Openclaw has been extremely popular this month, although it looks like a vibe-coding project done by one person with AI in 2 months, and is really just a wrapper of different LLMs that wants access to everything. But I wanted to give it a try nevertheless.
Given my very good experience with Gemini, I have tried to use gemini-2.0-flash as the model API for openclaw, but noticed there is no free allowance in Europe and UK. Unfortunately Gemini does offer free allowance, at the expense of customer’s data being used for training, but due to the privacy laws in Europe, (thank you so much AI regulator!) they can’t offer it here.
Then I tried to host Qwen3:8B locally using Ollama, and set up Openclaw to use it. Talking directly to Qwen on Ollama was fine, but it took Openclaw 30 seconds to respond to a single ‘Hi’ from me. I am guessing Openclaw is giving the model too much context to process, and my Mac is not good enough.
Finally I tried to switch to OpenRouter, which offers some free APIs. (Side note: why are there so many middlewares/wrappers? Afaik there is Groq, LiteLLM, OpenRouter, Langchain, Ollama, vLLM, and probably more.) gpt-oss-120b seems to be the best free API that OpenRouter offers. Optimisitcally, I gave Openclaw my Whatsapp, and gave it my first task: Send me an encouraging statement every 30 minutes. Unfortunately, Openclaw failed to do so due to some cron job error. And the thing with Openclaw is whenever it fails, it keeps infinitely re-trying, and it easily uses up the daily API allowance within minutes.
The other problem I noticed is it would actually reply to every incoming whatsapp message assuming someone is trying to give it a command (what a stupid default setting!). How awkward. The fix is to change the dmpolicy setting from ‘pairing’ to ‘allowlist’. I appreciate the fact that Gemini knows everything about Openclaw without the need to google, even though Openclaw is really just a month-old project.
