Install and first run

1. Install

Install the latest release (Linux x86_64 or an Apple Silicon Mac):

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Aivyx-Agent/aivyx-coder/releases/latest/download/aivyx-coder-installer.sh | sh

The installer checks the download against its published SHA-256, puts aivyx-coder in ~/.cargo/bin (beside Aivyx PA’s aivyx-pa, if you have it) and adds that folder to your PATH. Check it with aivyx-coder --version; if your shell says “command not found”, open a new terminal. Installed an earlier version into ~/.local/bin? Remove it (rm ~/.local/bin/aivyx-coder) so the new one is the one you run. To build it yourself instead, see Building and testing, or run it in a container — see Docker.

2. Start a local model server

aivyx-coder needs a server with a tool-capable model loaded: Lemonade Server, Ollama, llama.cpp’s llama-server, vLLM or Jan. If you don’t have one yet, Local model servers walks through it.

3. Run setup

From the project you want to work on:

aivyx-coder --setup

Setup finds the servers already running (Ollama, Lemonade Server, Docker Model Runner, or one on port 8080) and offers them first, lists the models, checks that the one you pick really answers, reads the context window it’s served with, and writes ~/.config/aivyx-coder/config.toml. You don’t strictly need --setup the first time: a plain aivyx-coder at a terminal runs the same setup when there’s no config yet.

Run --setup again whenever you want a different server or model; it asks before replacing the config and keeps the old one as config.toml.bak.

4. Start working

cd ~/projects/my-app
aivyx-coder

The current folder is the project. Ask for a change; every file edit and command waits for your approval (y to allow). The next chapter, A working session, explains what you’ll see.

Platform support

Linux (x86_64) — the full sandbox: every command the model runs is confined by the kernel (Landlock and seccomp).
macOS (Apple Silicon) — works the same, but there is no kernel sandbox: commands rely on the approval prompts, the deny list and checkpoints alone. The binary is unsigned and unnotarized; installing with the command above avoids Gatekeeper’s quarantine (a browser download may be refused).
Windows — no native build. WSL2 runs the Linux binary; if its kernel lacks Landlock, commands refuse to run rather than run unconfined (see require_enforcement in the configuration reference). Or use Docker.