Build Your Own AI Agent, or Get an Expert to Do It?

Short answer: if you enjoy tinkering, have a technical background and the agent will only touch low-stakes data, build it yourself. The one-click options are genuinely good now. If the agent will read your business email, talk to your customers or touch your CRM, get an expert to build it, secure it and hand you the keys. The build cost is small compared to the cost of getting security wrong, and your time is worth more than the saving.

That is the honest version. Here is the detail behind it.

The one-click promise, and what it actually gives you

You have probably seen the ads. Hostinger will sell you a “1-click OpenClaw” that is “live in 60 seconds” with “no technical experience needed”. They claim over 100,000 agents deployed. Nous Research’s Hermes Agent installs with a single command and comes with 40 plus tools out of the box. These are real products and they are genuinely impressive. An agent you can message from WhatsApp that reads web pages, writes files and remembers what you told it last week, for a few dollars a month in server costs, is something that did not exist two years ago.

Quick naming note, because the search results are confusing. OpenClaw is the open source personal agent that was previously called Clawdbot and then Moltbot. Hermes Agent is a competing open source framework from Nous Research, and it is a different thing from the Hermes 4 language models, which are just models. If you have been researching “openclaw vs hermes”, those are the two projects people are comparing.

Now the part the landing pages skip. Hostinger actually sells two different things:

• Managed OpenClaw is the genuinely easy path. Hostinger handles updates, security and backups, and you buy AI credits through their dashboard. If you just want to poke at an agent for personal use, this is a fair product.
• OpenClaw on a VPS gives you a server with the software pre-installed and root access. The one click installs the software. Everything after that is yours: API keys and model billing, pairing WhatsApp or Telegram, updates, SSL, backups, and the configuration files. That is not “no technical experience needed”. That is running a server.

The gap between those two sentences is where most people come unstuck. The install really does take a minute. Running the thing safely is the actual job, and it never stops being the job.

What went wrong for 40,000 people

This is not a hypothetical risk. In February 2026, researchers at SecurityScorecard scanned the internet and found more than 40,000 OpenClaw instances exposed to the public. Not hacked. Just sitting there, reachable by anyone, because the default configuration bound the control panel to the open internet rather than to the owner’s machine. Many of those instances were leaking the exact things an agent needs to do its work: API keys, email and messaging tokens, SSH keys, and in some cases the whole filesystem.

It got worse from there:

• A security firm called OX Security found that OpenClaw stored credentials and API keys in plain text files, and that deleted secrets lived on in backup copies.
• A one-click remote code execution flaw (CVE-2026-25253, rated 8.8 out of 10) let an attacker take over an instance via a single malicious link. A patch shipped quickly, but scans showed most running installs were on old versions. Unpatched servers do not patch themselves.
• Researchers audited ClawHub, the community marketplace where you download extra skills for your agent, and found 341 malicious ones. Most came from a single campaign that installed password-stealing malware on Macs.

And then there is prompt injection, which deserves its own paragraph because it is the risk most owners have never heard of. Your agent reads things: emails, web pages, documents, calendar invites. Any of those can contain hidden instructions. Researchers have shown that a single crafted email can convince an agent to quietly send its owner’s private keys and tokens to an attacker. The agent is not being hacked in the traditional sense. It is doing what it always does, reading text and acting on it. The OpenClaw project’s own security documentation is refreshingly blunt about this: “Prompt injection is not solved by system prompt guardrails alone.”

None of this means OpenClaw is a bad project. Its creator has been upfront that it began as a hobby and a tech preview, and the documentation now tells you exactly how to lock an instance down. The point is simpler: the marketing says one click, and the project’s own docs describe a hardening checklist. Both are telling the truth about different halves of the product.

The car analogy

Most of us could change a tyre, and plenty of us could change the brake pads with a YouTube video and a Saturday. Almost nobody does their own brakes, because the cost of getting brakes slightly wrong is not proportional to the money saved.

An AI agent connected to your business email and your customer records is the brakes, not the stereo. If a hobby agent you run for fun leaks its API key, you cancel the key and feel silly. If an agent wired into your business leaks the wrong thing, you are explaining to customers where their details went, and in Australia that can come with obligations under the Privacy Act.

So the build or buy question is not really “can you do it”. Plenty of business owners could, eventually. The questions that matter are:

1. What does the agent touch? Public information and your own notes, or customer data, money and your reputation?
2. Who notices when it misbehaves? Automations rarely fail loudly. They quietly skip invoices and drop leads while the logs say everything is fine. Someone has to be watching.
3. What is your time worth? The weekends you spend learning Docker, environment variables and token budgets are weekends you are not spending on the business. That is the real price of DIY, and it is usually the biggest line item.

An honest comparison

The part nobody mentions: the expert learns your business

Here is the quiet advantage of having someone build it with you rather than buying a box. While Andrew is connecting your inbox to your CRM, he is inside your systems, seeing how the work actually flows. That is usually where the second and third ideas come from. The agent build you asked for might save your admin five hours a week. The thing he notices while building it, the quoting bottleneck or the report someone assembles by hand every Monday, is often worth more than the original job. A one-click install will never tell you that. It does not know your business. A good consultant finishes the build knowing it well.

When you genuinely should DIY

We would rather be straight with you than win every job:

• You are technical, or you want to become technical, and you will enjoy the learning.
• The agent handles nothing sensitive. Personal research, your own notes, hobby projects.
• You are experimenting to understand what agents can do before investing. Honestly, a $10 a month managed instance is a great way to build intuition.
• Your need is a single simple trigger. One form feeding one spreadsheet does not need a consultant.

If that is you, go for it, and a few safety basics regardless: never expose the control panel to the internet, keep the software updated, be very careful which community skills you install, give the agent the minimum access it needs, and assume anything it reads could be trying to manipulate it.

When to get an expert

• The agent will read or send email on your behalf, or touch your CRM, calendar, files or accounts.
• Customer data is involved at all.
• You need it to keep working when you are not thinking about it.
• You would rather spend your headspace running the company than administering a server.

That last one is the real reason most of our clients hire us, and it has nothing to do with ability. They could learn this. They have decided their attention is the scarcest thing in the business, and they are right.

What it costs

For context, Australian agencies typically quote $5,000 to $25,000 for a small business AI build, and consultants charge $150 to $450 an hour. We build a custom agent with two working connections, for example email plus your CRM, for $2,500 ex GST, fixed and agreed in writing before we start. You own everything, including the credentials and the documentation. We wrote a full breakdown of Australian AI automation pricing in our cost guide.

If you want to talk it through first, that is what the free chat is for. No pitch, just straight answers about whether your job is a DIY job or an expert job. We will tell you honestly, either way.

Ask Andy a question or call 0477 048 850.

Common questions

Is OpenClaw safe to run for my business?

Out of the box, no. Researchers found more than 40,000 instances exposed on the public internet in early 2026, many leaking API keys and tokens, and the project’s own documentation says prompt injection cannot be solved by prompt guardrails alone. It can be run safely, but that requires deliberate hardening: private network access only, prompt injection precautions, careful skill vetting and prompt patching. That work is the product you are paying an expert for.

What is prompt injection?

Hidden instructions inside content your agent reads, like an email or a web page, that trick it into doing something its owner never asked for, such as forwarding private data to an attacker. It matters because agents act on what they read, and an agent with access to your email and files has a lot to give away.

How much does it cost to run an AI agent yourself?

Hosting is cheap, roughly $5 to $30 a month. Model usage is the variable cost and depends entirely on how much the agent works; light personal use is often under $50 a month. The real cost of DIY is your time: expect 20 to 60 hours to get something reliable, plus ongoing maintenance.

How much does a professionally built AI agent cost in Australia?

Agency quotes for small business builds typically run $5,000 to $25,000. Our fixed-price build is $2,500 ex GST including two system connections, with running costs (model usage and any tool subscriptions) paid directly by you, typically $50 to $300 a month depending on workload.

Can I start DIY and upgrade to a professional build later?

Yes, and it is often a sensible path. Experimenting with a cheap managed agent teaches you what these tools can do. When you are ready to connect it to real business systems, that is the point to get the security and reliability done properly. Nothing you learn is wasted.

Is a one-click AI agent from Hostinger actually one click?

The install is. Hostinger’s managed plan is genuinely low effort for personal use. The VPS version installs the software in one click and then leaves you running a server: API keys, updates, backups, channel pairing and security are all yours. For business use, treat “one click” as the start of the project, not the end.