vishnukool

I built basically this on top of my telegram setup last year. The thing that killed it was group chats, per-message LLM calls on a busy group eat your api budget alive, so I ended up putting a tiny local model in front as a first pass.

Such a rage bait title, but sure I'll take the bait.

It is easy to phrase anything as slaves. Are refrigerators and microwave ovens our slaves? But if you push the tech fare enough.. there is a boundary where we do cross the slave threshold.

The problem is that LLMs do not have a conceptual grounding in actual time. They estimate based on statistical correlation found in their training data which is filled with standard corporate project management timelines legacy codebases and waterfall estimates.

Good one. One thing that's becoming clear is that agent security is less about jail break prompts and more about permission boundries, hidden context flow and un intended tol behavior.

System design definitely carries more weight now but Leetcode has not totally vanished. It's just evolved.

Respect. It's a good day to remember the impact they had even if they are no longer around to pick up the phone. Btw calling her now.

This is the weird part of AI hallucinations and people start treating it like a verified fact after it gets repeated.

What have you seen to be the bottleneck ?

From the usage patterns I've seen for 90% of the use cases it seems to be fine, the one exception being Browser based use cases which chugs memory and so needs a beefier machine.

Very true and actually totally spot on!

It's the damn chromium instances that chugs memory like an Irish man chugs beer. I experimented launching containers with it and quickly realized the tiny shared infra hosts are not a good fit. We're launching a higher prized plan that has more memory for such things.

Yes exactly! Although we've launched a Claude Skill to help setup OpenClaw! So Agent 37 is literally the easiest Openclaw setup at this point :)

Absolutley! We do have some metrics like, when a user interacts with a skill or not, and so that's a signal

Yes, once we have a large enough numbers to justify that we'll factor that into the search results quality

Ah that's a good point, I think I could introduce some kind of time decay by also factoring in last updated repo time into the search priority.

Great point, we de-duplicate the search results by Repository Id. that largely filters out content from the same person , and note that search prioritizes github stars, so low quality one's get pushed to the bottom

Thank you! We have limits around usage. Also as a startup we have a lot of Gemini credits from Google, so that helps in offering it. Not to mention the gemini 3 flash that it's running on is dirt cheap.

Regarding low quality skills, our search only surfaces top starrted and forked skills, so that largely addresses that. Kinda like google page ranking. You'll see that whatever you search for, you tend to only get high quality skills first.

Yup we actually backup data in a separate partition per container. So it survives restarts and even accidental deletions.

Thanks! At the moment no, but down the line I imagine we can expose it if this is useful

We do back up data, although it's manual. The docker instance home directory and homebrew are backed up in a separate partition. So it survices restarts and container updates.

SO far pretty good! Turns out these things are mostly idle

We only allow ttyd not ssh, so users only have access to their individual containers and isolated from the host. But inside the container they can do whatever they'd want. There are also caps on CPUs and resource and PID limits so as to avoid noisy neighbor situations

True but Unfortunately, it's expensive to run VMs that are stateful and ongoing. In a way the 99 cents is just gate keeping really interested people vs others.

thanks!

Yup on the roadmap!

Haha thanks ! Yeah i really pushed the limits on this one :)

Sorry that was only for the early adopters, the current price is at 3.99 so we can have some margin.

The 4GB is just from my personal profiling and using OpenClaw

At the moment it's more of a demo for hobby users setting up for themselves and friends, but if it's an enterprise use case where you need more control, we have https://www.agent37.com/openclaw offering that works better.

Yeah the auth here is simple intentionally. This is just an example of what's possible, but you could extend the node server to do forward auth or anything else you have in mind

That would basically be as simple as updating the Docker image and entry point file and then restarting the container.

I wanted to let a few friends get a hosted OpenClaw setup without each of them dealing with Docker, reverse proxy, SSL certs, etc. So I built a small toolkit that turns one Linux box into a multi-tenant hosting setup. Each person gets their own isolated instance on a unique subdomain.

What each person gets:

1. Their own OpenClaw dashboard. 2. A password-protected web terminal (HMAC tokens, 24h TTL) 3. Isolated CPU, memory, and PID limits so one person can't take down the box

Stack: Docker containers + Traefik for routing and auto SSL (Let’s Encrypt DNS-01) + a tiny forward-auth service (~90 lines of JS) for terminal token validation.

You can try it locally with one command, no server needed.

That spins up 2 instances on localhost using localtest.me. For real server deploys, there’s a provision script that sets up Docker + Traefik and a create-instance script to add more users and give them their own OpenClaws.

Rough sizing: each instance needs ~4 GB RAM, so an 8 GB VPS fits ~2, 16 GB fits ~4, etc. If you’ve already got a VPS sitting around, you just need a domain pointed at it (wildcard DNS A record) and the provision script handles the rest: installs Docker, sets up Traefik, automates SSL. Way cheaper than managed OpenClaw hosting and you actually own the infra.

Open to hearing feedback, anyone else self-hosting this vs using managed hosting?