Local models only
Talks to Ollama, Lemonade, llama.cpp, Docker Model Runner, vLLM or Jan over their OpenAI-compatible API. It never calls a cloud API — model routing refuses any endpoint that isn't on your own machine or network.
A coding agent that never leaves your machine.
aivyx-coder is a terminal coding agent for local models. It assumes the model can be wrong — or manipulated — and was built around that.
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Aivyx-Agent/aivyx-coder/releases/latest/download/aivyx-coder-installer.sh | sh
cd your-project && aivyx-coder # finds your model, sets it up, opens the TUIThen start a local model server and run aivyx-coder --setup in your project — it finds the server, you pick a model, and it writes your config. After that, plain aivyx-coder opens the TUI. Docker and other platforms: see the README.
A security boundary built first
Every tool call passes a permission gate. Reads are automatic; writes and commands ask, and an approval covers only that exact file or command. Paths like .env and SSH keys are off-limits.
Commands in a sandbox
On Linux, commands run under Landlock and seccomp: inside the project, away from local daemons, with a private temp folder, and stopped when the call ends.
Plan mode
Ctrl+P makes the agent read-only. Tools that change things are hidden from the model, and the gate refuses them if it tries anyway.
Undo, diff, commit, test
/undo and /redo rewind a whole turn. /diff shows what changed, /commit drafts a message for you to approve, and /test runs the test command it detects in your project — or the one you configure.
Sessions per project
Several conversations per project. /sessions lists them, /resume N switches to one, and --resume picks one up at launch.
Specialist teams
A lead agent can hand work to specialists with narrower permissions, each in its own session.
In your editor, or behind another agent
Runs in Zed's agent panel or VS Code over the Agent Client Protocol, or as an MCP server another local client can hand coding tasks to, capped at an access level you choose.
Knows your codebase
A ranked map of the repo's symbols — Rust, Python, JavaScript and TypeScript — goes to the model each turn, sized to fit its context.
"The model may be wrong, or actively manipulated. The sandbox is built on that premise."