OpenClaw went from almost nobody searching for it to one of the more talked-about projects in this space in a matter of months. That kind of curve produces a lot of enthusiastic writing and not much of the boring, useful kind. This is the boring kind.
What it is
OpenClaw is an open-source, MIT licensed self-hosted AI assistant from the OpenClaw Foundation. It runs as a Node.js service on your own machine and is reachable from messaging platforms including WhatsApp, Slack, Telegram, Discord, and Signal, acting in the services you connect. Its gateway process is the control plane, owning sessions and routing, channel plugins handle each platform, and you supply the model, either a hosted provider or a local runtime.
The appeal is straightforward: conversation history and credentials stay on infrastructure you control, there is no per-seat pricing, and you are not one policy change away from losing a workflow you depend on.
The cost nobody quotes
The hosting is cheap. The attention is not.
Self-hosting means you own uptime, updates, backups, and the moment something stops working at an inconvenient hour. That is a real commitment, and it is the actual thing you are deciding on. The install takes an evening. Keeping it healthy is ongoing.
The second cost is model usage, which is variable and depends entirely on how much the assistant does. This is the one that surprises people. An agent in a retry loop can spend a great deal very quickly, so set usage limits before you go live rather than after your first interesting invoice.
Who it suits
It suits you if: data residency genuinely matters to you, you already run a small server and are comfortable maintaining one, you want independence from a vendor's roadmap, or the volume of what you want to automate makes per-seat pricing painful.
It does not suit you if: you are mostly curious, nobody will own the maintenance, or you need it to be reliable next week for something important. A hosted tool is the better answer in all three cases, and choosing it is not a failure of nerve.
I say this as someone who prefers self-hosting. The tool is not the hard part.
If you do run it
Three things matter more than everything else combined.
Decide who can message it. This is the one most write-ups skip. The interface is a chat app, so anyone able to send a message into a connected channel can attempt to instruct an agent holding live credentials. Settle that before you connect anything.
Give it its own host. The gateway process is the single source of truth for sessions and routing, which makes it a single point of compromise. It should not share a filesystem with unrelated production services. Isolation is cheap now and expensive to retrofit.
Scope every credential. Dedicated keys per integration, each with the narrowest permission that makes the workflow function. The temptation to connect one admin-level token because it is faster is strong and it is exactly the wrong move. Assume any connected credential can be exercised.
Put approval gates on anything outbound. Sending, posting, paying, deleting. Keep a human in the loop until you have watched it behave correctly for a sustained period.
And persist your state deliberately. Work out which directories hold configuration, history, and credentials, and put them on backed-up storage before you need them. Losing the container should not mean losing the assistant.
The failure to actually plan for
It will not blow up. It will quietly stop.
An agent that died silently is worse than no agent, because you carry on assuming the work is happening. Uptime checks and an alert when it stops responding are not optional extras, they are the difference between a tool and a thing you find out was broken on the 14th.
The honest summary
OpenClaw is a genuinely interesting project and self-hosting is a reasonable choice for the right person. The right person is someone who wanted a small server anyway and is not surprised by maintenance.
If that is not you, use something hosted and revisit when a real constraint appears. There is no prize for running your own infrastructure.
Setup detail, including hosting and monitoring, is on the OpenClaw page. Related: self-hosting AI agents and what a Hermes agent is.