Rui

At a glance

Rui is a terminal coding-agent runtime, written in Zig, that runs OpenAI models against your project with Bash, skills and web search. A background process called the Host saves every message, decision and command result in SQLite, so work keeps going after you close the terminal and survives a Rui crash. It never reruns a command whose outcome it cannot confirm. It is still in development and has no file-editing tool yet.

Good fit

A developer with a ChatGPT account or OpenAI API key who wants long agent investigations to keep running in the background and to come back to them later. It suits people who care about knowing exactly which commands ran: an interrupted Bash command is reported as indeterminate rather than silently repeated.

Consider something else when

You need the agent to edit files through a dedicated tool, want a model provider that lacks an OpenAI-compatible Responses API, or work on Windows or Intel Macs (Intel macOS and ARM64 Linux builds are experimental). Also skip it if you want sandboxing: commands run with your user's full access and the project folder is not isolated.

Start here

In ChatGPT security settings, turn on device code authorization for Codex (a workspace admin does this on team accounts). Install with the release script, run rui auth login codex and enter the code it prints. Then cd into a project, run rui, and enter /configure --permission-mode ask before your first request so each Bash command waits for Allow once or Deny.

Cost and requirements

Rui is a free download; model usage runs on the ChatGPT account or OpenAI API key you sign in with. It needs an Apple Silicon Mac on macOS 15 or newer, or x86-64 Linux with glibc 2.34 or newer. An OpenAI-compatible Responses API server such as Azure OpenAI can be used by setting RUI_OPENAI_BASE_URL before the Host starts.

Limits to know

By default Bash runs without asking, and /configure changes that only for the current Session. Rui installs no OS service and never restarts a crashed Host; recovery happens when you next start it, and it does not promise recovery from an OS crash or power loss. Only one model, gpt-6-luna, has passed live testing on both supported platforms. Failed model responses are kept in the local Store, never uploaded, and may contain private conversation content.

Related work

• Codex CLI: OpenAI's own terminal coding agent, using the same ChatGPT sign-in (alternative)
• kamchatka: a Linux terminal agent runtime that treats tools, permissions and requests as explicit state (alternative)
• OpenAI: Rui's built-in sign-ins are a ChatGPT (Codex) account or an OpenAI API key (context)

Still unclear

• The repository page shows no license, so the terms for reusing or redistributing Rui's code are unclear.
• How well Rui works with models other than gpt-6-luna, including its own default gpt-6.1-sol, has not been qualified by the project.