Vomit: a local LLM that rewrites Claude Code's output into plainer English

Vomit is a small Go command-line tool that intercepts the text Claude Code emits and reroutes it through a locally-run language model, which paraphrases the output into what its author considers more readable English. The project runs entirely on the user’s machine with no telemetry and no external dependencies, hooking into Claude’s output stream (via a ‘scrub’ hook mode) or watching sessions passively through a ‘tail’ command. Setup involves installing the binary from GOPATH, pointing it at a local model, and — if the user has none — running something like Llama.app with GPT-OSS 20B behind it.

The author is upfront that this is a rough, self-described ‘vibe-coded’ experiment tested only on Mac. Because the translating model sees only Claude’s outbound messages and has no visibility into the underlying files or tool actions, it fills gaps by hallucinating, and it can occasionally drop Claude’s real message entirely. It’s also slow. For recovering the untouched original text, the author points users to a separate tool, AgentsView.

The project is less a serious productivity utility than a pointed piece of commentary on how verbose and jargon-heavy agentic LLM output has become — enough that someone built a translation layer for it. It’s released under GPLv3.