i recently set up my own openclaw instance, and it was a wild ride.
actually setting it up
the actual installation is quite easy with like a single curl or package to install. after that comes the fun part.
the model choice
when i first installed it, i went with openrouter/free as the model since i didn't want to pay for any tokens. this worked for like a day or two when i hadn't really used it much. once i started to try to get it to even read my email, i gave up. i thought my memory setup was the issue. no, it was just the really bad models. at that time kimi k2.5 was still the newest since k2.6 wasn't out, so that was my choice. also, since it was quite cheap compared to other models like the claude or gpt lineup, this actually helped a lot.
refrain from using it
then i had it set up with a good model and working, but i had little use case and also i was paying for every token so every dm to it cost me cents to euros. that's why i rarely used it.
the codex subscription
a while ago i decided to bite the bullet and purchase the 20$ openai subscription. i connected this to my opencode too, and it was an actual gamechanger. i just set the model to gpt 5.4-mini which didn't take much from my usage and it was crazy. it's a super fast model that doesn't cost me anything directly and is not that much worse than the kimi or flagship models from other companies.
🦞