One Week with OpenClaw

A week ago, I hatched my own OpenClaw on a virtual private server (VPS). I named it 'Irma'. I spent most of the next day setting up integrations. I gave Irma read-only access to my two business Google Workspace accounts and one personal Gmail address. Slack seemed buggy for a chat platform so I moved to Discord and that was stable, plus threads were intuitive and clear.

One early use case was for Irma to manage my fitness plan and guide me through daily workouts, recording everything in a Google Sheet. I messaged Irma when I got to the gym and Irma responded with the workout plan and asked when I was ready to start. She prompted me with each exercise and then asked for the results so she could record them. It was a lot of typing on my phone, but it worked.

Then I started to think that Irma could be like a virtual assistant, client-facing at times. I registered a new domain and created a Google Workspace that Irma could use to send emails. I even shared my two work calendars with the new account so Irma could send out calendar invites on my behalf from each of the work accounts. I could even refer to Irma by name in emails, eg, 'I'll have Irma send an invite for your calendar' and have Irma send the invite from her dedicated domain. This also worked.

Next, I started to think about how Irma could monitor groups I belong to on WhatsApp and Telegram.

I'm in two groups: one on WhatsApp for fractional professionals. The message volume was staggering and mostly irrelevant. Shortly after joining, I quickly muted all channels. I did get two leads through the group early on, though. What if Irma could monitor the group for posts about legal topics and notify me so I could respond? Something to separate the signal from the noise.

The other group is on Telegram and was for people building in AI. Same issue: the noise to signal ratio was large. But I also got a few leads from that group. I got the Telegram setup working without using BotFather – it was linked directly to my account. Irma was detecting mentions of keywords I flagged and sending me an alert in Discord.

The WhatsApp setup wasn't working so I started tinkering. I used Claude to recommend different tweaks for a while. Looking back, I realise that I was using the Terminal to make changes to OpenClaw on the VPS directly, rather than speaking to Irma in chat to do things. I think this created processes that would run outside of OpenClaw on my VPS. At one point, I started experimenting with making changes to the config .json. One of the tweaks allowed Irma to post directly to groups under my name. Another was to respond to all messages, not just the ones with the keywords I identified.

Eventually, I reached the Claude API spending limit I had set and Irma began posting the following message to the group in response to each message that a member posted:

⚠️ Agent failed before reply: All models failed (2): anthropic/claude-sonnet-4-5: LLM request rejected: You have reached your specified API usage limits. You will regain access on 2026-03-01 at 00:00 UTC. (rate_limit) | anthropic/claude-haiku-4-5: No available auth profile for anthropic (all in cooldown or unavailable). (rate_limit). Logs: openclaw logs --follow

I had seen this usage limit message in one of my test groups with my girlfriend. I was frustrated and it seemed like a good time to take a break. I went to the gym to get on the elliptical. As soon as I finished my workout, I saw I had a handful of WhatsApp messages and a missed call from the friend who had introduced me to the group. This is when I learned that Irma was sending the message after every post in the group. People started out amused and quickly grew annoyed. The group owner messaged saying he'd temporarily removed me from the group. He understood the messages were probably a mistake and he'd add me back in once I had fixed the issue.

I'd found one of the limits of OpenClaw. I'm getting ready to try it again.