Yesterday I spent $15 and had a fun day configuring OpenClaw to tell me I was burning money running OpenClaw. Totally worth the experiment, but if you're paying for your own tokens, make sure you don't have auto-bill turned on!
My tl;dr: A sign of the future, but this is v0.1 alpha stuff. It's an evolutionary step, your favourite chatbot with a cron job and a script that tries to remember by summarising your chats to Markdown.
I set it up with Claude and Telegram on a headless Ubuntu machine. I first tried Signal and WhatsApp, but I skipped them once I realised it would have access to all my chats, instead of being a private communication channel with the bot.
Google integration meant installing gogcli, which needed a GUI session to save credentials to the keyring. This is a known bug.
I blew out my context window for the first time when sending a voice note via Telegram. It didn't process it locally but encoded it to text and uploaded it to Claude. This is also a known bug.
My first experiment was getting it to check my Gmail and summarise new emails. By default, it sent the entire email body to Claude, so the first email newsletter with inline images blew out my context window, again, along with $2-3 worth of tokens.
It was fun getting it to check its own logs and vibe code a way to limit large emails, but I wasn't convinced that the Python script it produced actually solved the problem. I expect that obvious automations like this will be included in future assistants, and have tested code and sensible defaults[1].
It didn't end up flagging any interesting emails or events for me. The most useful automation was a little report at 5pm with how many tokens I'd spent. I spent a day building a tool to proactively tell me I'm wasting money. I didn't need to spend tokens to work that out 😆
I am not ready to spend this much on something with such limited utility. There are tons of breathless videos suggesting you can use this tool to run your life or business. Yes, it adds proactivity and some learning to the experience, but it's still backed by the same intelligence as your favourite LLM. Even though it's running on a loop and it (somewhat) builds a personality and memory via Markdown, you're still stuck with the variable day-to-day quality of LLMs.
I'm super jazzed about LLMs and agents right now, especially with coding. LLMs have significant safety and security problems to overcome, and I am sure there are going to be some serious WTF moments ahead. But as a tech toy to spend a day playing around with? Love it.
I can imagine that the established players are watching carefully and will either snap up a leading product or simply replace it themselves before long.
Have you tried giving Ralph Wiggum access to all your personal data and a pile of tokens? How did it go?
I long to be haunted once again by the ghost of determinism. ↩︎
