What have you learned to be the best practices? The stuff I'm trying to build is just not working, despite my AI telling me that it's built and working properly all the time. Am I stupid or is my AI agent stupid?
Also, I've been using clawi.ai to host my agent in the cloud, but that's becoming rather expensive. So, any advice on hardware/ a link to a how-to guide on how to set up hosting for myself is appreciated as well.
Which model are you using? Most folks raving about claw are using the latest frontier models.
You mostly just need to make sure you have enough RAM. 16GB or greater should be fine. Nearly any old computer should do - delete everything from it then start from there.
Setup is still fairly technical and it's best to have a local friend help you get setup - someone who can sit next to you and pair up.
Be very careful what you give it access to - either on your local network or your accounts in various places. I recommend not giving it access to anything except segregated/parallel versions of things.
Are most people you know running frontier models on the local machine? Or are they plugging into an online API service?
Plugging into model providers. Running large frontier models locally requires a machine with a few GPUs and 1TB of RAM afaik.
At least in the office, everyone is using Codex-5.3 with a bit of Opus and Sonnet 4.6.
That's what I thought but you mentioned RAM so I wasn't sure. Coz if you're just running text based feeds to and from the model provider I didn't think hardware would matter much. But maybe there's a lot of additional orchestration that happens locally
OpenClaw is pretty bloated. If you're having it run web browsers and stuff, memory can get tight.
That's what you would take the Mac Mini M3 Ultra w/ 512GB RAM [1] for, or 4x M4 Pro with 128GB in a cluster, see #1360715 for the latter, which is perhaps the better setup (because you can add mac minis to it)
You'll run quantized GLM-5 (or Kimi K2.5 on a cluster of 8). Then you run your agent on a much lower spec box.
I'm still looking for a clone of openclaw that I can actually compile - maybe nullclaw, because with less
sloplocthe chance of it being unable to compile is lower 😂 Going to be "fun" diving into Zig tho, ugh.🥺 I remember when my new computer (in the late 80s iirc) had 512kB RAM and that was a beast. ↩
thanks dude! I think I'm gonna wait 1-2 more months before I dig into this again...once it's a bit less technical and scary! lol
I haven't done anything cutting edge openclaw, but I did try BrowserOS.
I'm using a video transcoding service provider (Mux) that can get really expensive so I wanted to keep tabs on my daily usage. They don't have a billing API so I was able to setup BrowserOS with a daily scheduled workflow that logs into the website and downloads the billing CSV file and summarizes the contents. The next step would be to send the summary to myself, maybe by setting up the browser agent with it's own e-mail account, or I could set up a webhook on my backend server.
Not having paid anything, I ran into API ratelimits. Looked into Claude pro, got sticker shock and noped so hard. I think I am going to hold off on agents until I'm able self-host.
My experience - API to openAI codex - super smart agent, no skill - no problem, he can create his own skills. Created a separate calendar for me and was planning my meetings, was filling in pdf documents for me and suggested to do company taxes. API costed me approx 10 usd per day, so I switched to openAI gpt5-mini and suddenly my agent became dumb. Cannot use nostr, says he cannot install any dependencies and I have to run cli commands myself. But cost is like 1 USD per day. For now I put it on hold as I need to understand if I want to pay 10usd per day or not. Probably you can optimise it and do small tasks via cheaper models, haven’t really digged into it. My feeling is - in the next 2 month we will have openclaw-like agent from one of the companies with 20-30 USD monthly flat rate, so the problem will be solved.