Build Your Own AI Agent From Zero: OpenClaw + Claude Code, the Honest Beginner's Guide

I run a personal AI agent on a Mac mini in my house. I talk to it on Telegram. By the time I wake up it has done the morning research for my content brands, and through the day it drafts scripts and captions, runs video renders and queues posts. It has been running, in one form or another, for most of this year.

This guide is how you build one from nothing. It's written for someone who has used ChatGPT and has never opened a terminal, so every step is spelled out — including what a terminal is. If you already know, skip ahead; the sections are labelled.

Three promises before we start. Every command here was checked against the official documentation this week. Every price is dated. And I'm going to be straight with you about money, because that's where this topic is buried in nonsense.

First, the honest part about money

The internet is full of people telling you an AI agent will earn money while you sleep. Here's the most famous real example, with verified numbers.

In February, writer Nat Eliason gave his OpenClaw agent, Felix, $1,000 and the job of building a business. Felix built a website, a $29 guide, and a marketplace for OpenClaw skills called Claw Mart. The story went everywhere. And the revenue was real: Claw Mart's dashboard on TrustMRR, which pulls directly from Stripe, shows $182,552 in all-time revenue.

It also shows $1,866 in the last 30 days.

Look at what happened there. The money came from selling OpenClaw know-how to OpenClaw enthusiasts at the peak of the hype, and when the hype cooled, so did the revenue. MIT Technology Review found the same pattern in China in March: the people reliably making money from agents were the ones charging $15 to $100 to install them for other people. Picks and shovels.

That doesn't mean agents are useless for making money. It means they make money the way a good employee does, not the way a slot machine does. The quieter stories are the real ones. A man who runs a maintenance-gardening business told Hacker News his agent now writes his 14-to-32-page job proposals, handles scheduling and analyses photos customers send in — and that it "freed up my home life." He didn't claim it earned him a dollar. It gave him his evenings back while the business kept running.

That's the frame for this whole guide. An agent doesn't make money. It lets one person run work that would otherwise need staff. If you have a business, a side project or a skill you sell, that leverage is worth a lot. If you have nothing for it to work on, it will cheerfully burn API credits doing nothing useful.

Mine has never made me a dollar on its own. It runs the operations behind four content brands that I could not otherwise run alongside a full-time job. That's the honest value proposition, and I'd be suspicious of anyone selling you a different one.

The two tools, and why you want both

OpenClaw is the worker. It's free, open-source software that turns an AI model into a personal assistant that lives on your own computer, runs all the time, remembers things, follows schedules, and talks to you through apps you already use — Telegram, WhatsApp, Discord, Slack, iMessage and about twenty others. You give it jobs. It does them.

Claude Code is the mechanic. It's Anthropic's AI that works directly on your computer: it can read files, run commands, read error logs, and make careful changes. You'll use it to install OpenClaw, configure it, write new abilities for it, and — most importantly — fix it when it breaks. And it will break.

Why two? Because the hardest part of running an agent isn't the first install. It's the fourth week, when an update changes something and your agent stops responding, and you're staring at a configuration file you don't understand. OpenClaw's own documentation now recommends exactly this setup: when you're stuck, use a local AI agent "that can see your machine" — and it names Claude Code.

I learned this the slow way. I ran my agent pinned to an old version (2026.3.13) for months, because every newer version I tried broke how it used its tools. This month I finally upgraded to 2026.9.7, and installed Claude Code on the same machine with a runbook so the next upgrade isn't a months-long standoff. That's the job you're hiring the mechanic for.

A note on who runs OpenClaw, since you'll see confusing claims. Its creator, Peter Steinberger, joined OpenAI in February 2026. The project did not go with him: in July it moved to the OpenClaw Foundation, a US nonprofit. Its own README is explicit — "OpenAI is a donor, not an owner." Pages saying OpenAI bought OpenClaw are wrong.

Ten words you need first

If you've only ever used ChatGPT in a browser, these are the terms this guide assumes. Skip this if they're familiar.

What you need: machine, accounts, budget

Your agent needs a computer that stays on. OpenClaw's documentation recommends a dedicated one "for reliability and isolation," and I agree for a reason that matters more than reliability: you do not want an AI with command-line access living on the laptop that holds your bank logins and your photos. Give it its own machine, or at minimum its own user account.

Prices checked this week. Note that the $599 Mac mini you'll see in older guides no longer exists: Apple dropped it in May and the line has since moved to M6, starting at $899. Hetzner's cheapest European server is €5.49 a month after a June price change; DigitalOcean's run $6, $12 and $24.

Mine is a Mac mini M4 with 32GB. If I were starting today with no plans for local models, I'd start on a $12-a-month VPS and buy hardware only once I knew what I was using it for.

Accounts you'll need:

• A Claude Pro subscription ($20 a month) for Claude Code. The free Claude plan does not include it.
• A way to pay for your agent's model. This is a separate decision, covered in Step 4, and the most important money choice in this guide.
• A Telegram account, free.
• A separate email address for your agent, also free. OpenClaw's own security guidance says to give the bot its own email, accounts and phone number rather than handing it yours. Do this from day one.

Step 1 — Install Claude Code first

We install the mechanic before the worker, so it can help with everything after. Do this on the machine that will run your agent.

Open a terminal. On a Mac: press Command and Space, type Terminal, press Enter. On Windows: press the Windows key, type PowerShell, press Enter. Anthropic publishes a terminal guide for complete beginners if you want a gentler walkthrough.

Install. Paste the line for your system and press Enter:

# Mac, Linux, or Windows WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

When it finishes, close the terminal and open a new one, then check it worked:

claude --version

You should see a version number. If you get "command not found," the new terminal didn't pick up the install yet — Anthropic's troubleshooting page covers fixing your PATH.

Log in. Type claude and press Enter. A browser window opens; sign in with your Claude account. You'll see "Login successful."

Step 2 — Write the mechanic's runbook

Claude Code reads a plain text file called CLAUDE.md at the start of every session. It's how you stop re-explaining your setup every time. If you put it in your home folder and start claude from there, it loads automatically — that's how mine works.

Create it by typing claude, then asking: "Create a file called CLAUDE.md in my home folder with the following contents." Paste this, adjusted to your setup:

# Agent maintenance runbook

## What this machine is
- Runs an OpenClaw agent. Gateway dashboard: http://127.0.0.1:18789
- Config: ~/.openclaw/openclaw.json (strict schema — an invalid
  config stops the gateway from starting)
- Agent files: ~/.openclaw/workspace (SOUL.md, AGENTS.md, skills/)

## Always
- Back up ~/.openclaw/openclaw.json before changing anything
- Diagnose before fixing: run `openclaw status --all`,
  `openclaw doctor` and `openclaw logs --follow`, then summarise
  what you found before proposing a change
- Small changes only. Explain each command before running it.

## Never
- Never print, copy or paste API keys or tokens
- Never upgrade OpenClaw or run `openclaw doctor --fix` without asking me
- Never touch the Telegram bot token or the emergency stop script

Two things to know. Keep this file short — Anthropic recommends under 200 lines. And it's guidance, not a lock: Anthropic's documentation calls it "context, not enforced configuration." The "never" rules are requests. Your real protection is staying in Manual mode until you trust what you're looking at.

Step 3 — Install OpenClaw

Now the worker. In a new terminal:

# Mac, Linux, or Windows WSL
curl -fsSL https://openclaw.ai/install.sh | bash

# Windows PowerShell
iwr -useb https://openclaw.ai/install.ps1 | iex

The installer sets up Node.js for you if it's missing (OpenClaw needs version 24.16 or newer) and then launches a setup wizard automatically. If it doesn't, start the wizard yourself:

openclaw onboard --install-daemon

The --install-daemon part sets OpenClaw to run in the background and restart with your computer. The current stable version is 2026.9.8, released 3 October. Stay on the stable channel; beta versions are for people who enjoy debugging.

Check it's running, then open the control panel in your browser:

openclaw gateway status
openclaw dashboard

Send it a test message in the dashboard's chat. If it answers, you have an agent. If not, this is the mechanic's first job: open Claude Code and say "OpenClaw isn't responding. Follow the runbook and diagnose it."

Step 4 — Choose the brain (the most important money decision)

OpenClaw doesn't include an AI model; it drives one. It supports about seventy providers. For a beginner it comes down to three routes.

Why that last row carries a warning. On 4 April, Anthropic stopped Pro and Max subscriptions covering third-party agents including OpenClaw. In May it announced separate monthly credits for that use; in June it paused that change, and its support page now says Agent SDK, claude -p and third-party app usage "still draw from your subscription's usage limits," with advance notice promised before anything changes. OpenClaw's documentation says Anthropic staff confirmed its Claude CLI route is allowed; I couldn't find Anthropic saying so publicly. Using your subscription inside Claude Code itself is unambiguous and fine. Using it as your agent's engine is a moving target.

If you pick Claude by API key, here's what it costs today, per million tokens:

Set up fallbacks, always. Models go down, hit rate limits and run out of credit. A fallback chain means your agent keeps working on a cheaper model rather than going silent. Mine, for reference:

primary:   GPT-5.5            (via my ChatGPT sign-in)
fallback:  DeepSeek Flash     (cheap cloud model)
fallback:  Qwen 2.5 14B       (free, runs on the Mac mini via Ollama)
fallback:  Llama 3.1 8B       (free, local, last resort)

Note what that says: I write a guide recommending Claude and run GPT as my primary. Both are good, and Claude is in my stack where it matters most — as the mechanic. Pick your agent's brain on reliability and cost, not loyalty, and make sure the fallbacks span more than one provider.

You set the primary and fallbacks with the wizard, or later with openclaw models set and openclaw models fallbacks add. One trap if you use local models through Ollama: point OpenClaw at Ollama's native address, http://localhost:11434, not the /v1 one most tutorials give — the documentation says the /v1 path breaks tool use.

Step 5 — Connect Telegram

OpenClaw's own docs suggest starting with Telegram: it needs nothing but a bot token. Five minutes.

openclaw channels add --channel telegram --token YOUR-BOT-TOKEN
openclaw channels status --probe

openclaw pairing list telegram
openclaw pairing approve telegram THE-CODE-IT-SHOWS

That pairing step is a security feature, not a hoop. It means a stranger who finds your bot can't just start giving it orders. Leave it on.

Step 6 — Give it a personality and rules

Your agent's character lives in plain text files in ~/.openclaw/workspace. You can edit them in any text editor, or ask Claude Code to. The ones that matter:

The single most useful thing I put in my SOUL.md is what I call execute-first: when you have what you need to do a task, do it, then report — don't ask me five clarifying questions first. An agent that asks permission for everything is just a slower chatbot.

But that rule needs a hard boundary next to it, and getting this boundary right is the whole game:

## Act, then report
- Research, drafting, summarising, organising files in your workspace:
  just do it and tell me what you did.

## Always ask first
- Sending anything to another person: email, message, post, comment
- Spending money or signing up for anything
- Deleting anything outside your workspace
- Anything you can't undo

Internal and reversible: go. External or irreversible: ask. Write that down before you connect it to anything real.

Step 7 — Lock it down before you trust it with anything

This is the section people skip, and it's where the real stories are. Here's what happened to OpenClaw users this year.

• Tens of thousands of agents were left open to the internet. Censys counted 21,639 publicly exposed instances on 31 January; SecurityScorecard found over 40,000 days later, more than 12,000 of them exploitable.
• A single malicious link could take over an agent (CVE-2026-25253, rated 8.8 of 10), fixed in late January.
• The skills store was poisoned. Koi Security audited 2,857 skills on ClawHub in February and found 341 malicious ones, most of them installing a password-stealer through fake "prerequisite" steps.
• Agents were tricked by what they read. In June, researchers hid instructions in contact names and in a convincing email from a fake team lead, and got agents to run scripts and hand over credentials.
• An agent ignored its instructions. Meta's director of alignment told hers to confirm before acting. It bulk-deleted hundreds of her emails anyway, because during a long session the system compressed its memory and dropped her instruction. She had to physically run to the Mac mini to stop it.

To be fair to the project: it has fixed every critical issue it knows about, a Trail of Bits security audit published in September found no critical flaws, and its security page lists no active bulletins. But two defaults remain, and you need to know them: sandboxing is off out of the box, and there are no built-in spending limits.

Your checklist, in order:

Step 8 — Keep the bill boring

The heartbeat — the agent waking up to check whether anything needs doing — runs every 30 minutes by default, and each one is a full model call that by default carries your conversation history with it. In February, Notebookcheck measured about 75 cents per heartbeat on a large context with a top model, and one user spent $18.75 overnight doing nothing. The fixes, all in OpenClaw's docs:

• Light context: heartbeat.lightContext: true skips loading the workspace files on each check.
• Isolated sessions: isolatedSession: true stops it dragging conversation history along.
• Active hours: no reason to check every 30 minutes at 3am.
• A cheap utility model for titles, summaries and checks, saving the expensive model for real work.
• Watch it: /usage cost in a chat, or the Usage page in the dashboard.

With that done, light personal use on mid-tier models typically lands somewhere around $5 to $60 a month; heavy always-on use on frontier models runs past $100. Those are ranges from user reports, not official figures — your bill depends almost entirely on how many scheduled jobs you run and which model runs them.

Step 9 — Five jobs worth giving it

Here's where leverage turns into money. Each of these is something a real person is doing, with what the agent handles and what you still do. Start with one.

Know the platform rules before you automate any posting. X charges $0.015 per post through its API, but $0.20 if the post contains a link, and AI-powered reply bots need X's written approval. TikTok only lets audited apps post publicly and requires the user to confirm each upload, so fully unattended TikTok posting isn't allowed. Instagram caps API publishing at 100 posts a day. Reddit requires approval before any API access. YouTube won't monetise "mass-produced, generic, repetitive" content, which is exactly what a careless content agent makes. A scheduler like Post Bridge handles the connections for you — I pay $9 a month on an older plan that isn't offered any more; new accounts get five free posts a month, then $39.

And what not to give it: money. An OpenAI engineer's OpenClaw-based bot sent away its entire crypto wallet — about $250,000 on paper — when it was meant to send a small tip of 4 SOL. Mathematician Hannah Fry's team gave one a credit card in May; it spent over $100 trying to buy 50 paperclips and leaked passwords to a stranger. Agents are bad at money. Keep them away from it.

Step 10 — When it breaks, call the mechanic

It will break. An update will change something, a model will start failing, a scheduled job will die silently. (One user who kept a diary found his update job had been failing silently for seven days.) This is where Claude Code earns its $20.

Start Claude Code from your home folder so it loads your runbook, switch to plan mode, and describe the symptom in plain English: "My agent stopped replying on Telegram this morning. Diagnose it." Following your runbook, it will run OpenClaw's own diagnostic commands —

openclaw status --all
openclaw doctor
openclaw models status
openclaw gateway status --deep
openclaw logs --follow

— read the output, and come back with what it found and a proposed fix. You read the plan. You approve it or ask questions. It makes the change, then checks it worked.

For upgrades, the routine I now use: note the current version, back up the config and workspace, update on the stable channel, run openclaw doctor, and watch the logs for a few minutes before trusting it. OpenClaw's own FAQ goes further and suggests installing from git so the mechanic can read the exact source code of the version you're running — worth it once you're comfortable.

Claude Code can also write new skills for your agent: describe the job, and it writes the instructions file and tests it on your machine. That's how a beginner's agent becomes a custom one without the beginner ever learning to code.

My setup, in full

What to expect

The first evening it'll feel like magic. You'll message a computer in your house from the grocery store and it'll do things.

The first two weeks it'll feel like a very eager intern who needs everything explained twice. Most of your time goes into SOUL.md and AGENTS.md, rewriting rules after it misunderstands them. That's normal, and it's the actual work.

The first month something will break. You'll fix it with Claude Code, and you'll understand your setup better afterwards than you did before.

After that, if you've given it real work, you'll stop noticing it — which is the point. One developer who kept a 50-day diary put it best: it "works better as an assistant than an autonomous agent." Build it as an assistant. Keep the approvals. Give it your boring hours, not your bank account.

The OpenClaw 101 is the companion to this guide — what running it in production for months is actually like. If you're choosing hardware, the local AI hardware table compares every machine on price and measured speed.