mvschwarz/openrig

What it is

OpenRig is an open-source orchestration system for AI coding agents. It takes tools like Claude Code and Codex — which normally run as separate terminal sessions — and bundles them into a persistent, organized "team" that you manage from one place. The core idea: a harness wraps a single AI model, and a rig wraps multiple harnesses into a coordinated group.

How it works

You define an agent team in a YAML file, then launch it with a single command. A "lead" agent talks to you about the outcome you want and coordinates specialist agents (builder, reviewer, QA, etc.) across the team. Each agent runs in its own tmux session with its own identity, model, and context. A terminal UI (TUI) shows the whole rig as a graph or a table of "seats" with runtime, model, and state. The system persists context between sessions, so work and conversations stay at the same addresses across restarts.

Built-in teams

• starter — one Claude Code builder + one Codex reviewer, for a single bounded change
• workshop — lead, builder, QA, reviewer for ongoing work in one repository
• factory — seven agents (lead, advisor, build, QA, design, two reviewers) for sustained product work

Practical details

• Requires Node.js 22+, tmux, macOS or Linux (Windows via WSL2)
• Installed as an npm package (@openrig/cli), with a guided setup script
• Works with your existing Claude Code and/or Codex accounts — no second subscription needed
• Manages permissions carefully: by default agents can run ordinary rig commands and read the project without prompting, but lifecycle commands and broader actions still ask
• Stores state in SQLite under ~/.openrig, writes provider hook configs for Claude Code and Codex, and can optionally use Herdr for richer terminal management

Why it matters

Instead of juggling many AI coding sessions yourself, OpenRig lets you talk to one lead agent, have it delegate to specialists, and get back reviewed results plus only the decisions that need your attention. It's positioned as the open-source foundation for running "AI civilization"-style multi-agent experiments.