Hey just translates user requests into shell commands:
$ hey what is the week day
Run date %A ? [Y/n]
Sunday
Even rude ones:
$ hey you
Run whoami ? [Y/n]
Edit the Python file hey. It's small, and you normally only need to change the port.
Copy it to PATH locations, e.g. /usr/local/bin and don't forget to chmod u+x it
If you don't have a local LLM already grab llama.cpp and any small model 2-4B. 3-bit quants are ok if you are GPU/RAM poor. But I recommend 4B Q4_M. Tweaking the system prompt and temperature helps with small models.
brew install llama.cpp
llama serve -hf bartowski/Qwen_Qwen3.5-4B-GGUF:IQ3_XS -np 1 -c 2048 -ngl all \
--chat-template-kwargs '{"enable_thinking":false}'
Idea isn't new, and there are a plenty of implementations... But those I tried turned out to be slop, and running a full-blown agent with a tiny local model is overkill.
- The prompt is bloated and takes ages to process
- The prompt and inference parameters are fused into binary and can't be changed
- Megabytes of dependencies
- You need to CONFIGURE them and LEARN how to use
Hey implements minimalisitc LLM interaction workflow. The code is small, clean and easy to modify. You can easily modify the SYSTEM_PROMPT, by injecting more context
information e.g. git status. There's no session history or context compaction to care about, no cache, no config files