Hey, it's Cindy 🌱 You commented CLAW, so here is the whole no-hardware setup. There are two routes that cost you nothing in hardware, and this guide walks both of them.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the context
What OpenClaw actually is
OpenClaw (previously named Clawdbot, then Moltbot) is an open source personal assistant that runs on a machine you control instead of inside someone else's chat window. It has persistent memory across conversations, it can drive a browser, it can read and write files and run commands, and you talk to it through a normal chat app rather than a dashboard. Its repo is at github.com/openclaw/openclaw.
The reason it went viral is what people started pointing it at. There is now a marketplace called Rent a Human where agents post paid tasks for real people to do in the physical world, with MCP and REST integrations so a bot can hire someone without a human in the loop. The site claims hundreds of thousands of registered humans across 100+ countries. Whatever you make of that, it is worth understanding what these things are becoming rather than reading about it for another six months.
The scary part of the setup story is a myth. OpenClaw's own install docs say it needs no dedicated hardware and runs on standard computers, including the laptop you are reading this on. The Mac mini thing is a preference some people have for leaving it running 24/7, not a requirement.
pick your route
Two ways in, no hardware
Route A: hosted in your browser Emergent spins the whole thing up for you from a template. No terminal, no config files, nothing installed on your computer. Best if you want to try it this weekend.
Route B: on the computer you already own One install command on your existing Mac, Windows or Linux machine. Free, fully yours, and it stops when your laptop sleeps. Best if you are comfortable with a terminal.
route a
The browser route, start to finish
🔗 Open Emergent. Go to Emergent and sign in with Google. You land on a home screen with a row of template chips.
🎯 Click the MoltBot chip. That is the OpenClaw template. Clicking it pre-fills the entire setup prompt for you, so you are not writing any configuration yourself.
▶️ Run the prompt. Press enter and leave it alone. Emergent provisions the infrastructure in the background. Their own tutorial puts this at around five minutes, sometimes two, sometimes eight. You do not need to babysit it.
🔑 Take the default keys, or bring your own. Emergent supplies default model keys so it works out of the box. If you already have an Anthropic API key you can add it instead, which gives you direct control over which model your agent runs on.
✅ Wait for "OpenClaw started successfully". That message is your green light. Your agent now exists and is reachable.
🚀 Hit Deploy if you want it always on. The Deploy button in the top right keeps the agent running continuously, which is what makes scheduled tasks and background automation possible. Skip this while you are still testing, because a running agent consumes credits.
Emergent runs on a credit system and the free allowance is small, so an agent you leave deployed and chatting all week will burn through it. Check the current plans on Emergent's pricing page before you deploy, and test in short sessions first.
make it reachable
Connect it to your chat app
This is the part that makes it feel real. Instead of opening a dashboard, you text your agent like a person.
🤖 Telegram: make a bot first. Open Telegram, message @BotFather, send /newbot, and follow the two prompts to name it. BotFather hands you an API token.
🔐 Telegram: pair it. Paste that token into Emergent, then send your new bot a test message. It replies with a pairing code. Enter the code back in Emergent and the connection is confirmed.
📱 WhatsApp: ask the agent to do it. Type "help me connect my WhatsApp" into the agent chat, then go to the channels or integrations section, choose WhatsApp, and click Show QR. Scan the code with WhatsApp on your phone and enter your number with the country code.
🧪 Send it one message. Say hello from your phone. If it answers, you are done with setup and everything after this is you deciding what it should do.
Use a spare or secondary number for WhatsApp while you are testing. Connecting your main personal number to anything with autonomous message access is a decision worth making deliberately, not by accident on a Saturday afternoon.
route b
Or run it on your own laptop
If you would rather it live on your machine, with your files and no per-credit cost, OpenClaw installs with a single command. You need Node.js 22.22.3 or newer (Node 26 is what the docs recommend) and no special hardware.
💻 macOS, Linux or WSL2. Run curl -fsSL https://openclaw.ai/install.sh | bash in your terminal.
🪟 Windows PowerShell. Run iwr -useb https://openclaw.ai/install.ps1 | iex. There is also a native Hub app if you would rather not touch PowerShell at all.
📦 Or use a package manager.npm install -g openclaw@latest --allow-scripts=openclaw works if you already have npm set up. Docker, Podman and Kubernetes deployments are documented too.
🧠 Point it at a model. Add an API key from your provider and set your default model. This is the step the next section is about.
Not sure which route is yours? Start with the browser route this weekend to see whether you actually like using an agent this way, then move it onto your own machine once you know what you want it doing. Trying it costs an evening. Buying hardware first costs a lot more.
the important bit
OpenClaw is the wrapper, not the brain
This is the part the viral clips skip. OpenClaw handles the plumbing: the memory, the chat connections, the browser control, the permissions. The actual reasoning comes from whichever model you point it at, and it supports 50+ providers, so the answer to "is this thing smart" is really "which model did you give it".
Its documentation tells you to use the strongest latest-generation model available, and the quick start in the providers docs sets the default with anthropic/claude-opus-4-6. That is what I would run it on too. A cheap model in an agent loop rarely saves you money, because it fails more often and burns extra tokens retrying.
copy these
Give it one real job
The mistake almost everyone makes is setting the agent up and then never giving it anything to do. Paste this into your agent chat as its first proper task.
🎯 my first real agent job
I want to give you one real task this weekend so I can see what you are actually capable of. Here is the job: [describe one small real task, for example "find three restaurants near me that take bookings for four people on Saturday night, compare them, and tell me which to book and why"].
Before you start, tell me exactly what you plan to do, which tools or websites you will use, and anything you will need permission or information from me for. Wait for me to say go.
While you work, message me at each meaningful step rather than going silent and returning twenty minutes later. If you get blocked, tell me what blocked you and what you would need to get past it, rather than guessing or making something up. At the end, tell me honestly which parts you did well and which parts a human would still have to redo.
Then use this one to work out what is genuinely worth handing over, instead of copying someone else's demo.
💡 find my three best agent tasks
I have a personal AI agent set up that can browse the web, read and write files, run commands, and message me on Telegram or WhatsApp. I want to find the tasks in my own life that are genuinely worth handing to it.
Here is my week: [describe your job, the admin you repeat, the things you keep forgetting, and the tools you already use]. Ask me any questions you need before answering.
Then give me three specific tasks worth automating, ranked by how much time they save against how likely they are to go wrong. For each one, describe exactly what the agent would do, what you would need me to approve before it acts, and one thing that could go badly so I can decide whether I want it running unsupervised. Rule out anything involving money, passwords or messages to other people unless it stops for my approval first.
read before you run it
Three safety rules
This thing has real access. Files, shell commands, a browser, and your chat apps. That is exactly why it is useful and exactly why it is different from a chatbot. Give it the narrowest access that lets it do the job, and never point it at a machine holding work data, client files or anything you cannot afford to lose.
Nothing autonomous with money or other people. Keep a human approval step in front of anything that spends, sends, posts, or messages someone who is not you. Reading and drafting are safe. Acting on the outside world is where an agent's confident mistake becomes your problem.
Watch what it costs while it is learning. On the hosted route that means credits, on your own machine it means API spend. Either way, run short sessions and check the usage after the first few real tasks, before you leave anything deployed and running all week.