I built Polyphon because I kept wanting a second opinion on what one AI told me, then a third, then I wanted them to respond to each other. It's a local-first Electron desktop app for macOS. You set up a "session" with any combination of voices — Anthropic Claude, OpenAI GPT, Google Gemini, local models via Ollama, or CLI tools like Claude Code and Codex. You send one message; every voice responds and reads what the others said. You can broadcast to all of them or @mention a specific one. A few things I care about that might interest HN:

No backend, no account, no telemetry. Your sessions are encrypted at rest with SQLCipher (AES-256). API keys are read from your shell environment and never leave the main process — the renderer makes zero network calls. There's an MCP server built in, so Claude Code and other agents can drive Polyphon programmatically. There's a poly CLI (@polyphon-ai/poly on npm) for terminal and CI use.

It's free and open source under Apache 2.0 as of this week.

https://polyphon.ai https://github.com/polyphon-ai/polyphon