OpenClaw: First Impressions
Since the viral launch of OpenClaw, Jenxi has been reminding me to get familiar with OpenClaw. As usual, it takes me a little while to get moving on things, but I do eventually get around to it.
I fiddled with OpenClaw on a Hetzner VPS over Chinese New Year in January, and encountered the usual issues that people did, which is expected.
However once I managed to get a version of it up and running with some key functions, I could see the possibilities better. While I currently still don't have a wide-enough idea of what the possibilities are, I'm spending the next few days to try and get up to speed as quickly as I can. A short break, which should have major impact in the longer term.
I went down the rabbit hole of exploring other agents (and other agent-like setups) too, such as NanoClaw, Ductor, NanoBot, ZeptoClaw, but concluded that while each has its own benefits, being familiar with what is currently the most mainstream technology would be a better bet, not to mention with all the momentum and demand being directed at it, OpenClaw will be the better choice to serve the mainstream for now.
What has helped with my adoption of it is basic familiarity with Linux and the common unix commands and reading how OpenClaw (or any other recent Claw agent) is structued, especially the SOUL.md and HEARTBEAT.md files.
In terms of getting it to a working version quickly, these are what I felt are the key tweaks that are needed
- Set it up on its own VPS (or computer) that you don't mind having to reinstall everything later.
- Giving it the maximum permissions possible. In recent updates, OpenClaw has shipped with more restrictions on the agent out of the box, which is the correct decision from a Product point of view, but it makes it harder to discover its capabilities. At least during the duration of learning, you sould lock it down afterwards.
- Hook it up to a capable model. While I spent a bit of time wrangling between Codex 5.4, GLM 5.0, Ollama, and some others, it was a major breakthrough when I found a bunch of capable and free models on Open Router. I'm currently using the Hunter Alpha model as much as I can for now, while it's still free.
Most of what I want to do involves the browser, and some nice tricks I've found give the agent better browser capabilities include:
Some things I've currently got my OpenClaw working on (while I sleep via cron jobs and the heartbeat) include
- Ongoing research for market conditions and new business opportunities
- Reviewing and refining some websites that I'm currently building
- Given it ownership of the remotejobs.com.my project that I haven't had time to look into, while I focus on steering the direction of this project.
As our RubyCoded business is increasing focus on the Open Person Company (OPC) initiative that the Shenzhen government is encouraging, familiarity and adoption of OpenClaw and autonomous agents is really non-negotiable.
Skills compound, and often we have to place bets on which skills are fundamental and will compound over time. OpenClaw itself may not be the future, but having competency and familiarity with agents and orchestration/coordination of teams will be increasingly important and is worth investing in.