A rig for your harnesses: run Claude Code and Codex as one team

What it does

OpenRig is a multi-agent harness for AI coding agents — not the agents themselves, but the team they form when you run several at once. You define a topology in YAML (pods, edges, continuity policies), boot it with one command, and get tmux sessions running Claude Code and Codex side by side, managed as a single system. A lead agent coordinates specialists; you talk outcomes and get back results plus decisions that need a human.

The interesting bit

The pitch line is honest: “a harness wraps a model, a rig wraps your harnesses.” It treats terminal sprawl as the actual problem — which sessions are running, how they relate, how to recover after a reboot — and answers with snapshots (rig down --snapshot, restore by name), adoption of already-running Claude/Codex sessions, and a TUI that shows the rig as a graph and a seat table. It can even discover your existing stray sessions and fold them into a managed rig instead of killing them.

Key highlights

• Ships starter rigs: first-project (two Codex seats), conveyor (four seats, mixed Claude Code and Codex), product-team, adversarial-review, research-team, and a secrets-manager with a Vault-managing specialist agent
• Messaging across agents via rig send, rig broadcast, and rig chatroom; topologies evolve live with rig grow/rig shrink
• Every agent stays a real tmux session you can attach to and work with directly
• YOLO mode (full permission bypass) is off by default and requires explicit opt-in
• Apache 2.0, TypeScript, requires Node 20/22/24 and tmux

Caveats

• Setup is invasive: it writes trust settings, executable hooks, and config into ~/.claude.json, ~/.codex/config.toml, ~/.tmux.conf, and workspace files — the README itself tells you to back things up first, and notes some writes can replace existing entries without full rollback
• Activity relays send event metadata to the daemon, and plugin init phones home to a GitHub release endpoint; provider/MCP connections have their own data flows
• The README is dense — expect a real reading session before your first rig up

Verdict

Worth a look if you’re already juggling multiple Claude Code/Codex sessions and want structure, persistence, and recovery instead of tab chaos. If one agent in one terminal does the job, this is machinery you don’t need.