OpenClaw Security Audit & Multi-Agent Setup Guide

"Seven days ago you were wondering what an AI assistant was. Now you have one that's online 24/7, knows you, can act, and works proactively. This is just the beginning."

Today's Goal

What You've Built This Week

What you have now is a digital presence that runs for you around the clock. Day 7 is about making it stronger, safer, and understanding where it goes from here.

Must Complete: Security Review

Your assistant touches email, calendar, files, and possibly your browser. Review this now, then revisit it monthly:

• SSH uses key auth, and password login is disabled.
• Firewall is enabled, with only the ports you actually need exposed.
• OpenClaw runs as a non-root user.
• All API keys live in environment variables or .env, not in committed code.
• .env is ignored by git.
• OAuth token files are locked down, for example chmod 600 token.json.
• Keys and tokens have a rotation reminder, ideally every ~3 months.
• SOUL.md explicitly says external messages require confirmation unless you intentionally want automation.
• Destructive operations such as deleting files or changing configs require confirmation.
• Your model provider dashboard has a monthly budget or usage alert configured.
• Heartbeat runs no more often than you truly need.
• Unused Skills are disabled or removed.

Verification check: you should be able to point to the exact file, setting, or dashboard control that satisfies each checked item. Anything you cannot verify stays unchecked.

Optional: Advanced Patterns

You do not need these to finish the beginner guide. Use them once the core system already feels stable.

Nodes: Multi-Device Collaboration

OpenClaw can pair with your phone, tablet, or other devices — giving it eyes, ears, and location awareness.

What Nodes enable:

• Camera — capture photos or short video clips from your iOS or Android device
• Audio — voice input and playback
• Location — share your current location for context-aware tasks
• Voice wake — trigger the assistant hands-free

Pairing a device:

openclaw nodes pending                         # List pending device requests
openclaw nodes approve <requestId>             # Approve by request ID

Once paired, you can ask: "Take a photo of what's in front of me" or "Where am I right now?" and the assistant acts on real sensor data.

Nodes overview: camera, audio, location, voice →

Write Your Own Skill

Community Skills do not cover everything. A custom Skill can be as small as one Markdown file.

1. Create the folder: mkdir -p ~/.openclaw/skills/weather
2. Open ~/.openclaw/skills/weather/SKILL.md with any text editor — on Mac/Linux: nano ~/.openclaw/skills/weather/SKILL.md, on Windows: open it in Notepad.
3. Paste the minimum template:

# Weather Query

## Capability
Query current weather for any city using the wttr.in API.

## Usage
curl "wttr.in/{city}?format=3"

## Output
Report temperature and conditions in one concise sentence.

4. Restart the service with openclaw gateway restart, or ask your assistant to "refresh skills" in chat — it picks up new skill files on the next turn.
5. Test: ask What's the weather in Tokyo? and confirm your assistant uses the new Skill.

This is how many community Skills begin: one workflow written down clearly enough for the assistant to reuse.

Full guide: creating skills →

Verification check: if the Tokyo test does not trigger the Skill, list loaded skills, confirm the folder path is ~/.openclaw/skills/weather/SKILL.md, then ask your assistant to "refresh skills" or restart with openclaw gateway restart.

Multi-Agent Patterns

Your assistant can delegate tasks to specialized subagents and run them in parallel. This is useful only when the task is large enough to benefit from parallel work.

Example: you send one message: "Research the competitive landscape for AI coding assistants." Your main agent spawns three subagents simultaneously:

• One searches arXiv for recent papers
• One scans GitHub for trending repos
• One reads Hacker News threads on the topic

All three return results to the main agent, which synthesizes them. That is faster than doing the same work sequentially.

Subagents: spawn and delegate → ACP: agent-to-agent protocol → Multi-agent sandbox →

Verification check: describe one real task you would split into at least two subagents and what each subagent would own. If you cannot do that clearly yet, skip this pattern for now.

Real Usage Examples

See what people are actually building with fully set-up assistants:

openclaw.ai/showcase — real user automations →

Filter by "Developer," "Automation," or "Productivity" to find multi-agent builds and advanced patterns.

Where This Is Heading

Models keep improving. Every few months, Claude, GPT, and Gemini get substantially better. Your assistant — without any changes — automatically benefits. The infrastructure you're building now compounds with every model upgrade.

Costs keep dropping. Running a capable AI assistant today costs $10–30/month in API fees. The trajectory is clearly downward. At some point this becomes negligible.

Multimodal becomes standard. Assistants that see (camera feeds), hear (voice conversation), and speak (natural voice replies) are already in early access. The text-only interface is transitional.

Multi-agent networks. The future isn't one assistant — it's a team. One agent that manages email, one that handles code review, one that monitors analytics, coordinated by an orchestrator that reports to you.

The First-Mover Advantage

The most important thing to understand: the earlier you start, the bigger your advantage.

Every day your assistant runs, it accumulates memory about your habits, projects, terminology, and preferences. A 6-month-old assistant is categorically different from a fresh install — not because the model changed, but because the context has 6 months of learning baked in.

There's no shortcut for that accumulation. Only time.

You started this week. That's 6 months of compounding that begins today.

Common Issues

• Custom skill not loading: confirm the file is exactly ~/.openclaw/skills/weather/SKILL.md, then ask your assistant to "refresh skills" or restart with openclaw gateway restart.
• Cron syntax error: use a standard five-field cron expression and refer back to the cron reference before re-running the command.
• Need to see which skills are loaded: ask the assistant to list loaded skills, or use its reload/list-skills workflow before assuming the file was ignored.

Cron jobs: full syntax reference and examples →

Your Next Steps

This week, every day:

1. Talk to your assistant for 10 minutes. Let it get familiar with your patterns.
2. Update SOUL.md whenever something feels off. Souls are refined, not installed.
3. Try one new Skill. See what sticks.

In a month, your assistant will be in a different state than it is today — not because you made big changes, but because it learned.

That's the difference between AI assistants and every other tool you've used. It grows.

Verification check: by the end of this week, you should have one completed security review and one small advanced experiment logged, or a deliberate decision to stay on the core setup until it feels routine.