Developer Builds Local Tool to Decode Claude's 'Token Vomit'
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Zach Ahn published "Vomit," a tool that routes Claude's garbled "token vomit" through a local LLM to convert it into English, with source code released under GNU GPLv3 at github.com/zachahn/vomit.
- Vomit runs entirely locally with no telemetry and no external dependencies, though Ahn warns the intermediary LLM "hallucinates a bit" because it can only see what Claude attempts to communicate — not its actions or files.
- The project is "vibe-coded" and only tested on Mac, with Ahn flagging that it is "pretty slow" and carries a risk of completely missing Claude's message.
- Vomit does not modify Claude's runtime — it only writes files to TMPDIR — and Ahn recommends AgentsView as a way to recover the original unprocessed messages when Vomit falls short.
- Installation requires
go install github.com/zachahn/vomit@latest; Ahn suggests using Llama.app with GPT-OSS 20B as the local LLM backend for users without an existing setup. - A non-invasive mode lets Vomit run alongside Claude via commands including
vomit list(session identifiers),vomit tail(translate or follow the latest session), andvomit scrub -claudefor hook-based output replacement.
Why it matters: Developers frustrated by Claude's verbose, repetitive output get a fully local, GPLv3-licensed workaround that doesn't phone home — but Ahn himself flags the trade-off: the LLM intermediary can hallucinate, meaning users may get plausible-sounding summaries of messages that don't exist, and in worst cases Vomit misses Claude's message entirely.
Ask SkimNews
