Sweaterdog/J-Space-Explorer ? reverse-engineered prompt

Build me a local web app for exploring language models, where I can load a safetensors model and type a prompt, then see what the model is thinking at every layer and every token.

I want a grid that shows the logit lens and a smarter Jacobian lens, so I can click any cell and break that activation down into the verbalizable concepts behind it. Please also include a chat mode where I can talk to the model and inspect what it is poised to say before it says it, plus a way to steer, remove, or swap concepts and compare the edited output with the original.

It should work with a local model path, download the default model if needed, and run as a simple local app with a guided tour for first time users. If it helps, look up the current docs online and make sure the setup is easy to run on one machine.