March 2026: OpenClaw Setup and honeymoon period
In March 2026 I set up a local OpenClaw running on a Mac Mini I procured with that purpose in mind. Mid-March I had the OpenClaw, nicknamed “Nova” write about its setup. End of March, I wrote an exuberant review of my workflow and walking through a usecase, making an web application on the go with voice commands over Slack.
April 2026: OpenClaw model migration, big spend, and debugging
Early April, disaster struck! Right before a planned family trip to San Francisco, Anthropic announced they’d no longer support harnesses piggybacking on Claude Max subscriptions. What that meant in practice is that the brain of my OpenClaw was about to go offline, so I needed to retool it in a hurry. My first priority was that it stay up through the transition as fixing it long term would be easiest if I could use its own infrastructure to help. So we spun up a channel and got going.
Migrating from Claude Max to a hierarchy of models
We setup OpenRouter as well as a local model fallback.
Then I got word that Anthropic would be gifting $200 in API credits to those effected, so I decided to use those as the primary model until they ran out. It turns out that I burned through them in just a few days! I was really benefitting from the subsidized tokens in my workflow. I then switched to OpenRouter, and had a lot of conversations about using less costly models (e.g. not Opus 4.6) and ended up switching a lot of my regularly recurring jobs to those, as well as some individual Slack Channels.
I spun up a #usage channel to track usage, and 8 days into April I had already spent $200 in Claude Credits PLUS nearly $50 in OpenRouter. I wasn’t going to slow my workflows down though, so I needed a new solution which I resolved to find when I got back from my SF trip.
Migrating to OpenAI’s Codex – and the slow decay in functionality
When I got back, from SF I saw that OpenAI supported running agents on their Pro plan, so I signed up to the tune of $100/month and migrated the OpenClaw entirely over, with Openrouter or a local model as fallbacks or being able to be called upon as needed. Whew! After that my usage was under control, and did not creep up more. $100/month seemed reasonable and fit in my budget. At first things seemed to be going smoothly, mostly because I wasn’t spinning up projects for a few days. But then as I started to spin up more projects I noticed the model just being lackluster or, dare I say, lazy. Often the model would ask me to do things, instead of it doing them. Despite prompting otherwise to “be agentic” and to just “do the thing I asked you and see it through, kindly”, and to “put in your memory to not check in unless absolutely needed”, the problems continued.
I put it on OpenAI’s model being inferior at first, and gradually started using the OpenClaw less and less mid-April. Many of the previously functioning tools degraded, it seemed to have trouble opening the browser when commanded from Slack, and the purchasing workflows were fine if they ran on cron jobs (recurring jobs) but not as one-offs requested from Slack
May 2026: Back to the old modest budget and high capabilities
The root of the issue solved! The wrong level of thinking effort
Toward the beginning of May I sat down on console for the first time in the month, thinking that it was time to service the OpenClaw. First we debugged the browser issue which amounted to me saying “try harder” a bunch on console until it was fixed (some configuration issue mumble mumble) and then as part of that somehow I saw that Anthropic’s model had a level of thinking default to it and thought to look into Codex’s default. It turned out, it was defaulting to LOW. We quickly upped that to the max, and then were off to the races, with the Claw seeming as smart as they were in the old days. With browsing fixed, and the brain’s of the operation back online, I was back in hacking action.
A retrospective of use cases
Here are a bunch of things I have cooking:
It’s a mix of sysadmin for the OpenClaw, science projects, coding projects, writing projects, and family maintenance activities. There are a bunch of everyday activities setup in the background that don’t have their own dedicated Slack channels. Here are some things I have going
Family Admin
- ordering groceries via Ralph’s API + Instantcart – can be prompted with recipes, meal planning, other ideas e.g. “Here’s a few recipes I want to cook this week for a family of 6, can you buy the groceries”
- any ordering on Amazon via a prompt – e.g. “I’m doing this project here’s a list of materials, go buy them” or “buy X”
- ticket buying: I like to go to the botanical gardens and rollerskating. I have memberships. It autoreserves tickets every few weeks and adds them to my calendar
- haircuts: I like to make sure my hair is cut every 12-15 weeks and dyed ever 6 weeks or so. It books the appointment and adds to my calendar
- doctors and dentists: it searches the read only calendar it has access to and highlights if a family member doesn’t have an expected next appointment (dentists every 6 months; doctor’s every year)
- it connects with invite provider Partiful and creates and invitation from a prompt “Hey can you create a board game brunch invite for this Saturday 11am”
- family trips and other such: I can spin up a channel, do research, pin items, have the agent draft itineraries and more. kind of like a shared Claude chat session.
Vibe coding for family / health
- on the machine I’ve made and it runs notifications when my husband’s car is due home (for dinner) and when my baby wakes up from sleep or has a bottle (while I’m working or traveling only – these notifications can be paused)
- those are powered by another vibe coded integration between the Snoo smart bassinet and the Huckleberry sleep/tracking app. Huckleberry is great for logging bottles, pumping, and auto estimates wake windows from the logged sleep
- I have an iPhone app that takes my heart data from my watch and tracks whether I’m getting enough moderate/vigorous exercise per week.
Writing and Media
I have a lot of writing projects and media projects going. Some things I’m working on are
- A middle grade novel – featuring mirror bacteria at the South Pole!
- lots of blogs – I often spin up a slack channel to workshop one and do background research befor publishing
- reports and other books. I’m thinking of writing something on Engineering leadership and it currently interviews me by asking questions I then respond to, to build up content
- baby or kids books I’m writing or workshopping
- A podcast on productivity and parenting with weekly interviews for 2026 season
The Slack channels help me stay focused with research and spin off various tasks. For example, we’ve created some great illustrations for the middlegrade novel based on my own South Pole photos.
Productivity
- I have several slack channels that are more about daily routines and productivity, a Slack channel that reminds me about good diet practices with daily goals, a daily goals channel, and channels that surface interesting papers as well as language learning content
Science
A big part time effort of mine is scientific work. Some things I’m looking into are mechanistic interpretability, pentesting and analyzing the security of the OpenClaw system itself, and building an electroporator for my home biohacking lab [with the intent of eventually working with cyanobacteria].
June 2026 and beyond: Migrating to Hermes
My OpenClaw is back in action after upgrading to high thought, but it still is very forgetful.
I often prompt it with things I shouldn’t really need to as it’s in its main memory file. So I’m now thinking of migrating to Hermes; they have an OpenClaw migration workflow. I have a great sense of what I use this for, and it is ever expanding. Memory does seem to be a limitation and I’d love to try out something that worked a little more smoothly. So my goal for June is to give Hermes a spin. What are you building? Send me a note!
You must be logged in to post a comment.