Entries for July 16, 2026

I deleted my agent accounts. I don't need them anymore for secure access to GitHub and Hugging Face repos Instead, I am using brokerkit, a credential broker swiss army knife which can build an approval gate around any system My agents automatically get read access to all my repos, unless I specify certain ones to be excluded They then have to open PRs which only I can merge. Unless I allowlist them on certain repos to push to the main branch Best feature: Timed requests. e.g. "I want to be able to push to the main branch next 2 hours", "I want to release this package only once in the next 5 minutes" Which saves me from the hassle of changing a config for some short term change, but also protects me from the woes of long term prompt-injection risk My openclaw instance has its own linux user on my workstation, and is no longer root. It uses gh-broker and hf-broker from inside its account. Can also run sudo through sudo-broker, to install stuff. Secure access for free, without having to set up a separate server It ingests quite a bit of information from the internet every day, and the remaining lethal trifecta risk is leaking of some of my private repos, and nothing else Repo: github.com/osolmaz/brokerkit

I deleted my agent accounts. I don't need them anymore for secure access to GitHub and Hugging Face Instead, I am using

Your ML agent on Hugging Face infra: ML Claw 🦞🤗 Quickstart: npx mlclaw bootstrap If you have Hugging Face PRO, you can run your agent on a Space, with the highest possible download speed to Hugging Face models, datasets and buckets, for fast experimentation 🚀

When I started building this, the docker space was actually free as well. But due to some bad actors exploiting the free tier at a mindblowing scale, free docker space had to be retracted from the free tier ☹️ However, you can still run ML Claw locally, but with your state backed up to a Hugging Face bucket for free, up to 100 GB of free private storage and 8 TB of public storage! Repo if you would like your agent to scan it before you run the command: github.com/osolmaz/mlclaw

Codex and Claude code teams should learn something from Cursor team when it comes to queueing/steering

I asked ML Claw to name itself. Of course in German/Alman Lame suggestions: Ulf, Strich, Fritz, Lex I intervened and called it GoePT Goethe rolling in his grave