You
The model runs on your own GPU via WebLLM. No API key, no account, and nothing leaves your device. The first load downloads the weights (880 MB–5 GB depending on model) and is then cached by the browser. Needs WebGPU: latest Chrome or Edge, or Safari 26+. Firefox will not work.
Your key is kept in this browser's localStorage on this device and is sent only from your browser to OpenRouter. SlopBin never receives it, and you can delete it at any time with Forget key.
Three cartoon avatars sit down with you, four faces in all. Mira runs the table as a warm Table Topics master, Dev is a nervous new joiner, and Priya has an anecdote for everything. Start with a 60–90 second warmup of small talk, then run a five-minute Table Topics session with a visible countdown. Your turn comes round about a third of the time; speak out loud and the others react to you.
Everything runs client-side. There is no account, no server, and nothing to install — the portraits are drawn on canvas and the voices come from your browser's built-in speech synthesis.
You choose the brain. Run locally and WebLLM drives the three characters on your own GPU through WebGPU: no API key, no bill, and nothing leaves your device. Or bring your own OpenRouter key, which is kept in this browser and sent only from here to OpenRouter.
Yes. Choose Run locally and the three characters are driven by a WebLLM model running on your own GPU through WebGPU. No key, no account, no bill, and nothing leaves your device. The first load downloads the model weights (880 MB to 5 GB depending on the model) and the browser caches them after that. It needs WebGPU, so recent Chrome or Edge, or Safari 26+; current Firefox cannot.
Only if you pick the OpenRouter option. Then it is kept in your browser's localStorage on this device and sent from your browser straight to OpenRouter. SlopBin never receives it, because this page has no server at all, and you can delete it at any time with the Forget key button.
Not if you run it locally. With the OpenRouter option you do need a key, for the AI to answer over the network. Without one the room still loads and the avatars animate, but every turn falls back to a short canned line instead of a real answer. Free OpenRouter models work; several free tiers get rate-limited often, which is why a paid model is offered as an option too.
Press the Forget key button in the OpenRouter panel. It clears the key from localStorage immediately and the room falls back to canned lines. Closing the tab does not delete it, so use the button. If you were running locally, press Unload instead to release the model from GPU memory; the cached weights stay in the browser until you clear site data.
Yes, in Chrome, which supports the Web Speech API the room uses for recognition. Your browser asks for microphone permission when your turn starts. Four seconds of silence ends your turn, or press Done. Other browsers get the text box instead.
The tool itself is free and has no hosting cost. Running locally costs nothing at all. The OpenRouter option goes through your own account, so you pay only OpenRouter's rates, or nothing if you pick one of their free models.
Not from the page. Each avatar loads a portrait image, and a per-face calibration stored in your browser lines up the eyes and mouth for blinking and talking. Replacing the images is a matter of dropping new files in the faces folder.