob/main/skills/vps-forge/SKILL.md): the skill itself, the workflow and safety rules the agent follows - [References](https://github.com/mountain-beaver/vps-forge/tree/main/skills/vps-forge/references): one file per phase (SSH, firewall, Docker, web, VPN, backups, alerts, audit mode) - [Scripts](https://github.com/mountain-beaver/vps-forge/tree/main/skills/vps-forge/scripts): audit.sh (read-only audit), deadman.sh (automatic rollback), external-check.sh, install-alerts.sh ## Benchmark - [Results and caveats](https://github.com/mountain-beaver/vps-forge#benchmark): scores per scenario, per check and per distro - [Methodology and harness](https://github.com/mountain-beaver/vps-forge/blob/main/bench/README.md): how the VMs, the agent sandbox and the outside-in judge work - [Per-run results](https://github.com/mountain-beaver/vps-forge/blob/main/docs/bench/results.md) ## Optional - [CHANGELOG](https://github.com/mountain-beaver/vps-forge/blob/main/CHANGELOG.md) - [SECURITY](https://github.com/mountain-beaver/vps-forge/blob/main/SECURITY.md): how to report a flaw privately
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

