5 Mistakes People Make When Self-Hosting an AI Assistant

Self-hosting an AI assistant is a great idea. But there are a handful of mistakes that almost everyone makes — and they're all avoidable if you know what to watch for.

We've seen these in support channels, GitHub issues, and community forums. Learn from other people's pain.

1. Running It on Their Personal Laptop

The Problem

"I'll just run it on my MacBook." Famous last words.

You set up OpenClaw, connect Telegram, have a great first conversation. Then you close your laptop to go to bed. Your AI assistant? Gone. Close the lid, it stops. Suspend, it stops. Restart, you have to manually start it again.

Even if you keep it running, your laptop is doing double duty: your work machine and your AI server. Context windows consume RAM. Background processes consume CPU. Your fans spin up during video calls because Claude is processing someone else's message in your group chat.

Why It Happens

It's the path of least resistance. You already have a computer, npm is already installed, why not just run it here? Setting up a remote server feels like an unnecessary extra step.

How to Avoid It

Use a dedicated server. A €4/month Hetzner VPS is always on, always connected, and doesn't care when you close your laptop lid. It's a dedicated machine with a dedicated job.

Your options:

• Self-host on Hetzner — €4-8/month, full control
• lobsterfarm — managed, no setup required
• Raspberry Pi — $50 one-time, runs on your desk

Any of these is better than your personal laptop. The AI assistant should be running somewhere you don't think about — silently, reliably, 24/7.

2. Not Securing the Gateway

The Problem

OpenClaw's gateway accepts connections from messaging channels and processes commands. If you expose it to the internet without authentication, anyone who finds your endpoint can interact with your AI — on your API key, with your money.

We've seen instances where people:

• Left the gateway port open with no auth token
• Used 0.0.0.0 binding without a firewall
• Shared their gateway URL publicly for "testing"

One user discovered their AI had been having conversations with strangers for three days. The API bill was… educational.

Why It Happens

Security isn't the exciting part. People rush through setup to get to the "talking to AI" part and skip the "locking the door" part. Default configurations sometimes prioritize ease of setup over security.

How to Avoid It

Set a gateway auth token. OpenClaw supports token-based authentication for the gateway. Set it during onboarding or in your config:

openclaw config set gateway.token "your-secret-token-here"

Configure your firewall. Only expose the ports you need:

sudo ufw allow ssh
sudo ufw allow 443/tcp  # HTTPS if needed
sudo ufw enable

Don't bind to 0.0.0.0 unless you need to. If the gateway only needs to be accessible from localhost (which is often the case), bind it to 127.0.0.1.

Use fail2ban. It's free, it's easy, and it blocks brute force attempts:

sudo apt install fail2ban
sudo systemctl enable fail2ban

Read our security guide for the full picture.

3. Using OAuth Instead of API Keys

The Problem

Some people set up OpenClaw with OAuth tokens for services like Google, GitHub, or Slack. OAuth tokens expire. When they expire at 3 AM, your AI assistant silently breaks.

Worse: OAuth token refresh can create race conditions. If two processes try to refresh the same token simultaneously, one gets a new token and the other's refresh token is invalidated. Now you have an auth error that requires manual intervention.

We've seen this pattern repeatedly:

1. Set up OAuth for a service
2. Works great for 1-6 hours
3. Token expires
4. Refresh race condition
5. AI assistant is broken, user doesn't notice for hours
6. Manual re-authentication required

Why It Happens

OAuth is the "standard" auth method for many services, and web tutorials default to it. It's also what you get when you click "Sign in with Google" in a setup wizard. API keys feel old-school by comparison.

How to Avoid It

Use API keys wherever possible. They don't expire (or expire on long timelines), they don't have refresh race conditions, and they work reliably for always-on services.

For AI providers specifically:

• Anthropic: Use API keys (from console.anthropic.com)
• OpenAI: Use API keys (from platform.openai.com)
• Google services: Use service account keys instead of OAuth where possible

When you must use OAuth (some services don't offer alternatives), implement proper token storage with file locking to prevent race conditions, and set up monitoring for auth failures.

See our OAuth troubleshooting guide for more details.

4. Ignoring Context Window Costs

The Problem

AI models charge per token — both input and output. The context window is everything sent to the model: system prompts, conversation history, memory files, and your actual message.

Here's where it gets expensive: by default, OpenClaw loads your conversation history into the context. A long conversation means a large context. A large context means more tokens per message. More tokens means higher costs.

We've seen users shocked by API bills because:

• Their conversation sessions never reset, accumulating thousands of messages
• Large memory files (50KB+ MEMORY.md) were loaded into every single API call
• Group chats with many participants created huge contexts
• File uploads (PDFs, images) bloated the context

One user had a 100-message conversation session in a group chat. By the end, every new message was sending the entire 100-message history to Claude. The last message cost more than the first 50 combined.

Why It Happens

Context costs are invisible. You don't see a price tag on each message. API dashboards show daily totals, not per-message costs. It's easy to not notice until the monthly bill arrives.

How to Avoid It

Reset conversations regularly. Use /reset in Telegram or the equivalent in your channel to start a fresh context. You don't lose memory (MEMORY.md is separate), just conversation history.

Keep memory files lean. Review MEMORY.md periodically. Remove outdated information. Aim for under 10KB for daily-driver memory.

Use topics/threads. In Telegram and Discord, separate topics maintain separate contexts. Your coding discussion doesn't carry the context of your grocery list.

Monitor usage. Check your API provider's dashboard weekly:

• Anthropic usage
• OpenAI usage

See our API costs guide and cost reduction guide for detailed strategies.

5. Not Setting Up Auto-Updates

The Problem

OpenClaw gets regular updates — bug fixes, security patches, new features. If you install it once and never update, you're running outdated software on an internet-connected server.

This isn't hypothetical. Security vulnerabilities in Node.js dependencies are discovered constantly. An unpatched server is an invitation.

Beyond security, missed updates mean missed features and missed bug fixes. That Discord resume loop that was fixed three versions ago? You're still hitting it because you're running the version from six months ago.

Why It Happens

"If it ain't broke, don't fix it." The AI is working, conversations are flowing, why touch it? Updating feels risky — what if the update breaks something?

Also, there's no auto-update mechanism built in. You have to manually npm update or git pull, which means you have to remember to do it.

How to Avoid It

Set up a weekly update check:

# Add to crontab (crontab -e)
0 4 * * 0 cd /path/to/openclaw && npm update -g openclaw && openclaw gateway restart

This updates OpenClaw every Sunday at 4 AM and restarts the gateway. Simple and effective.

Keep your OS updated too:

sudo apt install unattended-upgrades
sudo dpkg-reconfigure unattended-upgrades

This automatically installs security patches for your OS — the single most impactful security measure for any server.

Subscribe to release notifications. Watch the OpenClaw repo on GitHub to get notified of new releases. Major updates might need manual intervention, but patch updates are usually safe to auto-apply.

The Meta-Mistake

There's a sixth mistake that encompasses all of these: treating self-hosting as "set it and forget it."

A self-hosted AI assistant is a server. Servers need maintenance: updates, monitoring, backups, security reviews. If you don't want that responsibility, that's completely valid — it's exactly why managed services like lobsterfarm exist.

But if you choose to self-host, commit to the maintenance. Thirty minutes a week keeps everything healthy. Skip it, and you'll eventually hit one of these five mistakes.

Try lobsterfarm (managed) → · Self-host on Hetzner →