One thing that’s not in the article: using your own endpoints can be a bit bothersome, even if you have (like me) a LiteLLM proxy to abstract away all your keys. Zed has trouble setting that up (especially with authentication) and seems to be unable to set custom models to “reasoning” mode and hide or compact the <think> tags, which makes for a messy chat window.