Build an OpenClaw agent

OpenClaw is the runtime people mean when they say "a personal AI agent." Not a chatbot in a web page and not a coding assistant in an editor — a process that stays running, holds context, reaches you on the chat apps you already use, and does things on a schedule whether or not you are looking.

It is also the heaviest of the small runtimes, and the one most likely to be overkill. This guide is about knowing which of those two you are dealing with.

What makes OpenClaw different

Most agent runtimes are a loop around a model with some tools bolted on. OpenClaw is a gateway: a long-lived process that owns a set of chat channels, a scheduler, a skills system and a JSON-RPC API, and routes work between them. That architecture is why it can do three things the lighter runtimes cannot.

It is genuinely multi-channel

The same agent, with the same memory, answers on every channel in the fact card above. This matters more than it sounds. An agent that is only on Telegram is a Telegram bot; an agent that is on Telegram and Slack and IRC and can be reached over its API is a service. You tell it something in one place and it knows it in the others.

It acts without being asked

Cron jobs are first-class. An OpenClaw agent can wake at 07:00, go and check the things you told it to check, and message you the result — no request, no trigger, nothing for you to remember. Most of the value people get out of a personal agent turns out to be this rather than conversation, and it is the single biggest reason to pick OpenClaw over a lighter runtime.

It fans work out to sub-agents

A large task gets decomposed and handed to sub-agents rather than crammed into one enormous context window. Practically: it stays coherent on jobs that would otherwise fall apart halfway through, and the cost per task is legible because each sub-agent's work is a separate, attributable unit.

Building one

1. Describe the job, not the runtime. "Answer my Telegram overnight and send a digest at 8" gets you a better agent than "set up OpenClaw," because the brief becomes the agent's actual instructions. Mention the channels you want and anything that should happen on a schedule.
2. Choose how the model is paid for. Bring your own API key and you pay your provider directly; use the managed model and tokens are forwarded at provider cost with no markup. Either way the model bill and the hosting bill stay separate — see what an agent actually costs.
3. Deploy. One click puts it on its own Firecracker micro-VM with a private network and a persistent volume. Nothing is shared with another tenant, including the kernel.
4. Connect Telegram first. It is the least painful channel to set up — a bot token from BotFather — and the fastest way to find out whether the agent's instructions are right before you wire up anything harder.
5. Add the schedule last. Get the agent answering well in conversation before you let it act on a timer. A cron job on a badly briefed agent is just an alarm clock that sends you nonsense.

Writing the brief so the agent is actually useful

The brief is not a prompt you throw away — it becomes the standing instructions the agent works from on every turn and every scheduled run. Three things separate a brief that produces a useful agent from one that produces a chatty one:

• Say what it should do without asking. "Escalate anything from my landlord immediately" is an instruction. "Be helpful" is not.
• Say what it must never do. An always-on agent with chat access and a scheduler benefits enormously from an explicit list of things that require you to say yes first.
• Say what "done" looks like for the recurring jobs. A digest with no shape becomes a wall of text within a week.

Where it goes wrong

Cron plus a vague brief burns money quietly. A scheduled job that runs every hour and does slightly too much is the classic always-on failure: nothing breaks, nothing alerts, and the model bill is four times what you expected at the end of the month. Start with a daily schedule, look at what it actually cost, then tighten.

It is heavier than the job often needs. If you want one bot on one channel that answers questions, OpenClaw is more machine than the job deserves — PicoClaw does that on a smaller, cheaper box. If your job is really a multi-step business process with hand-offs and approvals, Hermes models it properly instead of you rebuilding a workflow engine inside a personal assistant.

Multi-channel is a security surface, not just a feature. Every channel you connect is another way to reach an agent that can act. Add them one at a time and for a reason.

Watching it once it is running

An agent that acts on its own schedule is exactly the kind you cannot usefully supervise by reading a chat log. Every agent built here reports to ClawMetry, so you get cost per conversation, what it did and when, and a way to pause or stop it mid-run. For a runtime whose whole point is doing things while you are not watching, that is not an optional extra.

Is OpenClaw the right pick?

Not sure? Describe the job rather than the runtime — the builder's solutions engineer reads the same capability matrix this page does and will argue for a different one if it fits better.