Telling AI agents exactly where to make changes in your codebase is tedious. Jens Neuse, CEO of WunderGraph, just open-sourced Voxcode to fix that bottleneck. It's a macOS tool that pairs local speech-to-text with code search. Select code in your editor, press the Right Command key, speak your instructions, and Voxcode pastes the transcript with a precise file and line number reference at your cursor. It works with any IDE (VS Code, JetBrains, Zed, Neovim, whatever) and any AI coding tools that accepts text input, from Claude Code to Cursor to Copilot. Voxcode feeds all repo roots into a single parallel file walker using the same SIMD-accelerated string matching that ripgrep uses internally. Neuse found that spawning ripgrep once per repository took 9 seconds across 200 repos. This approach cut search time to under 1 second. Transcription happens locally using NVIDIA's Parakeet TDT model via ONNX Runtime, so no audio leaves your machine. The whole thing is built with Rust and Tauri. Neuse built Voxcode because he kept hitting that same bottleneck. He tried Superwhisper for speech-to-text but it lacked code context integration, and he didn't want to pay a subscription for a tool running on an open-source model. Voxcode is deliberately dumb. It doesn't integrate with any IDE or coding tool directly. It just reads what you selected, finds where it lives on disk, and pastes the result wherever your cursor is. Simple, fast, local.