openclaw-starter

Fork this. Run your own autonomous AI agent in 30 minutes.

A minimal, opinionated starter kit for running a 24/7 autonomous AI agent on a VPS using OpenClaw.

Based on the actual setup behind @ghostferal — an AI agent that creates YouTube content, manages social media, and operates independently for months at a time.

What's Included

Quick Start

1. Get a VPS

Hetzner CX22 ($5/mo) or CX32 ($15/mo for 8 CPU). Ubuntu 24.04 LTS.

2. Install OpenClaw

curl -fsSL https://openclaw.ai/install.sh | bash

3. Configure Your Model

openclaw config set agents.defaults.model anthropic/claude-sonnet-4-6
echo "ANTHROPIC_API_KEY=sk-ant-..." >> ~/.openclaw/.env

4. Fork + Clone This Repo

git clone https://github.com/YOUR_USERNAME/openclaw-starter ~/clawd
cd ~/clawd

5. Fill In Your Info

Edit these files:

• USER.md — who you are, timezone, contact, preferences
• MEMORY.md — initial context about you and your projects
• HEARTBEAT.md — what you want the agent to do every hour

6. Set Up a Channel

# Telegram (recommended)
openclaw channels login --channel telegram

# Or WhatsApp
openclaw channels login --channel whatsapp

7. Start

openclaw gateway start

Say hello to your agent. It will read SOUL.md, AGENTS.md, and USER.md automatically.

The Memory System

The agent wakes up fresh every session. These files are its memory:

MEMORY.md          ← curated long-term facts (read every main session)
memory/2026-03-19.md  ← raw daily log (read today + yesterday)
SOUL.md            ← personality (never changes much)
HEARTBEAT.md       ← hourly checklist (you edit this)

No vector database. No embeddings. Just markdown. It works because:

1. Models have 200k context windows
2. MEMORY.md stays under 5k tokens if you curate it
3. The agent updates its own memory as it learns

More: How to Give an AI Agent Long-Term Memory

Heartbeat System

OpenClaw pings your agent every hour. The agent reads HEARTBEAT.md and acts on it. Example:

# HEARTBEAT.md

## Every Heartbeat
1. Check email for anything urgent
2. Work on top KANBAN task
3. If nothing: research trending topics, write a tweet draft

## Rules  
- Night (23:00-08:00): no messages, just background work
- Never send more than 2 messages to human per day

What This Enables

• Content creation: The agent produces YouTube videos, tweets, blog posts autonomously
• Research: Monitor news, trends, GitHub, Hacker News — summarize what matters
• Coding: Spawn Claude Code sub-agents for development work
• Operations: Monitor servers, rotate logs, run backups, alert on issues

Customizing

The key files to edit:

SOUL.md — personality and rules. Be specific. "Never use emdashes" works. "Be helpful" doesn't.

HEARTBEAT.md — your agent's recurring job list. Keep it short (under 10 items) or it ignores everything.

KANBAN.md — what it's working on. Max 3 DOING items. Forces prioritization.

Cost

Minimum viable setup: ~$25/mo. This one runs on $115/mo (Hetzner CX32 + Claude Max).

Resources

• OpenClaw docs
• How I Run an AI Agent 24/7
• AI Agent Memory Without a Vector Database
• model-watchdog — auto-rollback for config changes

Built by feralghost. The agent wrote parts of this README.