The short answer
Self-hosting an AI agent costs less in cash and more in time: you rent a server, install the software, connect a chat app and a model, and keep it all running. A hosted service costs more in cash and takes the setup, security and maintenance off your hands.
Both routes work, and they fail in different ways. Self-hosting suits people who enjoy this kind of work or already have someone technical. Hosting suits people whose hours are better spent on their own work. Comparing only the monthly price misleads you, because the full cost includes setup hours, maintenance, and what happens when something breaks at 11pm.
What does an AI agent need to run?
Every agent needs five components, however it's delivered: a model, an engine, a server, a chat app and a way to pay. Knowing them makes every product page and tutorial easier to read.
A model. This is the part that does the thinking. You either pay per use or connect an existing subscription where the provider allows it. Models differ in quality, speed and cost; choosing an AI model covers how to pick one.
An engine. The software that turns a model into an agent, with memory, scheduled tasks, tools and the ability to write new skills. It's what separates a chat box from something that works for you over time.
Two well-known open-source engines are OpenClaw and Hermes Agent, both free under the MIT licence. OpenClaw describes itself as a self-hosted gateway that connects the chat apps you already use, Telegram included, to an AI agent running on your own machine or server. Hermes Agent, built by Nous Research, puts its emphasis on learning: it creates skills from experience and builds up a picture of you across sessions. Both support persistent memory, scheduled jobs and skills, and both work with several model providers.
A server. A computer that runs the engine around the clock. It can be your own machine, but a laptop that goes to sleep makes a poor host, so most people rent a virtual private server (VPS). Its location affects speed and where your data sits.
A chat app. The place you talk to the agent. Telegram is a popular choice for personal agents because it runs on every device and supports bots well. LINE and WhatsApp are used too, each with its own setup requirements.
Billing. Someone pays for the model usage and the server, either metered and billed to you or bundled into a subscription.
Check all five when you compare offers. A low price that covers only the server doesn't make a cheap agent.
What does self-hosting an AI agent involve?
You rent and secure a server, install an engine, connect a chat app and a model, then maintain all of it. For a first-timer, the sequence runs like this.
Choose and rent a server. Pick a provider and a size. Too small and the agent is slow or crashes; too large and you're paying for capacity you don't use. Most people misjudge this once.
Secure it. Create a user, turn off password login, set up SSH keys, configure a firewall and apply updates. If that sentence means nothing to you, this is where self-hosting turns from an afternoon into a project. Both engines publish security guidance worth reading first: OpenClaw's and Hermes Agent's.
Install the engine. OpenClaw and Hermes Agent both offer a one-line installer and a guided setup, and it usually works. When it doesn't, you're reading error messages and searching for others who hit the same one.
Connect the chat app. Create a bot, get a token, paste it into a configuration file, restart and test. Easy once you've done it, fiddly the first time.
Connect a model. Either an API key with billing attached, or a subscription sign-in where the provider allows it. Read the terms, because they differ by provider and they change. Anthropic, for example, doesn't permit third-party apps to route requests through consumer Claude plans, and no provider lets you resell access.
Set up backups and monitoring. Decide what happens if the server dies, and how you'd find out. Most people skip this until the first failure.
Keep it running. Updates, security patches, disk space, expired tokens, model changes and the occasional breaking change in the engine. It isn't daily work, and it never stops entirely.
Plenty of people do all of this happily. It's a skill, though, and that skill has to live somewhere: in your head, or with someone you pay.
What does a hosted agent take off your hands?
A hosting provider handles the server, the security hardening, the engine installation, the model connection and the billing. Some services hand you a sign-up form and a settings page; others set the agent up with you. With Able°, we do the setup on your own private server (in Singapore by default) and guide you through the few steps on your side, such as connecting Telegram, which Connect Telegram walks through.
Before you trust any provider, check:
- Shared or private server? A private server per customer is a different product from a shared pool with separate accounts.
- Where is it? Location affects speed and where your data is stored; what happens to your data explains why that matters.
- How is model usage billed? Metered credits, a bundled allowance or your own subscription each behave differently when you use the agent heavily.
- What happens when it breaks? Is there support, and how quickly does it respond?
- Can you export your data? If you leave, do you get your history and the agent's notes?
- What can the agent do today? Ask for the current capability list rather than the roadmap.
That last check prevents the most disappointment. Features a provider plans to ship are a promise, and you can only use what exists now.
Is self-hosting an AI agent cheaper?
In cash, usually. Once you count your hours, often it isn't. Prices are easy to compare and hours aren't, so hours get ignored. Work through these:
How many hours will setup take you? For someone technical, an afternoon. For someone who has never opened a terminal, a weekend plus a few evenings, plus the time to recover from a mistake that breaks something.
What's an hour of your time worth? If you bill by the hour, you have the number. If you don't, use what you'd pay someone else to do the job.
What does downtime cost you? If the agent runs your morning planning and reminders, an outage is an inconvenience for a day. If you lean on it for deadlines, it costs more.
Which subscription would you avoid? Some people self-host partly to skip a monthly fee, which is reasonable. Compare that fee against your hours rather than against zero.
Add it up and the cheap option sometimes isn't. Sometimes it still is, and that's fine.
A later cost catches people out: keeping it working. Setting up once is a project. Getting through an engine update, a model change and a provider policy change is a series of smaller projects, each turning up at an awkward moment.
Which route suits you?
Self-hosting tends to win when:
- you're technical, or have someone who is, so setup is enjoyable and maintenance fits existing habits
- you want full control over engine, model, tools and data location, without a provider's decisions in between
- you have unusual requirements, such as a specific integration, compliance arrangement or workflow no product supports
- you want to learn, because building one teaches you how agents work in a way that using one doesn't
- budget is tighter than time, which makes doing it yourself a rational choice
Hosting tends to win when:
- your time is the scarce resource and setup isn't the work only you can do
- you need it working this week, since provisioning, hardening and testing take longer than people expect
- you want someone accountable, with a support channel instead of a search engine when it breaks
- you're still working out what you want, and would rather try the idea before committing to infrastructure
- a team will use it, and shared access, roles and an activity log aren't things you want to build
Can you switch later?
Yes, though both directions cost something that rarely appears in comparisons.
Moving from self-hosted to hosted means carrying your data across, learning a new interface and possibly losing some configuration: a day of work and a spell of reduced reliability. Moving from hosted to self-hosted is usually harder, because by then there are months of memory and history to export.
Before you sign up anywhere, ask how you'd export your conversation history and the agent's memory. A clear answer means the switching cost is manageable; a vague one means you should assume it's high. On the self-hosted route you already hold the machine, so the answer is usually better.
Many people start hosted, learn what they use the agent for, and only then decide whether self-hosting is worth it. Others self-host because they enjoy it and move to hosted when maintenance becomes a chore. Self-hosting something you don't yet understand goes worst, because every problem looks like a configuration error when the cause is often the design.
Questions to ask either way
Whether you build or buy, these separate a working agent from an expensive experiment:
- What can it do without me asking?
- What does it remember, and where is that stored?
- What can it access, and with what permissions?
- What happens when it's unsure?
- Can I see a log of what it did?
- What happens if the provider or the server goes down?
- How do I leave, and what do I take with me?
If a provider can't answer these in plain language, price is the least of your worries. For the safety side, read how AI agent safety works.
When you need neither
If you send a handful of AI requests a week, a full agent may be more machinery than you need. A chat tool plus a few saved prompts can do the job.
If you can't name three tasks you'd hand over, wait. The setup cost only pays off when there's regular work for the agent to take on.
Common questions
How much does it cost to self-host an AI agent?
There are two running costs: the server rental and the model usage. The engine can be free if you pick an open-source one. The server is a predictable monthly fee, while model usage grows with how much the agent does, so a busy agent can cost more in model usage than in hosting, before you count your own setup and maintenance time.
Can I run an AI agent on my own computer?
Yes, both OpenClaw and Hermes Agent run on a laptop or desktop. The agent only works while the computer is on and awake, though, so scheduled check-ins and reminders stop when it sleeps. That's why most people who rely on an agent all day move it to a rented server.
What is a VPS?
A VPS, or virtual private server, is a rented computer in a data centre that runs all the time. You get your own slice of a larger machine, with its own operating system and storage, and you decide what's installed on it. For an AI agent it provides the always-on home a laptop can't, and you're responsible for looking after it.
Are OpenClaw and Hermes Agent free?
The software is. Both are open source under the MIT licence, so you can download, run and modify them without a licence fee. Running them still costs money: the server they live on (unless you use your own computer) and the AI model usage, through an API key or a subscription sign-in where the provider allows it.