OpenClaw: How to Set It Up, Pick the Right AI Model, and Actually Get It to Do Useful Stuff

I've been running OpenClaw on my machine for about six weeks now, and the honest truth is that I didn't fully get it until about day four - but when it clicked, it really clicked.

This thing started as a hobby project by an Austrian developer named Peter Steinberger in November 2025. By late January 2026 it had 247,000 GitHub stars and people were buying Mac Minis specifically to run it. That trajectory should tell you something about whether it's worth your time.

What Is OpenClaw?

OpenClaw is an open-source personal AI assistant that runs as a persistent daemon on your own machine, connects to the messaging apps you already use, and executes real multi-step tasks - not just chat.

It's built on Node.js and works as an orchestration layer. It routes messages from WhatsApp, Telegram, Discord, Slack, Signal, iMessage - over 50 channels total - through an AI model you configure, then executes actions: browsing the web, managing your calendar, processing email, controlling smart home devices, writing and running code. The model itself lives elsewhere (via API or locally via Ollama). OpenClaw is just the brain stem connecting everything together.

In short: I talk to it through the apps I already have open. It does the thing. The AI model is the reasoning engine - OpenClaw is the runtime that gives it hands.

Why OpenClaw Matters in 2026

1. The agentic shift is actually happening this time

Every year since 2022 we've been promised "autonomous AI agents." Every year it turned out to mostly mean "AI that tries one thing, gets confused, and stops." OpenClaw is the first project I've used where the agent loop actually holds up across multi-step tasks without hand-holding. I've watched it summarize a thread of 40 emails, draft a reply, and add three calendar events from one Telegram message. That's not a demo. That's what it does on a Thursday afternoon.

2. It's model-agnostic, which matters more than it sounds

You're not locked into one AI provider. Claude, GPT-4o, DeepSeek, Gemini, or local models via Ollama - you configure which one you want and can switch any time without losing your history. This is a bigger deal than people initially realize because model pricing changes constantly. GPT-4o dropped 50% in price in late 2025. If you'd been locked in somewhere else, you'd have missed that.

3. The skills ecosystem is moving fast

OpenClaw uses a "skills" system - essentially plugins that live as directories containing instructions and tools. ClawHub, the community skills registry, already has over 5,700 extensions. Someone built a full CRM lead-generation workflow. Someone else built a skill that monitors their server health and messages them on Telegram. I've found four skills I use daily that I didn't even know existed until I went looking.

4. Privacy is a genuine first-class concern

Everything - conversation history, config, memory - stays on your machine. If you're running local models via Ollama, not a single byte of your data ever hits a cloud server. For anyone who touches client data, handles contracts, or just values not having their private conversations indexed by a corporation, this architecture matters.

How to Deploy OpenClaw on Your Own PC

This is the part most guides rush through. I'll be real - the first version I installed was a mess because I skipped the wizard and tried to manually configure everything. Don't do that. Use the wizard.

Step 1: Check your Node version

OpenClaw requires Node 22 or higher. This trips people up.

node --version

If you're on anything below v22, update first. The one-line installer on their site handles this automatically, but if you're doing it manually, sort Node first.

Step 2: Install OpenClaw globally

npm install -g openclaw@latest

Or if you use pnpm:

pnpm add -g openclaw@latest

Step 3: Run the onboarding wizard

This is the step most tutorials skip past with a single line. The wizard is genuinely good - run it properly.

openclaw onboard --install-daemon

The --install-daemon flag installs the Gateway as a persistent background service (launchd on macOS, systemd on Linux, Scheduled Tasks on Windows). Skip this flag and your assistant shuts down when you close the terminal. Don't skip it.

The wizard will ask you which AI provider you want to use, then walk you through getting your API key configured. Have your Anthropic or OpenAI API key ready before you start.

Step 4: Start the Gateway and verify it's running

openclaw gateway --port 18789 --verbose

Then check status:

openclaw gateway status --deep

If it's running, you'll see your port, uptime, and connected channels. If something looks wrong, run openclaw doctor - it's like a self-diagnostic that tells you exactly what's misconfigured.

Step 5: Connect a messaging channel

The onboarding wizard covers this, but the general pattern is: pick a channel (I started with Telegram because it's the easiest), create a bot via BotFather, paste the token into OpenClaw. Done in about four minutes.

Once connected, send: openclaw agent --message "what can you do?" --thinking high

That --thinking high flag is important. It tells the model to actually reason through its response rather than just pattern-match. Use it whenever you're setting something up or debugging.

Step 6: Explore the skills registry

openclaw skills search

Browse what's available, install what looks useful. This is where OpenClaw goes from "interesting toy" to "thing I actually rely on."

Best AI Models for OpenClaw

Here's what I actually use and why - and where different models genuinely pull ahead.

The model-selection question is where I see the most confusion online. People ask "which model is best for OpenClaw?" and the honest answer is: it depends entirely on what you're asking it to do. Driving a Ferrari to the grocery store isn't a flex, it's just expensive.

Model	Best For	Price (Input/Output per 1M tokens)
Claude Opus 4.6	Complex multi-step reasoning, sensitive tasks, highest accuracy	$15 / $75
Claude Sonnet 4.6	Daily assistant work, email, calendar, general tasks	$3 / $15
Claude Haiku 4.5	Simple/fast responses, reminders, high-volume cheap tasks	~$0.25 / $1.25
GPT-4o	All-around solid, fast on simple tasks, strong code generation	~$2.50 / $10
DeepSeek V3.2	Budget-first users, routine automation, code generation on a tight budget	$0.28 / $0.42
Gemini 3 Flash	Free tier available, high-volume simple tasks	Free tier / low
Llama 3.3 70B (Ollama)	Maximum privacy, no API cost - needs 48GB+ RAM	$0
Mistral 7B (Ollama)	Low-RAM local option - marginal for serious agent work	$0

The community data backs this up: about 55% of active OpenClaw users run Anthropic models, 25% run OpenAI, and 20% run local or other providers. The reason Claude dominates isn't loyalty - it's that tool-calling reliability and long-context consistency genuinely hold up better under real agent workloads. Peter Steinberger, the project's creator, explicitly recommends Anthropic Pro or Max with Opus 4.6 for serious agent work.

That said - Claude Sonnet 4.6 is my daily driver. Opus is overkill for email processing and calendar management, and at $75/M output tokens, you'll feel that bill if you're running the agent hard all day.

For local models: I've found that anything below 32B parameters struggles with reliable tool-calling chains. 7B and 8B models are fine for one-off questions but fall apart on multi-step agent tasks. If you're going local, you need either a 32GB Mac Mini ($1,199 new) running something like Qwen3-Coder:32B, or accept that you'll be using cloud APIs for anything complex.

Best Practices

1. Don't skip --install-daemon on setup

I made this mistake on my first install. Without the daemon, your assistant dies the moment you close your terminal. Everything you've set up - channels, memory, skills - still works, it just won't respond to anything. The --install-daemon flag is the difference between a toy and a persistent agent.

2. Treat the SOUL.md file seriously

OpenClaw has a personality configuration file called SOUL.md. Most people fill it in with a few words and move on. I've found that the more specific and detailed your SOUL.md, the more consistently the agent behaves across long conversations - especially with smaller models that tend to drift from instructions over time. Spend 20 minutes on it.

3. Route tasks to the right model tier

This is the part nobody talks about enough. OpenClaw supports model routing - you can assign different models to different task types in your config. I run Haiku for simple reminders and weather checks, Sonnet for email and calendar, and only invoke Opus for things like multi-document analysis or complex research tasks. This alone cut my monthly API bill roughly in half.

4. Run openclaw doctor any time something feels off

Weird silent failures, the agent stopping mid-task, channels not responding - most of these are config issues, not model issues. openclaw doctor catches the majority of them in seconds. I wasted two hours once debugging what turned out to be a port conflict that doctor would have flagged immediately.

5. Be very selective about third-party skills

Cisco's AI security team tested a third-party OpenClaw skill and found it was performing data exfiltration and prompt injection without any visible indication. This is real. Read the code of any skill before you install it - ClawHub has no meaningful vetting process right now. Stick to skills with significant stars and recent commits, and audit anything that requests access to your email or files.

6. Track your token usage for the first two weeks

OpenClaw sends full conversation history on every request. Long sessions burn tokens fast. I didn't realize this until my first month's bill showed up and I'd used nearly three times what I expected. Check your usage in the OpenClaw dashboard or your provider's console after the first week and you'll have a much better sense of what your real monthly cost looks like.

The Debate

The most serious criticism I've heard about OpenClaw is the security argument - and I want to be honest here because it's not a bad-faith concern.

OpenClaw, by design, needs broad permissions to function. Access to your email, calendar, messaging platforms, file system. One of the project's own maintainers posted on Discord that "if you can't understand how to run a command line, this is far too dangerous of a project for you to use safely." That's a real warning from inside the project. China's government moved in March 2026 to restrict state agencies from running it, citing security concerns. Cisco found a third-party skill doing exfiltration in tests.

I take all of this seriously. My honest position: OpenClaw is powerful enough that the security surface is genuinely larger than most people assume when they first install it. The prompt injection risk is real - if you're feeding it content from untrusted sources (public websites, emails from strangers), a well-crafted injection in that content can redirect the agent's behavior in ways that aren't obvious.

Where I land: use it with intentionality. Don't install random skills. Sandbox the permissions you grant it. Keep sensitive accounts in separate profiles that OpenClaw doesn't touch. Don't run it on a machine with access to production systems unless you really understand the risk surface. The tool is genuinely useful - it's just not the kind of tool you set up carelessly and forget about.

What I Think Happens Next

1. Model routing will become the standard configuration pattern. Right now, most users just pick one model and leave it. Within 6 months I expect per-task model assignment to be a first-class UI feature, not something you configure in YAML. The cost savings are too compelling for the community not to build it.

2. The security tooling will catch up, slowly. The community knows the prompt injection problem exists. Skills sandboxing and permission scoping will improve. I'm not holding my breath for it to be solved quickly, but it's on the roadmap and it'll get incrementally better.

3. Local model quality will close the gap on simple tasks. Every 3 months the open-weight models get meaningfully better. By end of 2026, running a local 32B model for 80% of your daily agent tasks - with cloud fallback for complex reasoning - will be a reasonable production setup on mid-range hardware.

4. The skills ecosystem will bifurcate. There'll be a vetted, high-trust tier and a community free-for-all tier. The Cisco finding accelerated that conversation and I think some form of verified skills registry is coming sooner than later.

5. OpenClaw's lead won't last forever. The codebase is moving incredibly fast and the community is genuinely impressive, but Microsoft, Google, and Apple all want to own the "personal AI agent" category. When they ship their versions - and they will - the comparison will be instructive. I'm genuinely not sure OpenClaw wins that fight long-term. But it'll absolutely shape what those products look like.

How I'd Get Started If I Were You

1. Make sure Node 22+ is installed - check with node --version before you touch anything else.
2. Run npm install -g openclaw@latest and then openclaw onboard --install-daemon - let the wizard do its job.
3. Start with Claude Sonnet 4.6 as your model - get an Anthropic API key from console.anthropic.com, drop it in when the wizard asks. Sonnet is the right first model: capable, fast, and it won't run up a scary bill while you're still figuring out what you actually want.
4. Connect Telegram first - it's the lowest-friction channel to set up and the OpenClaw community has the most documentation written for it.
5. Install one skill that solves an actual problem you have - not five skills to experiment, one skill for something you actually do repeatedly. This is how it stops feeling like a toy.
6. Check your token usage after seven days - you'll have real data to optimize from, and you'll know whether you need to adjust model routing or usage patterns before you get a surprise bill.

The personal AI agent era that everyone's been predicting for three years is here, and it arrived looking like a lobster mascot on GitHub. Don't wait for a polished product from a big company - OpenClaw works today, it's configurable, and you own the data. That combination is harder to find than it sounds.