@janikvonrotz I still need to try out one of these... Currently I'm too cozy with Claude code.
Are you using a CLI tool, or just chat?
@janikvonrotz I still need to try out one of these... Currently I'm too cozy with Claude code.
Are you using a CLI tool, or just chat?
@ligasser I am using the Pi.dev harness. I recently did a training on the work-with-pi-workflow: https://wiki.mint-system.ch/input-prompt-md-und-pi/
I highly recommend to just not use anything by Anthropic or OpenAI (and BigTech). It is so much more interesting.
To give you an example: The basic Pi installation gives you the absolute basics. For harness features like todo lists or subagents you install extensions. One of the extenions I like is https://pi.dev/packages/pi-context-view?name=Context+usage It deciphers the entire context windows / usage
@ligasser You can also use Claude with Pi. One little step to take back control of the ai stack.
@janikvonrotz I decided some time ago that I don't want to do that :) My goal is not to understand LLMs, that's not where I want to learn stuff. I want to use them to do my work:
- fixing security issues
- explore and write code
- discuss ideas
- grammar check on my emails...
So far the Claude code is the one CLI where I trust it the most not to do stupid stuff in my terminal, and I really like the workflow. Now it can do things it failed to do 2 years ago, so I don't want to go back...
@ligasser Ok, that is an interesting attitude. And by interesting I mean questionable.
I believe the least thing you can do is not using Claude but everything else. And as I understand you have the privilege to access so many alternatives. So why not use them?
@janikvonrotz Does "I tried, but couldn't get it to work the way I want, and don't want to fiddle around with LLMs, but do other stuff, like fiddling around with coopcloud.tech recipes" count?
@ligasser @janikvonrotz I think you can plug in EPFL RCP models directly in claude code. I personally use OpenCode, and it's decent. I tried Pi, permission system wasn't built in, and you need a plug-in, so I have up quickly.
@yildiz @ligasser I did use OpenCode, but did not like:
- cli (run without interaction)
- access session info (token count, model)
- configurability (system prompt, extensions)
- assumes a lot (todos, mcp, plan/build modes)
The permission system of opencode is not a real safeguard. You should rather run the thing in sandbox that on the computer.
What are the EPFL RCP models?
@yildiz @ligasser Ah this: https://www.epfl.ch/research/facilities/rcp/ai-inference-as-a-service/
140 models! Very cool.
Do you have access to DeepSeek-V4-Flash?
@janikvonrotz @yildiz Yes, we have DeepSeekV4Flash - we link it to our C4DT web chat...
@janikvonrotz @ligasser oh yeah I wouldn't dare unleash it on my work machine or prod environment. I run in rootless podman,with only access to code I'm working on, modifying container image depending on use case.