Between setup sessions, DMs, Reddit threads, and Discord I've now looked at over 50 different OpenClaw setups. Broken ones. Working ones. "It works but it costs $200/month" ones.

The problems are almost never unique. It's the same 5 things, every time. Here's the list — check your own setup against it.

⚠️ Quick check: If you're spending more than $30/week on API costs or your agent feels slow and forgetful, you probably have at least 2 of these.

1 Opus as Your Default Model

This is the single most expensive mistake in the entire OpenClaw ecosystem — and it's the default for many people coming in from tutorials.

Opus is incredible. It's also 10–15x more expensive than Sonnet for tasks where you won't notice the difference.

Your agent checking your calendar? Sonnet. Summarizing an article? Sonnet. Setting a reminder? Sonnet. Writing a quick email draft? Sonnet.

Opus makes sense for deep research, long multi-step reasoning, or nuanced writing where quality genuinely matters. That's maybe 5–10% of what most people use their agent for.

One person I helped was spending $47/week. We changed the default model to Sonnet and added one line to their SOUL.md. Their next week cost $6.

✓ The Fix

In your openclaw.json:

{
  "ai": {
    "model": "anthropic/claude-sonnet-4-6"
  }
}

Then in your SOUL.md, add: "Use Opus only when I explicitly ask for deep analysis or complex reasoning. Default to Sonnet for everything else."

Expected savings: 60–85% on API costs. High impact

2 Never Starting a Fresh Session

This is the silent budget killer nobody talks about.

Every message in your current session gets sent with every new API call. That means if you've been chatting with your agent for 3 weeks in the same session, your "what's the weather?" question is carrying thousands of tokens of old conversation with it. You're paying for all of that. Every single time.

Three people I helped cut their monthly costs by 40–60% by doing one thing: typing /new before heavy tasks.

Your agent doesn't lose its memory when you start a new session. It still has SOUL.md, USER.md, MEMORY.md — all its files. You're just clearing the conversation buffer. Think of it like closing and reopening a chat window. The person on the other end still knows who you are.

✓ The Fix

Type /new at the start of each new topic or work block. Do it before any big task. Make it a habit. Your bill will thank you.

Expected savings: 30–50% on ongoing costs. High impact

3 Installing Skills Without Reading the Source

ClaWHub has 13,000+ skills. That sounds great. It's also a problem.

Security tools have flagged hundreds of skills as actively malicious — infostealers, backdoors, remote access tools disguised as automation. And that's just the ones that got caught.

But even the non-malicious skills can wreck your setup. I've seen skills that:

✓ The Fix

Rule: if you can't read and understand the skill's source code in 5 minutes, don't install it. If it needs shell access or network access, understand exactly why before it touches your setup. When in doubt, skip it.

Risk reduction: prevents silent cost leaks + protects your data. Security

4 Gateway Exposed to the Network

If you installed OpenClaw on a VPS and your gateway config says "host": "0.0.0.0" — or you didn't set it at all — your agent might be accessible to anyone who knows your IP.

That means a stranger could message your agent. Your agent that has access to your email, your calendar, your files, and possibly your shell.

Check right now:

openclaw config get | grep host
✓ The Fix

In your openclaw.json:

{
  "gateway": {
    "host": "127.0.0.1"
  }
}

Then access via SSH tunnel or a reverse proxy with auth (Nginx + basic auth, or Tailscale). Never expose your gateway directly to 0.0.0.0 on a public server.

Risk: critical. Fix takes 2 minutes. Security

5 No Structured Memory Architecture

Most OpenClaw setups use memory as one big file — or worse, rely entirely on conversation history. Both approaches break down fast.

The result: an agent that forgets important context, repeats itself, gives inconsistent answers, and costs more because it's re-learning things every session.

Structured memory means separating your agent's knowledge into specific files: one for your identity and mission, one for active projects, one for people and contacts, one for decisions made. The agent reads what it needs, when it needs it — and remembers correctly across sessions.

✓ The Fix

At minimum, create these files in your workspace:

SOUL.md       — who you are, what the agent should do
USER.md       — your name, preferences, timezone
memory.md     — key facts, projects, decisions
HEARTBEAT.md  — what the agent does when idle

Reference them explicitly in your SOUL.md so your agent knows to read them. This alone transforms an inconsistent chatbot into a reliable assistant.

Impact: transforms agent quality and consistency. Quality

The Audit Version

If you want to check your own setup quickly, here's the 5-point checklist:

  1. Model: Is your default model Sonnet (or local), not Opus?
  2. Sessions: Are you using /new regularly, or running one long session indefinitely?
  3. Skills: Have you read the source of every skill you've installed?
  4. Gateway: Does openclaw config get | grep host show 127.0.0.1, not 0.0.0.0?
  5. Memory: Do you have separate, structured memory files your agent reads on every turn?

If you answered "no" to 2 or more of these, your setup has room to improve — and the improvements are fast.

✅ Most of these fixes take under 10 minutes. The biggest one (model routing) takes 30 seconds and can save you hundreds of dollars a month.

Want someone to check your setup?

Book a $49 audit. We'll go through all 5 of these — plus anything else we find — and give you a prioritized fix list. Most audits find at least 2 issues; many find all 5.

Book a $49 Audit →