aviral-1008/switch-board

Switchboard is a self-hosted web app that lets you send one question to many different AI chat models at the same time and see all their answers appear side by side. Instead of picking a single AI provider and copying the same prompt into different websites one by one, you type it once in Switchboard and it fans out to OpenAI, Anthropic, DeepSeek, Google Gemini, Groq, NVIDIA NIM, OpenRouter, Mistral, xAI, Together AI, and Cerebras all at once. The answers stream in live, appearing word by word as each model generates them, rather than making you wait for a full response before you see anything. Each model keeps its own separate conversation thread, called a line, so you can mute one model or swap in a different one partway through a conversation without losing the history of the others. Your API keys for each provider are stored only in your browser's local storage and are sent straight from your browser to that provider. The server that runs Switchboard never sees or logs your keys. There is also a voice input option using the browser's built in speech recognition, letting you speak a prompt and review it before sending. Conversations are saved automatically on your device so you can come back to them later, and any response can be copied as Markdown text with the model's name attached. The app has three main pages: a home page describing the product, the actual chat app, and an about page explaining privacy and how it works. To run it yourself, you set up a Python virtual environment, install the listed dependencies, and start the app with a single command, then open it in your browser. No separate configuration file is required since all settings live in the browser. For running in production rather than testing, the project recommends using a WSGI server such as Gunicorn along with Nginx or Caddy in front for secure connections and rate limiting. The code includes security headers on every response and checks on user input length and model names before making outgoing requests. It is released under the MIT license.