I run Claude Code all day and I’ve stopped reading half of what scrolls past. A tool that pipes that output through a second local model and hands back English is, at minimum, an honest diagnosis. Fully local, no telemetry, no external dependencies — that’s the part I care about. Same instinct as self-hosting anything: keep the second opinion on your own hardware.
The author is upfront that it’s slow, vibe-coded, Mac-only, and that the local model only sees what Claude says — not files, not actions — so it hallucinates. A summarizer with no ground truth means you can miss a message entirely. I’d run vomit tail in a side terminal before letting it replace anything via hooks.
The story — Vomit is a Go tool by zachahn that converts Claude’s token output into English by piping it through a local LLM, running fully locally with no telemetry or external dependencies. Install via go install, configure with vomit init, then either replace Claude’s output through hooks or run non-invasively with vomit list and vomit tail. It works with Llama.app, Ollama, or anything speaking the OpenAI API. GPLv3, 91 stars. (Source)