sodachat running in your browser

The same models the terminal and Discord frontends use, exported to ONNX and run client-side by onnxruntime-web. Nothing you type leaves the page — after the model downloads, this works with the network off.

Pick a model and press Load.

A reply is 12 sampled candidates ranked by MMI relevance — the same engine as engine.py, so it is slower than one greedy pass and much less generic.