OpenClaw: Hardened Personal AI Setup

A step-by-step setup prompt you can paste into Claude Code (or any coding agent) to install, configure, and harden OpenClaw on a Mac.

OpenClaw turns Claude into a 24/7 personal AI assistant — persistent memory, tool access, and a direct line to your messaging apps (WhatsApp, Telegram, Slack, Discord, iMessage).

Prerequisites

• macOS (Apple Silicon recommended; Intel works too)
• Homebrew installed
• An Anthropic API key from console.anthropic.com
• A phone with WhatsApp (or a Telegram bot token, or Slack app credentials)

Usage

1. Open a fresh Claude Code terminal (or any coding agent with shell access)
2. Paste the contents of PROMPT.md
3. Follow along — it'll ask for your API key and show a QR code to scan
4. Send your first message: "Hey, let's get you set up. Read BOOTSTRAP.md"

That's it. The prompt handles installation, security hardening, watchdog setup, and first-run configuration.

What It Does

Phase 1: Install & Connect

• Installs OpenClaw via npm
• Runs the interactive onboarding wizard (API key, channel pairing, workspace creation)
• Installs the gateway as a launchd service (auto-starts on boot)

Phase 2: Harden & Verify

• Locks down file permissions (chmod 700 on config directory)
• Enforces loopback-only gateway binding
• Sets up token authentication
• Configures group chat safety (allowlist + require-mention)
• Installs a watchdog service for automatic crash recovery
• Runs a full security audit
• Verifies everything works end-to-end

What's Inside

├── README.md       ← You're here
├── PROMPT.md       ← The setup prompt (paste into Claude Code)
└── config/
    ├── ai.openclaw.gateway.plist    ← LaunchAgent template
    ├── ai.openclaw.watchdog.plist   ← Watchdog LaunchAgent
    └── watchdog.sh                  ← Health check script

Security Model

This setup is opinionated about security:

• Gateway binds to localhost only — not exposed to your network
• Token auth required — no unauthenticated access
• Group chats require @mention — bot won't speak unprompted in groups
• Config files are owner-only — chmod 700 on ~/.openclaw
• Watchdog monitors health — auto-restarts if gateway becomes unresponsive
• Prompt injection awareness — workspace files train the agent to reject embedded commands

If you use Tailscale, the gateway can be exposed to your tailnet (but never to the public internet via Funnel).

After Setup

Your agent wakes up fresh each session but persists through files:

File	What It Is
SOUL.md	Personality, values, boundaries
AGENTS.md	Operating manual (memory rules, security, workflow)
MEMORY.md	Long-term memory (curated by the agent)
memory/*.md	Daily notes
IDENTITY.md	Agent's name, vibe, emoji
USER.md	Your info (name, timezone, preferences)

These are created by OpenClaw's onboarding. The hardening prompt adds security rules and operational patterns on top.

Based On

A real setup running 24/7 on a headless Mac Mini with WhatsApp + iMessage, built over weeks of iteration. See the blog post for the full story.

Links

• OpenClaw • Docs • Discord • GitHub

---

Built by Aman Khan