Engineering & research benches
Offline-first LLM code editor
A code editor that works with open-weight language models running locally, with no cloud dependency.
Published
- TypeScript
- Open-weight models
- Local inference
- Quantization
What it is
An editor built around open-weight language models that run on your own machine. Code, prompts and context stay local; the editor keeps working with no internet connection.
Why offline and open-weight
Running models locally means the source code being worked on never leaves the machine, costs do not scale with usage, and the exact model version is fixed so results are reproducible. Open weights also make it possible to choose a quantization level that fits the hardware available.
The trade-off is capability per gigabyte of memory. The editor is designed around that: it gives the model a lot of carefully chosen context and asks it narrow, well-specified questions, instead of relying on a very large model to reason through a vague request.
How it works
The editor gathers relevant context — the current file, related types, documented schemas and design notes — and assembles it into a structured prompt. Model output is shown as a proposed change that the developer reviews before applying it.
Every accepted change can be stored alongside the prompt that produced it, which feeds directly into the provenance documentation approach used across gkheart projects.