Trixie
Voice-activated desktop agent for Linux, macOS, and Windows. Every capability is also a standard MCP server, so the same tools work from Claude Code, Codex, or any other MCP host.
Project Details
Client
Bismit Panda (Me)
Year
2026
Role
Developer
Status
Live beta
Ownership
Independent project
Technologies
The Challenge
A desktop voice agent has to own audio capture, wake word, transcription, an LLM tool loop, and speech synthesis, and still reach windows, files, and apps the same way on three operating systems and several Wayland compositors.
The Solution
Built a single Tauri application over a Rust workspace. cpal captures audio, with Silero VAD and a wake word in front of local Whisper or a hosted transcription endpoint, and Piper voices through sherpa-onnx for speech. An orchestrator runs Anthropic, OpenAI-compatible, or Claude Code loops over an MCP client registry. Platform differences sit behind traits, with an optional GNOME Shell extension for window access where Mutter withholds the toplevel protocols.
Highlights
- Fully local voice pipeline option, with wake word, VAD, STT, and TTS on-device
- Tool surface shipped as a standalone MCP server over stdio or HTTP
- Cross-platform window, workspace, and app control behind per-OS traits