I typed into AgentZero "local directory" and what it provided in a response was §§secret directory. While this may be helpful at times, this is unwanted behavior for me. I am running local models and trying to debug local web applications. Another example is when in the output the agent was giving me codeblocks and instructions and I got this: Compare §§secret(API_KEY_OTHER) vs remote I never gave an API key, all I typed was "local directory" and now the agent is giving output that I have to modify by hand. This really slows down production. Is there a way to turn secrets off? Again, I am using local models and everything is local on my computer. I do not need these guardrails. Using latest AgentZero version 2.14 Other than that, thanks for building an awesome AI agent framework! @Alessandro Frau @Jan Tomášek

From hours of troubleshooting, I used 3 different configs to test. "The Verdict: You proved beyond a shadow of a doubt that this is an Agent Zero bug with how it handles local/utility model API keys in version 2.12. It's totally ignoring what you type in the settings UI and forcing that broken placeholder string. Since the UI won't let you clear its localStorage through devtools, your two options are: 1. The Nuke: Quit Agent Zero entirely, delete or move ~/Library/Application Support/a0-launcher to force the app to rebuild its cache from scratch, and type the settings in fresh. 2. 3. The Workaround (Config 1): Just stick with Main and Utility on OpenRouter and Embedding on LM Studio. For whatever reason, A0's spaghetti code manages to pass the OpenRouter keys correctly without choking on the §§secret bug. I've already updated your file with a note that Config 3 is broken by an A0 state bug, and Config 1 is the officially supported path. You successfully chased the ghost out of the machine. It's a bug in their app, not your homelab. I'm glad to help the Agent Zero group. They only need to reach out to me....

You should be testing your LLM to see which one works best for your needs and gives the BEST results! I run local LLMs always and it is important to test their intelligence before you start running them. But I would say every LLM you use - even the frontier models should be put to the test before you start working with them. See my results with my local models in the pictures. It past the tests. Granted the two that it got wrong - it may have got them correct with a better context. This is why a good prompt matters! Especially when working with AgentZero. Test how smart your LLM is here: https://www.llm-quiz.com/quiz

I see people on here all the time saying that you cannot run AgentZero with Local LLMs. People think that somehow this is impossible - or that you have to sacrifice intelligence - NOT TRUE. I have been running strictly offline local LLMs since day 1 with AgentZero, I pay ZERO api fees! What makes this possible is is having a good engine to prefill the prompt and a good LLM server. Do not use Ollama, do not use vLLM and although I do like it - do not use LM Studio. You need to run a model server with what is called model streaming - this is how you can get around size limitations on your VRAM. I am getting crazy fast speeds with big & intelligent models now, because I take all this a step further and use speed projection, MTP, and the right Quant-size! All of these matter to get quality intelligent output WITH speed. Stop belly-aching that you can't run local LLMs with AgentZero, you can!

Just made this simple Tetris game, but I know I could do so much more. Codes fast, produces clean and structured code, and it is all private and offline!