OpenClaw was the first project to make always-on agents genuinely popular. I ran real server operations on it, and then switched that machine off on 2026-07-13. This page covers what it is and what the shutdown taught — things I stopped also stay on this site. That is the house rule.
In this pageshape and capability, my usage and shutdown, three lessons. Not in this pagethat machine's address, what it carried, any config or schedule.
| Shape | A personal agent running on your own machine, not a cloud service |
| Entry point | The messaging apps you already use — WhatsApp, Telegram, Slack, Signal and others |
| Capability | Shell commands, browser automation, email, calendar, file operations |
| Self-start | A heartbeat scheduler wakes it at an interval — it does not wait for you to speak |
| Extensions | A large library of preconfigured skills |
| Models | Model-agnostic: bring your own key, or run local models entirely |
Two design choices went on to be widely copied: ① the entry point lives in a messaging app (no new interface to open) and ② heartbeat wakeups (it can work while you are not talking to it).
Its scale is an outlier: repo openclaw/openclaw, TypeScript,
386,998 stars / 81,287 forks (GitHub API, checked 2026-08-21),
created 2025-11-24 — under ten months.
I ran a real block of server operations on it. Not a demo — work with real consequences that had to keep running. It did the job, and I am not going to qualify that.
Then: on 2026-07-10 that work moved to a different toolchain; on 07-13 the machine was shut down and the knowledge base archived in full.
Not "it was bad." Inventing a technical flaw to sound balanced would be lying.
The real reason is mundane: the work moved, so the machine had no reason to exist. And an always-on machine with no work is not free — it still costs money, still waits to be maintained, and still occupies a slot in your head labelled "how is that box doing".
House rule: if the log did not record a reason, write "no reason recorded". This one was recorded, so here it is.
It is the attention that "it exists" consumes. Every long-running process opens a background thread in your head: is it alive, when did it last update, does that error need looking at.
That cost never appears on a bill, and it is real. Which is why the three tests on page 10 belong before you build, not as willpower afterwards.
At shutdown I archived the whole knowledge base, purely out of caution at the time. It turned out to be the right move — not only as "in case I want to look back", but because those assets remain usable in the next tool.
Concretely: its successor can lift the entire ~/.openclaw directory in one command —
persona, memory, skills, settings. Delete it and that path is gone.
See page 14.
Ask it before you build: the day I do not want this, how do I take it apart cleanly?
If the answer is "no idea, it works fine right now", that system will eventually become something you dare neither touch nor stop. Genuinely removable systems share three traits: state concentrated in a few directories · nothing scattered across the system · nothing else falls over when it stops.
"It does not wait for you to speak" is the headline feature and the source of the bill. Every heartbeat can be a full model call, and cost is superlinear in turns.
Day one: set the heartbeat interval to the granularity you actually need, and know where to read usage. The default is a demo setting, not a long-run setting.
Anyone who can message you can message it. And it can run shell commands.
Those two sentences are the entire risk. Which is why the first lesson for this class of agent is always lock down who can talk to it before discussing features. Page 15 is about nothing else.
openclaw/openclaw, TypeScript, 386,998 stars / 81,287 forks,
created 2025-11-24, still being pushed to on 2026-08-21. GitHub API, checked that day.
GitHub reports its license field as Other (non-standardised) —
the licence is whatever the repository says; this page does not classify it.