Goulash: An LLM-Powered Copilot for Your Terminal That Respects Your Flow

If you've ever found yourself halfway through a command, realize you forgot the exact flags, and then fallen down a rabbit hole of browser tabs and chat threads, Goulash might be the tool you've been waiting for. It's a new open-source project that positions itself as an 'LLM-aware overlay' for your shell, offering suggestions and executable commands right below your prompt—without ever taking control away from you.

The core idea is simple: Goulash watches your session and uses a local LLM to offer advice, but every keystroke and command remains yours. You run goulash and use your shell exactly as before. Suggestions appear below the prompt, and you can pull them up with the Down arrow—kind of like a 'future history.' Type # followed by a question at the prompt, and the LLM suggests a command. Press Down to drop it onto your prompt for editing, then Enter to run it as your own.

Under the hood, Goulash integrates with zsh and bash using a ZDOTDIR trick or --rcfile wrapper, injecting its hooks without touching your existing rc files. It supports local engines like Ollama and LM Studio via an OpenAI-compatible API, emphasizing privacy and locality. There's even a chat panel (## to flip into it) and the ability to pin files—like a vendor's command reference—so the model can suggest correct commands for tools it has never seen.

One of the standout features is how it handles large files. When you pin a big file, Goulash doesn't wait for ingestion; it immediately shows a structure-only outline, then spawns an LLM compression in the background to swap in a summarised version. This keeps the flow uninterrupted.

The author notes it's not yet a stable surface—config keys and settings may change between releases—but it's already usable daily on macOS with zsh, with Linux and bash over SSH as secondary platforms. For developers who spend hours in the terminal, Goulash could be the antidote to context-switching, bringing the power of AI coding assistants directly into the command line.

Getting Started

To try Goulash, you'll need a local LLM engine like Ollama or LM Studio. The project is MIT/Apache-2.0 dual-licensed and available on GitHub.

git clone https://github.com/chaboud/goulash.git
cd goulash
go build
./goulash

Then, in your shell, just type # how do I extract a tar.gz? and press Down to grab the suggestion. It's that straightforward—and it might just keep you in the zone a little longer.