sthamann/nyx-local-ai

Nyx is a coding assistant extension for VS Code and Cursor that runs entirely on your own computer or your own local network, instead of sending your code to a cloud AI service. It gives you a chat style agent panel that can read your codebase, propose file edits as reviewable diffs, and act on approved changes, similar to the built in agents in Cursor, but without any account, telemetry, or per message billing. The extension connects to AI models you are already running locally through tools like Ollama, LM Studio, or llama.cpp, or to a fleet of your own machines such as a Mac Studio or a GPU workstation exposing an OpenAI compatible endpoint. Nyx calls this a machine manager: it can talk to several of your own machines at once, automatically discover which models are available on each one, detect what each model supports, and switch to another machine if one becomes unavailable. A large part of the project's design goes into making these local, sometimes smaller or less capable models behave reliably as agents. It works to recover tool calls even when a model's output is messy or doesn't format things the expected way, repairs malformed responses, matches file edits loosely so whitespace differences don't break them, and gives weaker models a simpler set of available tools. On top of that, it offers approval cards that show you a diff before anything is applied, checkpoints you can edit and rerun, a job queue for running multiple tasks in sequence, project memory, and a privacy report button that lists every outside address the session actually contacted during use. Getting started is meant to be simple: once installed, Nyx detects models you already have running locally, and if something is missing, like Ollama not running or no coding model installed, it walks you through fixing that with one click. It is meant to sit alongside, not replace, a cloud based agent such as Cursor's own, so you can use the cloud agent for harder problems and Nyx for everyday work that should stay on your own hardware. The full README is longer than what was shown.