I got OpenClaw running. First impressions were genuinely good — it felt like the kind of local-agent setup I’d been waiting to try properly, not just demo in a browser.
Then the honeymoon ended. It got stuck often. Things broke in ways that weren’t obvious to unwind. I’d get partway through a workflow, hit a wall, and spend more time recovering the agent than using it.
The part that really slowed me down was local models. I couldn’t install or use them reliably. Between packaging quirks, resource fights, and half-working paths, “run a model on this machine” never settled into something I trusted day to day.
Useful experiment. Not something I stuck with once the friction outweighed the upside.