oxi
oxi
oxi keeps your workspaces, sessions, and keys on your machine, runs local models from LM Studio or Ollama — even over SSH — and hands you the system prompt so it can be a focused coding agent or whatever you point it at.
macOS (arm64) · Linux (x86_64) · Windows (x86_64) · MIT licensed
No Electron, no webview runtime, no cloud dependency. Just a lean native app that talks to the models you choose.
Rust + egui, shipped as a single binary with a low memory footprint — no heavy app shell.
Settings, workspaces, sessions, OAuth tokens, SSH passwords, and tool execution all stay local.
First-class profiles for LM Studio and Ollama, with model discovery straight from the UI.
Run a model on a stronger LAN machine and let oxi open the tunnel for you automatically.
Read, write, edit, grep, find, list files, run shell commands, and fetch/search the web.
Each workspace has its own chats, active session, and tool execution rooted in that folder.
oxi ships with a sensible coding-agent prompt — but it's just an editable template. Rewrite it and the agent becomes whatever you need.
One editable template with a {tools_list} placeholder. Keep the coding defaults, or replace them entirely.
Point it at a repo as a coding assistant, or repurpose it into a reviewer, researcher, or writing partner.
Nothing hard-coded by a cloud provider. Your data, keys, and instructions never leave your machine.
Homebrew for macOS (Apple Silicon) and Linux (x86_64), or grab a precompiled archive.
Prefer a manual download? Precompiled archives for macOS, Linux, and Windows are attached to every GitHub release.