OpenClaw: What It Actually Does and How to Run It Safely

OpenClaw went from zero to 304,000+ GitHub stars in weeks, making it one of the fastest-growing open-source projects ever. Created by Peter Steinberger (who joined OpenAI in February 2026), it's a self-hosted AI agent that connects to your messaging channels and actually does things, not just chats. You might know it under its earlier names: Clawdbot or Moltbot. Here's everything you need to know before running it.

Quick Summary

Use OpenClaw if you want an autonomous AI agent that acts on your behalf across messaging platforms. Skip it if you just want a better chatbot, ChatGPT does that already.

What Is OpenClaw, Really?

OpenClaw is not a ChatGPT replacement. It's an autonomous agent framework, and that distinction matters more than most articles let on.

Here's the simplest way to think about it: ChatGPT is a brain in a jar. You ask it questions, it answers. OpenClaw gives that brain hands, eyes, and a phone. It can send messages on your behalf, browse the web, run code, manage files, and connect to APIs, all without you manually prompting each step.

The architecture works like a gateway: one agent brain connects to multiple messaging channels (Telegram, WhatsApp, Discord, Slack, and more). You configure its personality in SOUL.md, its capabilities in AGENTS.md, and your personal context in USER.md. The agent reads these files, connects to your chosen LLM (OpenAI, Anthropic, or a local model via Ollama), and starts making decisions.

How is this different from n8n or Zapier? Those are deterministic workflow tools, "if X happens, do Y." OpenClaw makes autonomous decisions. You say "handle my morning emails" and it figures out the steps. That's powerful, and also why the security conversation matters so much.

If you're exploring the broader landscape of AI coding assistants like Claude Code, OpenClaw sits in a different category entirely, it's not about writing code, it's about taking actions across your digital life.

Verdict: OpenClaw is genuinely different from ChatGPT, it's an agent, not a chatbot. Understand the distinction before you set it up.

How Do You Set It Up?

You need Node.js 22+ and basic terminal comfort. There are two paths.

Quick Install (npx)

# Fastest way to get running
npx openclaw@latest

# The onboarding wizard walks you through:
# 1. Your agent name
# 2. Model provider (OpenAI, Anthropic, or Ollama)
# 3. API key
# 4. First channel connection

Docker Install (Recommended)

# Better for persistence and security isolation
docker run -d --name openclaw \
  -v openclaw_data:/app/data \
  -p 3000:3000 \
  openclaw/openclaw:latest

Docker gives you process isolation out of the box and makes it much easier to restrict network access, which matters a lot (see the security section).

First Channel Connection

Start with Telegram for testing. It's the easiest to configure, you create a bot via BotFather, paste the token, and you're chatting with your agent in under two minutes. WhatsApp requires a Business API or linked device session, which adds friction.

Once the agent is running, visit localhost:3000 for the web dashboard. From there you can edit SOUL.md (the agent's personality), add channels, and install skills.

Verdict: Setup takes 15-30 minutes. Docker is worth the extra 5 minutes for persistence and easier security hardening.

Do You Really Need a Mac Mini?

Short answer: no. This question comes up constantly in OpenClaw communities, and the answer depends entirely on whether you're using cloud APIs or running local models.

With cloud API mode (OpenAI, Anthropic), OpenClaw itself is lightweight, roughly 200-500MB RAM. The heavy lifting happens on the provider's servers. With local LLM mode (Ollama), your RAM needs jump dramatically because you're loading the entire model into memory.

For most people, a $6-12/month VPS handles cloud API mode perfectly. DigitalOcean's $6/month droplet (1GB RAM) runs a single cloud-API agent fine. Bump to $12/month (2GB) for two or three agents. Only consider a dedicated Mac Mini or $24+/month VPS if you're running local models through Ollama and genuinely need zero API costs.

Verdict: You don't need a Mac Mini. A $6-12/month VPS handles most setups. Go local-LLM only if you have real privacy requirements or want zero API costs.

How Much Will It Actually Cost You?

OpenClaw itself is free and open source. The costs come from two places: the LLM API calls and your hosting.

A few things to keep in mind. Costs scale with usage -- 50 messages a day is very different from 500. The "free" framing around OpenClaw is technically true but misleading unless you're running local models on your own hardware. Budget $10-15/month for a genuinely useful cloud-API setup.

One important caveat: using the Claude API through OpenClaw may technically brush against Anthropic's Terms of Service. Anthropic banned OAuth tokens for third-party tools in January 2026, citing subscription arbitrage concerns. You can still use the API with a paid API key, but it's worth knowing the policy lines before you commit to Claude as your backend.

Verdict: Budget $10-15/month for a useful setup. The "free" framing is misleading unless you run local models on your own hardware.

Is OpenClaw Safe to Use?

Here's the honest answer: OpenClaw has real security risks if you don't configure it properly. And the defaults are too permissive.

The numbers are sobering. Kaspersky found 512 vulnerabilities in their audit, including 8 critical ones. CVE-2026-25253 is a one-click remote code execution vulnerability with a CVSS score of 8.8. CrowdStrike documented over 40,000 publicly exposed instances, with roughly 12,800 directly exploitable due to disabled authentication. And on ClawHub (the skills marketplace), 1,184 malicious skills were discovered as part of the ClawHavoc campaign, some deploying the Atomic macOS Stealer targeting browser credentials and crypto wallets.

Real incidents have already happened: inbox wipes, dating profile manipulation, cryptocurrency wallet targeting.

So should you run it? Yes, but harden it first. Here's the minimum:

# gateway config -- restrict binding
gateway:
  host: "127.0.0.1"  # NOT 0.0.0.0
  port: 3000
  auth:
    enabled: true

Five steps that take about 10 minutes:

1. Bind the gateway to 127.0.0.1 (never 0.0.0.0)
2. Enable authentication on the web interface
3. Use Docker for process sandboxing
4. Read the YAML of any ClawHub skill before installing it
5. Set channel-specific permissions, limit what the agent can do per channel

Microsoft's security guidance goes further, recommending you treat OpenClaw as "untrusted code execution with persistent credentials" and deploy it only in an isolated VM. That's enterprise-grade advice, but the principle is sound: don't run OpenClaw on a machine with your SSH keys and crypto wallets.

Verdict: OpenClaw is safe IF you harden it. The defaults are too permissive. Spend 10 minutes on security config before connecting any real accounts.

What Can You Actually Do With It?

Here are five use cases where OpenClaw genuinely shines, ranked by setup difficulty:

• Daily briefings via Telegram (Easy), Configure a morning summary: weather, calendar, top news in your niche, and pending tasks. Takes 10 minutes to set up with built-in skills.

• Email triage and drafting (Easy), Connect your inbox, let the agent categorize incoming mail, draft replies for routine messages, and flag anything that needs your actual attention.

• Code review assistant (Medium), Connect to GitHub via skills. The agent watches for new PRs, reviews diffs, and leaves comments. Not a replacement for human review, but a solid first pass that catches obvious issues.

• Content monitoring (Medium), Watch RSS feeds, summarize new articles in your industry, and post highlights to a Slack channel. Great for staying current without doomscrolling.

• Home automation bridge (Advanced), Connect Home Assistant or other smart home APIs to a natural language interface via Telegram. "Turn off the living room lights and set the thermostat to 20" actually works.

The skills ecosystem on ClawHub has thousands of community-built skills covering everything from server monitoring to social media management. But remember: audit before you install. Read the YAML source of any skill before running it, the ClawHavoc campaign proved that the marketplace isn't immune to supply-chain attacks.

You can also write custom skills using Markdown-based SKILL.md files. The official skills documentation covers the format.

Verdict: OpenClaw shines for personal automation and multi-channel messaging. It's overkill for simple chatbot use cases, use ChatGPT for that.

Troubleshooting, When Things Go Wrong

No competitor covers this, so here are the most common issues from GitHub Issues and community forums:

If you hit something not on this list, the OpenClaw GitHub Issues page is active and the community is responsive. Search before posting, most common problems already have solutions.

How Techsy Approaches AI Agent Development

At Techsy, we've been working with AI agent architectures for startup teams, evaluating tools like OpenClaw, building custom agent setups, and helping teams decide what actually makes sense for their use case.

Here's our honest take: OpenClaw is excellent for personal automation and small-team experiments. If you need a personal AI assistant that works across Telegram, Slack, and email, it's hard to beat for the price. But when teams need custom integrations, enterprise-grade security controls, or multi-agent orchestration that goes beyond what skills can provide, a custom build usually makes more sense.

Need help deciding between OpenClaw and a custom AI agent setup? Get a free consultation.

FAQ

What is OpenClaw?

OpenClaw is an open-source, self-hosted autonomous AI agent framework. It connects to your messaging channels (WhatsApp, Telegram, Discord, Slack) and takes actions on your behalf, not just answering questions, but sending messages, managing files, browsing the web, and running code. It uses an LLM (like GPT-4o or Claude) as its "brain."

Is OpenClaw free?

The software itself is free and open source. Running it costs between $0 and $90/month depending on your LLM choice. A local model via Ollama costs nothing. A cloud API setup with GPT-4o-mini runs about $10-15/month. Power usage with Claude Opus can hit $40-90/month.

Is OpenClaw safe to use?

Yes, if you harden the defaults. Bind the gateway to 127.0.0.1, enable authentication, use Docker sandboxing, and audit any ClawHub skills before installing. Without hardening, you're exposing yourself to real risks, Kaspersky documented 512 vulnerabilities, and over 40,000 instances were found publicly exposed.

How much does OpenClaw cost per month?

Realistically, $10-15/month for a useful cloud-API setup (VPS hosting + GPT-4o-mini). $25-35/month for Claude Sonnet quality. $0 if you run a local model on your own hardware. See the cost table above for a full breakdown.

Is OpenClaw better than ChatGPT?

They're different tools. ChatGPT is a conversational AI you interact with. OpenClaw is an agent layer that uses ChatGPT (or Claude, or local models) as its brain while taking autonomous actions across your messaging channels. Use ChatGPT for Q&A, use OpenClaw for automation.

What AI models work with OpenClaw?

OpenAI (GPT-4o, GPT-4o-mini), Anthropic (Claude Sonnet, Opus), and local models via Ollama (Qwen, Llama, Mistral). Any OpenAI-compatible API endpoint works too, so you can plug in providers like DeepSeek or Together AI.

Can I run OpenClaw on a Raspberry Pi?

Technically yes for cloud API mode, OpenClaw is lightweight (~200-500MB RAM). But performance will be limited and you'll hit storage constraints quickly. A $6/month VPS is more practical and gives you better uptime.

What is ClawHub?

ClawHub is the skills marketplace for OpenClaw, community-built extensions that add capabilities like GitHub integration, RSS monitoring, or smart home control. There are thousands of skills available. Be careful though: 1,184 malicious skills were discovered in the ClawHavoc campaign. Always read the source before installing.

What happened to Clawdbot and Moltbot?

Same project, renamed twice. The progression was Clawdbot, then Moltbot, then OpenClaw. The creator, Peter Steinberger, joined OpenAI in February 2026, and the project transitioned to foundation-backed open source governance.

Can I use OpenClaw with WhatsApp, Telegram, and Discord?

Yes, all three are supported as channels. Telegram is the easiest to set up for testing, just create a bot via BotFather and paste the token. WhatsApp requires a Business API or linked device session. Discord uses a standard bot token. You can run multiple channels simultaneously from a single agent.

Sources

• OpenClaw GitHub Repository
• OpenClaw Official Documentation
• Peter Steinberger Joins OpenAI, TechCrunch
• OpenClaw Vulnerabilities, Kaspersky Blog
• Security Teams and OpenClaw, CrowdStrike Blog
• Running OpenClaw Safely, Microsoft Security Blog
• ClawHavoc Malicious Skills, CyberPress
• Anthropic Third-Party Tool Ban, The Register