Setup is covered on the OpenClaw page. This post is about why you would pick an assistant you run yourself, and what that asks of you in practice.
What makes OpenClaw different
With ChatGPT, Claude or Gemini the model runs at the provider. With OpenClaw you run the assistant on your own hardware or on a server you manage. That changes who is in the chain: the conversation does not leave your environment.
For organisations with strict requirements on where data ends up, that is the argument. Not because the other parties are careless, but because some obligations simply require you to show there is no third party in between.
What it asks of you
Operations. You run a model yourself, keep it current and make sure it is available. That is work the provider does for you with the other assistants.
Hardware. A model good enough for administrative work needs serious compute. Do not underestimate that when weighing the cost.
Realistic expectations. A self-hosted model is generally weaker at understanding vague questions than the large commercial models. Be more concrete in your phrasing than you would be with ChatGPT.
What stays the same about the integration
The MCP integration itself does not change. Ledger Botje offers the same tools regardless of which assistant calls them. The data cache also stays in the EU and authorisation runs through the same OAuth flow.
So what you gain by self-hosting sits on the assistant side, not the integration side. That distinction matters when you have to explain the argument to whoever decides.
Where to go next
To sharpen the trade-off between self-hosting and a hosted assistant, see the security page for what is arranged on the integration side. For the tools themselves: the tool catalogue.
Who this is and is not for
Self-hosting is rarely the cheapest or easiest route. It is the route you take when there is a reason that outweighs convenience. That could be a sector obligation, a tender requirement, or internal policy that does not allow external processing of certain data.
Without such a reason, self-hosting mainly produces extra work. A hosted assistant with a data processing agreement is then more practical and usually better at understanding your questions too.
What capacity you need
A model that reliably handles administrative questions is not a small model. Budget for a machine with a solid GPU and enough memory, plus someone to keep it running. Do not underestimate upkeep either: models get replaced, and what is good enough today may not be in a year.
Make those costs explicit in the trade-off. A hosted assistant costs a subscription per user; self-hosting costs hardware, power and operations time. Which comes out cheaper depends heavily on user count.
Start with a limited scope
If you want to try it, start read-only and with a single administration. You quickly find out whether the model understands your type of question well enough, without a misunderstanding having consequences. If it disappoints, you have that conclusion within a week rather than halfway through an implementation.