Goddamn terminal, permissions and all the stuff I know too less about to use OpenClaw
How I finally got the webbrowser option to run and ended up being used by the machine
Lately my posting here has focused on work-related tasks. But I found on this very weekend a task that seemed quite AI-ready. In a club, someone was playing a record that I really liked, and I only saw that the record’s label was green and red. So, I thought, this should be a relatively easy task for an agentic AI.
Hey, can you do me a favor and please search on Discogs for records with a red and green label, electronica genre, and from the last 10 years.
Hey, can you do me a favor and please search on Discogs for records with a red and green label, electronica genre, and from the last 10 years.
I thought, this should be manageable. But of course, this was not going anywhere while I was at the club. Instead, OpenClaw explained to me that it needed more permissions, and so I was pushed back to give this task a try today. Most importantly, I needed to give the server with OpenClaw on it the possibility to use a browser.
What I did not know is that OpenClaw itself is so focused on the implementation on a MacMini that it recommends not-working workflows and permissions for the case of being run on a VPS instance. The important bit is actually, to really install Google Chrome itself and not Google Chromium, as explained in the linked Reddit post.
After that and lengthy troubleshooting, I finally got OpenClaw the capability to use a browser and also the Discogs API to get the images of records as planned. Also, I got it to build a script that checks how much green and red is contained in the cover art.
The problem: for the example year 2024, it did not find matching examples. And when I wanted it to look for matches in the last 10 years, the script would keep on crashing. The result was that OpenClaw was happily giving me instructions to paste one line of code after another into the terminal to get the process done. But whenever I asked it, to plan and execute the commands by itself, it would trip.
This is an all-to-familiar outcome, like the time, when I tried to get OpenClaw to summarize audiofiles autonomously. So in the end, the fastest route for me was to switch back to ask OpenClaw time and again, what I should do next in the terminal. And in the end, we actually got a first shot of results.
Unfortunately, the record I was looking for could not be found. But at least I got somewhere. However, I would count this not only a failure because I could not find the sought-after record. But more importantly, at least in the context of OpenClaw I still massively struggle with, letting agents do things autonomously without me being needed to use the terminal.
Comments 0
No comments yet — be the first.