The problem with local AI agents

OpenClaw is an open-source personal AI agent. It connects to your tools — Slack, Notion, Gmail, AWS — and runs tasks on a schedule or on demand.

Running it on your laptop has two problems. First, your laptop sleeps. The agent dies mid-task and has no idea what happened when you come back.

Second, and worse: the agent runs with full access to your machine. SSH keys, browser cookies, local code. If it gets compromised, everything is exposed.

The $5 fix

A dedicated Lightsail instance. The micro_2_0 bundle: 1GB RAM, 1 vCPU, 40GB SSD, 1TB transfer. $5/month flat. No data transfer surprises, no NAT gateway fees. First 90 days free.

Zero open ports with SSM

Lightsail can't attach IAM instance profiles like EC2. Instead, we use SSM hybrid activation. The instance self-registers with Systems Manager on first boot. From that point, you connect via Session Manager. Port 22 stays closed. No SSH keys to manage.

Someone scans your IP? They find nothing.

GitHub as the time machine

The ~/.openclaw directory becomes a Git repo. Every config change, every new skill, every memory update — version controlled. Break something? git checkout to when it worked.

A deploy key scoped to a single repository handles auth. Least privilege, applied to an SSH key.

The inception part

OpenClaw runs a cheap model for daily tasks. But when you want to write new skills, you use Claude Code on your local machine. It connects via SSM and deploys skills directly. A more capable AI improving a less capable one.

What it costs

$5/month for Lightsail. GitHub, SSM, IAM — all free. The variable is LLM usage.

Start with GPT-5 nano or Gemini 2.5 Flash Lite. Do not use a frontier model for an always-on agent. OpenRouter's Auto mode once picked Opus with extended thinking and burned hundreds of dollars in minutes. Set a billing limit before you connect anything.

🚀 Read the Full Tutorial →