Running OpenClaw with local models through Ollama
OpenClaw works with local models as well as hosted ones. When a local model makes sense, what hardware it needs, and the trade-offs against a hosted model.
By OpenClawPro. Published October 25, 2026. Updated October 25, 2026.
OpenClaw's documentation lists both hosted and local model providers, and Ollama is the most common way to run a model on your own machine. A local model keeps every message on your hardware and removes the per-use bill. In exchange, you need capable hardware, and today's local models are usually less capable than the best hosted ones for complex tasks.
When does a local model make sense?
When the data must not leave your infrastructure, when usage is high and steady, or when the tasks are simple enough for a smaller model: sorting, short summaries, routing. For long reasoning or delicate writing, a hosted model often still does better.
What hardware does it need?
It depends on the model's size. Small models run on a recent computer or a server with plenty of memory; larger ones need a graphics card with enough video memory. Check the model's requirements on its page before buying hardware, and test with your own tasks.
Is it free?
The model is free to download; running it is not. You pay for the hardware, its electricity and your time. A local model replaces a variable bill with a fixed one, which is predictable but not necessarily smaller.
Can you mix both?
Yes, and it is often the best setup: a local model for routine, private tasks and a hosted model, with a spending cap, for the hard ones. OpenClaw lets you configure the providers you want.
How we set it up
On a self-hosted installation, we can connect OpenClaw to a local model on your hardware, or to a hosted provider on your own account, or both. Your server needs at least 4 GB of memory for the agent itself, plus what the model requires. See the self-hosted offer.