This is the browser version of the helper lane already running in the Kindle app. It keeps local session memory, lets you fan one prompt across multiple models, and exports thread bundles, SFT rows, and feedback rows so stronger-model conversations can become training material for your smaller SCBE assistant.
The value here is not generic chat. It is that the same governed conversation can become a portable bundle for later ingestion into the SCBE training pipeline.
Keep the conversation concrete: one customer issue, one bookkeeping lane, one inventory problem, or one compliance review pass.
Use the compare-model field in Settings to capture better replies from larger systems in the exact same thread instead of starting over elsewhere.
Export thread preserves provenance, Export SFT gives quick training rows, and Export feedback records what helped versus what needs work.
The shell below is the same reusable chat engine used in the Kindle app lane. On this public page it starts in safe mode: no hidden token, no hidden proxy, and no claims that a backend already exists.
If you run this through a real server-side proxy later, keep the Hugging Face token off the page and store only the proxy endpoint in browser settings.