llm A field note by Vikrant Sharma
Someone built a chatbot that lives inside Emacs and I am jealous
Yayster is an LLM that runs locally in Emacs buffers. No API calls, no cloud, just a model sitting in your editor watching you code.
Someone on Hacker News shipped Yayster, a locally-run LLM that lives inside Emacs. Not a plugin that calls OpenAI. Not a cloud wrapper. A model that sits in your editor, talks to you in buffers, and does not phone home. The interesting bit is the architecture. Most editor AI tools are thin clients around an API. You type a question, it fires an HTTP request, you wait. Yayster runs the model on your machine. The trade-off is obvious: slower inference, smaller models, but zero latency spikes when your internet drops and zero risk of your company’s codebase leaking to a third party. The repository shows it hooking into Emacs Lisp directly. You can ask it to refactor a function, summarise a Git diff, or explain a compiler error. The model sees the buffer context, not a sanitised snippet you paste into a web form. That context window matters when you are debugging something three files deep. I spend most of my time in VSCode, which has a dozen LLM extensions. All of them are cloud-backed. None of them let me run the model offline. Yayster makes me want to spin up Emacs again just to try local inference during a flight. The limitation is model size. You are not running GPT-4 on a laptop. The models that fit in 8GB of RAM are closer to GPT-3.5 or worse. Fine for boilerplate, not fine for complex reasoning. But for someone who values privacy or works in an airgapped environment, this is the only option that works at all. I would fork this if I used Emacs daily. The idea of a model that never leaves my machine and integrates at the editor level is worth the inference slowdown. Most of my prompts are repetitive anyway. Faster is not always better if it means sending every keystroke to someone else’s server.
