I tried to ignore the hype. Curiosity got the best of me, and a free weekend finished the job.
OpenClaw went viral, 200k+ GitHub stars in a matter of days, spawning dozens of alternative projects like Nanobot and Hermes. I tried different setups, learned a lot, and landed on a setup that actually handles the boring parts of my day. Now Anthropic’s Claude Cowork is shipping the same patterns, and we must acknowledge where those came from. Behind the hype were real user needs and usage patterns that mainstream products had missed. There are more exciting things Claws do that should serve as an indication of what’s to come.
So, what are the Claws?
Your ChatGPT lives in the browser (or an app), each session separate. Very little carries over between conversations through a limited memory system and files you share with it one by one.
Claude Code lives in your repo, extendable through skills, plugins and MCP. The main focus is code, but you can point it at your notes, too.
The Claws are “Claude Code” for everything else in your life, living on your machine, with your data, interacting with services you use, reachable outside of the terminal and autonomous.
Original OpenClaw, Hermes Agent for self-learning, and Nanobot for its simplicity.
My experience
I tried to ignore the hype, but curiosity got the best of me. The main triggers were Nanobot with the promise of everything OpenClaw does, but in roughly 4000 lines of Python, which meant I could understand most of it in a few hours, and a weekend free from work. Soon I had my own agent that could search the web, pretend to be a cat, and remember that I once ate pancakes. Not very useful, far from the hype, but Pippin the Cat agent was born.
The first real use-case appeared when my partner and I started looking for a new apartment: 3 weeks of back-to-back viewings every Sunday. Pippin became a way to collect and structure apartment listings, add our feedback and rank them after viewings.
Next came the recruitment tracker for my Deerdata interviews, identifying gaps and important questions to ask 10 minutes before the interview starts.
Pippin helped me with calendar management, coffee brewing optimizations, it even helped with collecting data and harassing my ISP for bad service. Soon it all needed structure, so I revived my old Obsidian vault, and together with Pippin we migrated my more recent notes to it.
The first real issue was memory corruption. It was just a couple of markdown files, and it relied on the agent to extract entities and facts from sessions and update its own memory files. This worked fine until it didn’t, and when it failed, it completely overwrote the memory with the most recent irrelevant conversation.
The next big issue came from extending the agent with tools, mostly using MCP servers. The number of tools had grown to hundreds, overfilled context with noise; The agent never picked the right tools and burned the tokens.
I soon realised that introducing more users to it also comes with a whole new class of issues in every step of the process.
The Validation of the Claws
Recent features added to Claude Code and Cowork have clearly been inspired by the likes of OpenClaw, Hermes and Nanobot:
- Working with your own files
- Persistent memory and context across projects
- Skills and extensions to interact with your browser, calendar, Jira and more
- Scheduled actions, routines and loops
- Remote dispatch sessions from your phone or messenger
Now nicely packaged for everyone to use without spending nights tinkering. Behind the hype there were real needs and use cases.
What’s next
I have rewritten this article five or so times, just ask Saide – my chief editor. Either I go too deep and try to cover everything until it’s a disconnected, incoherent mess with no message or simplify it until it’s too vague.
Besides Claws, there are exciting things happening with agents being able to evolve their own skills from use, either through “dreaming” or more active nudges, that I am sure will eventually make it to consumer products.
But back to using Claws. The issues I have mentioned for Pippin the Cat agent either have been resolved or I am actively exploring solutions. And I would like to share my journey covering topics like memory, context rot, multi-user environments, speech and sovereignty in a series of articles.
Pippin is still running. Mostly behaving. If any of this sounds familiar or intriguing, stick around for future deep dives into specific topics or reach out to compare notes.