mrshu

While the project is very interesting in and of itself, the process section that has been edited a bit later to the post is even more revaling:

> A little bit more on my process. I use Claude Code (desktop app for Mac). I use Opus 5.5 on Medium effort as my default. I don't actually do anything resembling traditional engineering anymore (read/write code, run commands in a terminal, or etc.), I only talk to Claude and have it do everything. In this case, the only "manual work" I did was: 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box and 3) I bought the domain name. I worked on this in between Claude's actual work so it still feels a little bit like team work :) Once the box was up I gave Claude the IP and told it update, deploy, secure, harden and optimize the code, and it did the whole thing.

It really does sound like you do not need to run models like Opus 5.5 on high reasoning effort for most tasks.

I guess this unfortunately means HuggingFace won't be "the first company to go public with an emoji instead of the three-letter ticker" as the cofounders originally intended:

  "When we started the company, a running joke with my co-founders was that we wanted to be the first company to go public with an emoji instead of the three-letter ticker when you go in the NASDAQ."
  
  -- Clem Delangue (https://unsupervisedlearning.substack.com/p/the-future-of-open-vs-closed-source)

The historical uptime has been getting better but it is on a downwards trajectory in August:

https://mrshu.github.io/github-statuses/?view=all

This is why the "reduce thinking" approach of BottleCapAI is both quite interesting and very useful in practice:

https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B

I built a multi-agent alternative with Stream Deck and cmux/orca and it works pretty well -- particularly when you pair it with CodexBar and can easily see your token spend across multiple subscriptions.

It is all open source https://github.com/mrshu/muxboard -- feedback would be greatly appreciated!

By far the most impactful product of the Apretus project are the people. To quote a memorable line from Dominique Paul (https://www.thisiscrispin.com/):

> What most people miss IMO is that this is not a team who is doing this for the fourth time like virtually any other LLM provider and who could learn from its own past experiences. I bet if the team would do another model training they could get way better results at one fourth of the costs.

You are right, I am not sure why I did not realize Python is the whole point of the image. This is indeed much faster and easier.

An old habit that unfortunately makes one indistinguishable from LLMs these days...

It's pretty rough to learn I sound like Claude. Will need to do something about it then.

(For what it's worth I did write the message above manually but I understand why no one would believe that now. At least I did not call netcat "load-bearing" [https://mareksuppa.com/til/load-bearing/] or something...)

Yes, a very minimal image indeed. Had it been a custom image, curl would be one of the first things I would make sure it contains :)

> No, it can not. Bash lets you open TCP sockets.

Very fair pushback -- I did get carried away and will update the article to be more precise. Thanks for raising it!

> For less insane, non-bash shells there is always nc which is usually probably the wiser choice.

For completeness, `nc` or any netcat equvialent I could think of was not available in the image I was trying this with. It would certainly be a better option though.

Totally!

I was really just trying to see if intra-container connectivity works, and this ended up being a very quick way of doing so. (The alternative being building and deploying a new image, which would likely take significantly longer.)

That is indeed a solid pushback! :)

For what its worth, this container used `python:3.12.2-slim-bookworm` and I really would not expect that sort of an image to bundle `curl` -- even if it is intended for production.

I ran into this while checking connectivity between containers on an internal Docker network where the image had neither curl nor wget.

The main surprise was that Bash has /dev/tcp which lets you do the equivalent of an HTTP request with a bit of shell magic, for instance:

  exec 3<>/dev/tcp/service/8642
  printf 'GET /health HTTP/1.1\r\nHost: service\r\nConnection: close\r\n\r\n' >&3
  cat <&3

Where `service` is just the hostname of whatever you’re talking to and 8642 is the port you are trying to talk HTTP to.

Pretty cool!

I have also had positive experience with doing this multiple times via multiple model families, and then to recursively have the fixes reviewed too.

It's called review-anvil and does find significant amount of problems that might pop up:

https://github.com/mrshu/agent-skills/#review-anvil

Their uptime for just Git operations has been hit pretty hard in the past 90 days:

https://mrshu.github.io/github-statuses/

Well, some of the original commits are indeed quite telling!

https://github.com/amantus-ai/vibetunnel/commits/main/?after...

Author here. I wrote this after setting up Claude Code with MiniMax and Z.AI and realizing their docs all tell you to paste API keys into settings.json in plaintext -- which is risky given that Claude Code has been known to read .env files and leak contents into session transcripts. I already use KeePassXC, so I wrote a shell wrapper that fetches the key at invocation time and passes it as an inline env var. Nothing is written to disk. The same pattern works with any password manager CLI -- op read for 1Password, bw get password for Bitwarden, pass show for pass. Happy to answer questions.

It really seems the primary benefit of MCP servers was to force companies to establish some externally exposed APIs if/when they did not have them.

To be fair, it's not like he did not read a single line of code that ended up being generated.

This (multiple major outages a day) has unfortunately been happening for quite a while now -- on the 2nd of February, 2026 for instance.

The GitHub Status Page does not visualize these very well but you can see them parsed out and aggregated here:

https://mrshu.github.io/github-statuses/

Here it is. It looks like they are down to a single 9 at this point across all services:

https://mrshu.github.io/github-statuses/

GitHub no longer publishes aggregate numbers so here they are parsed out. It looks like they are down to a single 9 at this point across all services:

https://mrshu.github.io/github-statuses/

Some projects (like the pi coding agent) use a gated approach for first-time contributors:

https://github.com/badlogic/pi-mono/blob/main/CONTRIBUTING.m...

Here is what it looks like in practice:

https://github.com/badlogic/pi-mono/issues/1218

Peter Steinberger, the author of Clawdbot / Moltbot

https://steipete.me/

It was not built by Claude -- Peter no longer uses it for coding -- he builds exclusively with Codex now: https://steipete.me/posts/2025/shipping-at-inference-speed

Here is an example: https://www.dailymail.co.uk/news/article-15476463/Trump-tell...