the hackable AI agent framework for people who'd rather make human slop
I've been hearing a lot about OpenClaw lately. OpenClaw can read your email, OpenClaw can control your smart home, OpenClaw earned me $67 trillion on Polymarket, get in on the $OPENCLAW crypto drop before it's too late, OpenClaw is AGI, OpenClaw is 1000 security vulnerabilities wrapped up into one neat npm package, etc, etc, etc. All the Twitter tech bros seemed to be JUMPING on this little weekend project made by some random dude in, of course, San Francisco. It's open source. Actively developed. And who knows, I think there are some free inference providers out there. Maybe it could finish my ELA assignment for me. Wait, I still need to do my ELA assignment.
I'm extremely internally conflicted with generative AI, specifically LLMs.
On one hand, they're an incredible technology that is extremely technically impressive behind the scenes and has mind-boggling potential for humanity. OpenAI and Anthropic could launch us into a utopia, where everyone earns a generous UBI, where there are zero limits to creating things anymore, where humanity can shape and mold the world to our whims. In the shorter term, AI makes things that were absolutely unfathomable years ago normal (case in point: in a recent ARG I ran, I seriously considered the possibility of rolling my own full IdP platform for one single puzzle), makes learning anything easy and simple, eliminates tedious and boring busywork, and globally distributes knowledge and possibility in a way that could never have happened before.
Yet it also has a ton of downsides. People stop thinking, instead relying on black boxes created by for-profit corporations. We get addicted to "You're absolutely right!". Jobs are lost. Peoples' work is stolen and shoved into this training process without permission or care. We lose the purity, the determinism, of software (one of the reasons why I love it so much) and trade it for temperature and random predictions of the next word. Humanity is launched into a fever nightmare of grotesque AI slop, purple-to-blue gradients, environmental problems, and AI psychosis. We build datacenters, huge ones, to stay ahead of the competition - to appease the AI. We launch not humans, not animals, not experiments, but GPUs into space. Maybe the bubble bursts, and the working class is plunged into a recession while Sam Altman and his buddies (enemies?) are laughing all the way to the bank. Or maybe the bubble doesn't burst, and a misaligned ASI rapidly improves and kills us all.
I don't know. AI is cool, I guess. I think the best analogy I can find for generative AI is Temu; yes, it's so cheap, so easy, and you're able to find exactly what you need. Yes, it makes building things so much faster, more affordable, and simpler. But peel back the curtain and you'll stumble upon layers upon layers of ethics concerns, environmental problems, and terrifying truths. And your package is gonna take a long time to arrive, and it might work for a little bit, and then it'll either break, or end up killing you.
To be clear: I'm talking about LLMs specifically here, not image or audio generation. I hate that stuff.
So, well, while the AI bubble is still inflating, I might as well give it a spin. It would be funny - and useful - to have "Claude with tools and cron" in my discord sidebar. And I can burn some of that investor money by using the free models. Maybe I can hook it up to Google Classroom to do my ELA assignment. So, on my Latitude E6400 that I've been using to run a Mattermost and Gitea instance, I run the suspiciously simple setup script. I wait for Homebrew to install. Then pnpm. Then all the dependencies. I arrow-key through the menus. Paste my Groq key. Paste my Discord bot key. And then I start using OpenClaw.
And oh my god, it's horrible.
The CLI is messy. The tools are janky. The web panel is half-baked. The OpenRouter support is incomplete. The docs hinder more than they help. In the troubleshooting section, the first step is literally "clone the repo and ask Claude Code or Codex or whatever". WTF??? I tried to make it work. But who thought, in the age of Go, Zig, Rust, hell, even Python, that Node.js would be the best choice for this?? The logs are hard to read. It's incredibly slow. There is zero - absolutely zero - explanation for many things. There are probably 67 zero-days sitting in the repo as I write this. The Android app is mentioned multiple times in the docs, yet there's no download for it. The github repo has 4.4k open issues (at time of writing), many of them generated by agents themselves. It kills my soul to see that.
OpenClaw is not "the future" - it is a vibecoded mess that's being marketed as a product.
Yet I'm still incredibly intrigued by the original promise of OpenClaw (and, by extension, Poke, which did all this already months before OpenClaw yet didn't really make a splash). What would happen if you gave an LLM access to a wide range of tools and had it loop over and over? Would it be any useful? Would it at least be funny? Can one middle schooler with his Claude free tier account and too much time on his hands make a better product than OpenClaw?
So I'm building Bramble - a simple, lightweight, polished, documented harness for your favourite LLM that connects it to Discord, the internet, and a wide variety of tools. I'm not trying to sell anything, or "change the world", or get bought up by an AI lab, just trying to experiment and do weird stuff with these funni next word predictors. And use up my OpenRouter credits.
And get someone to do my ELA assignment for me.
~ hex4
Woo! You made it through the dramatic expository bit! Here, I'll let you in on a secret for reading this far: I have absolutely no clue how to use Go. There will be much more AI assistance than I'd usually admit to in this project. But hey, at least I understand the code. I don't think the OpenClaw guys even know there is code.
wait, what's "security"?
Caution
jokes aside, giving an LLM access to API keys, dangerous tools, and other sensitive information, and then letting it run free on the interwebs is not a good idea. assume that the LLM will either do something stupid, or someone else will make your LLM do something stupid. i've tried my best to build in good-ish security practices, but I'm just a middle schooler with a free-tier Claude account and less-than-average amounts of homework. please be extremely careful when using Bramble (as well as OpenClaw and variants!!!), and don't give your agent access to anything that would cook you if publicly released. and maybe don't let random people talk to your agent. and maybe don't hyper-publicize your agent's identity.
one more thing: while I've tried my best to make setup and operations as simple and straightforward as possible, bramble is a technical project. if you are not well-versed with computing, sysadmin, and the CLI, please avoid giving Bramble access to possibly-dangerous tools, and DO NOT paste commands or prompts from untrusted or sketchy sources.