An always-on home for your agents, not your laptop.

Step 1: order, create an unprivileged agent user

Pick Ubuntu 24.04 at checkout; the VPS is ready 30 to 60 seconds after payment. The agent never runs as root: give it its own user without sudo.

ssh root@YOUR_IP
apt update && apt upgrade -y
adduser --disabled-password agent
mkdir -p /home/agent/.ssh && cp ~/.ssh/authorized_keys /home/agent/.ssh/ && chown -R agent:agent /home/agent/.ssh
ufw allow 22/tcp && ufw allow 80/tcp && ufw allow 443/tcp && ufw enable

Step 2: keys in an env file, not in the shell history

Every provider key goes in one file only the agent user can read. Shell exports end up in .bash_history; a file does not.

su - agent
install -m 600 /dev/null ~/.agent.env
cat >> ~/.agent.env <<'ENV'
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
ENV

Step 3: a coding agent that survives a dropped SSH session

tmux keeps the session alive when your connection drops. Detach with Ctrl-b d and the agent keeps working.

apt install -y tmux        # as root, once
tmux new -s code
set -a; . ~/.agent.env; set +a
cd ~/project && claude    # or any other coding agent CLI
# later, from a new SSH login:
tmux attach -t code

Step 4: a systemd unit for the gateway, MCP server or LangGraph app

Anything that must survive a reboot gets a unit. This one runs a Python app as the agent user and restarts on failure.

# /etc/systemd/system/agent-app.service
[Unit]
Description=Agent app
After=network-online.target

[Service]
User=agent
WorkingDirectory=/home/agent/app
EnvironmentFile=/home/agent/.agent.env
ExecStart=/home/agent/app/.venv/bin/python main.py
Restart=on-failure
RestartSec=5

[Install]
WantedBy=multi-user.target

systemctl daemon-reload
systemctl enable --now agent-app
journalctl -u agent-app -f

Step 5: a Docker sandbox for code the agent writes

Code generated by a model runs in a throwaway container with no network, a memory cap and a CPU cap. Install Docker from the official apt repository first.

docker run --rm --network none --memory 1g --cpus 1 \
  --read-only --tmpfs /tmp -v /home/agent/work:/work:rw -w /work \
  python:3.12-slim python task.py

Step 6: HTTPS in front of the MCP server and webhooks

Caddy handles the certificate. Basic auth is the minimum; a bearer token check in the app itself is better.

# /etc/caddy/Caddyfile
mcp.example.com {
    basic_auth {
        agent JDJhJDE0JC4uLg   # output of: caddy hash-password
    }
    reverse_proxy 127.0.0.1:8080
}

Step 7: Ollama, if you want a small local model

curl -fsSL https://ollama.com/install.sh | sh
ollama pull YOUR_MODEL
free -h    # the model must fit in this RAM