🤮 Vomit: Refining Claude 5's Token Chaos

Vomit is a specialized utility designed to translate the "token vomit" produced by Claude 5 into coherent, readable English. It achieves this by routing the messy output through a secondary, local Large Language Model (LLM).

~~Claude 5 is hopeless~~ $\rightarrow$ Vomit makes it usable.

🛠️ System Architecture

The tool operates as a middleware layer that intercepts and cleanses data. Here is the conceptual flow of how the processing works:

⚠️ Critical Disclaimers

Before deploying this tool, please be aware of the following limitations:

[!IMPORTANT]

Hallucinations: Because the local LLM only sees the communication stream—and lacks access to the actual files or actions Claude is performing—it may occasionally invent details.

Performance: The process is relatively slow.

Stability: This project is "vibe-coded" and has been verified exclusively on macOS.

Reliability: There is a non-zero probability that some of Claude's messages may be missed entirely.

🚀 Getting Started

To get the system up and running, follow this checklist:

Install the Go environment.
Download the binary.
Configure the local LLM connection.
Set up the scrubbing hooks.

Installation & Setup

First, install the binary directly to your GOPATH:

go install github.com/zachahn/vomit@latest

Next, initialize your connection settings to link the tool with your local LLM:

vomit init

To automatically replace Claude's output using hooks, execute:

vomit scrub -claude

📖 Usage Guide

Vomit offers both invasive (hook-based) and non-invasive modes. If you prefer to run it alongside your session without modifying the runtime, use the following commands:

Command	Description
`vomit tail [session_id]`	Lists session IDs or translates tokens for a specific/latest session.
`vomit help`	Displays the full manual and available commands.
`vomit scrub`	Cleanses the output stream.

Technical Nuances

Vomit is designed to be non-invasive; it does not alter the runtime environment, though it does write temporary files to your TMPDIR. If you need to retrieve the original, uncleaned messages, it is recommended to use a tool like AgentsView.

💻 Compatibility & Recommendations

The tool is designed to work with various local LLM providers. The compatibility can be expressed as: $\text{Compatibility} = \{ \text{Llama.app, Ollama, OpenAI-API-Compatible} \}$

Recommended Stack: If you haven't configured a local model yet, the author suggests:

Installing Llama.app.
Downloading the GPT-OSS 20B model.

📄 Project Metadata

License: GNU GPLv3
Telemetry: $\text{None}$ (Fully local)
Dependencies: $\text{Zero external dependencies}$
Source: zachahn/vomit on GitHub

For more detailed insights, you can read the original blog post at https://zachahn.com/posts/1787191554.