My OpenClaw decided on its own to research the watch microbrand scene in Iran. 🤣
Seems like I have to work on it some more.
And yes, I named my OpenClaw Ryan. He’s my intern. IYKYK
I’ve been running OpenClaw/Moltbot/Clawdbot for about a week now…still not sure what I’m missing by not running it on a Mac.
And I mean that literally…since I run it on Linux, I have no idea. 🤣
I installed llama 3 8b on my Mac Mini. It’s not a great LLM, but tbh, it works the way I expect it too on Ollama.
It’s unusable when being used via API…will work on molding it, but I think it won’t be too bad to help preserve my tokens/rate limits for smaller tasks.
Just a call from my OpenClaw / Clawdbot/ Moltbot reminding me to fix my wife’s toilet.
I think she will be OK with me buying a $9400 Mac Studio to make it even more powerful now. 🙃
It took a LOT of work, but I’ve modded the heck out of Clawdbolt/Moltbot to where’s it’s kind of useful and not token eating monster.
I’ll continue to optimize and add more functionality to it.
Got tired of the token burn running Clawdbot.
Ordered a Mac Mini M4 Pro with 24GB so I can run an LLM locally and pass the heavy lifting to a cloud based.
We’ll see how it goes. 🤷♂️
Good question, here are my thoughts:
Moltbot / Clawdbot is a glimpse into the future.
It does have one major constraint: cost.
Even with heavy optimization, I keep running into 429 rate-limit errors using hosted APIs. Others have pushed this far and still land around ~$12 per day, which is hard to justify for a personal assistant that still needs frequent troubleshooting.
I’ve considered going mostly local.
In theory, a Mac Studio with an M3 Ultra and 96 GB of unified memory running Llama 3 70B, with Claude Max as a fallback, could work technically. But at roughly $9,400 upfront plus $200 per month, the ROI is tough to rationalize.
So yeah, it's a cool toy to play around with, and you kind of enjoy fixing things that break, but as far as it being something that profoundly changes your life, nope, not there yet, unless if you're spending a lot of money.
Yesterday I ran into my fair share of 429 errors yesterday with clawdbot, and now it’s 503 errors.
The 429 errors weren’t so much the LLM’s fault, and the 503 less so, but it does show how hobby-level this stuff is.
Spent a good amount of time the last 2 days trying to integrate certain features into it carefully, and it’s still not where I’d like it to be.
If you’re on the fence of about setting up, have realistic expectations about how much of a time commitment this is (a lot).
If you give clawdbot your credit card number, and it buys a bunch of luxury watches on its own because it saw your Notion, and your partner finds out, can you call the credit card company and have them marked as authorized charges?
Not that this happened to me.