mvschwarz/openrig

OpenRig is worth a look if you already run more than one coding agent and spend time babysitting terminals; skip it if one agent in one window is enough for you.

The problem

You want a feature that touches backend, tests and UI, so you open several Claude Code or Codex sessions in different terminals. Each session forgets what the others decided, so you paste context between windows, forget which one owns the migration, and after a reboot cannot reconstruct who was doing what. OpenRig changes the unit of work from a terminal session to a declared rig: you describe pods, roles and continuity policies in YAML, boot them with one command, then talk to a lead agent about the outcome while it routes work to specialists. The result is a persistent team whose context and owned work stay at the same addresses between sessions.

What's different

OpenRig is a harness for harnesses, not another agent. It runs as a local daemon plus CLI, TUI and MCP server over SQLite and tmux, with adapters for native Claude Code and Codex sessions and RPC runners for Pi. The point of tmux as transport is that existing agent sessions become seats in a topology the daemon can inspect, message and recover after a restart, rather than processes you have to track by hand. Agents can query and manipulate their own topology through MCP tools such as rig_up, rig_ps and rig_send, which is what makes a lead-and-specialist arrangement possible without a human relaying every message.

The trade-off is that you are adopting a runtime layer with real footprint. Setup writes provider trust settings, executable hooks and workspace state to your machine, and the project tells you to review what OpenRig changes before setup or a managed launch. The topology config is another artifact to learn and keep coherent: a rig with pods, edges and continuity policies rewards teams that actually have separable roles over people who just want parallel tabs. The README also marks the older React web UI as maintenance-only, so the interface to judge is the CLI/TUI/MCP path, not the browser view.

Who should try it

It fits a developer who already uses two or more agents and has a repeatable division of labour: an implementor, a QA seat, an adversarial reviewer running one Claude and one Codex, or a lead that dispatches to both. The author described exactly that kind of everyday setup in a Show HN reply, including a rig kept running continuously for about four days on a large implementation spec. The MCP tools also make it interesting if you want agents to manage their own team rather than you driving every handoff.

It is a poor fit for a solo coder who wants one assistant in one project, or for anyone unwilling to have a daemon and tmux wrapping their agent sessions. If you only need to delegate a single task, the YAML topology, seat vocabulary and recovery model add ceremony without a payoff.

Requirements and limitations

You need Node.js 22 or 24 (22 on Apple silicon, per the compatibility note), tmux, and macOS or Linux; Windows is supported only through WSL2, where the project says its automated tests do not yet run, citing one user's working setup. At least one signed-in provider is enough to start. Setup can install a missing Claude Code or Codex, writes tmux defaults, and installs Herdr by default on macOS and Linux unless you pass --no-herdr.

Installation itself is not the hard part: rig setup --dry-run previews the plan, and the self-install route is documented. Ongoing upkeep is where attention goes. Provider login checks share setup's logic and, as the README says, do not prove a working agent, so you still inspect running nodes with rig ps --nodes; rig doctor is the diagnostic path after machine changes. Pinning the version matters too, since repository guidance can be ahead of the published npm package.

HotRepo's take

HotRepo has tracked OpenRig since 2026-05-10; it appeared on the daily board seven times, gaining between 114 and 781 stars per recorded day from 2026-09-27 to 2026-10-03, and ranked #17 on the 2026-10-27 monthly TypeScript board with +7,695 that month. Those are board snapshots, not a continuous series, so persistence beyond the recorded dates cannot be judged from them. The Show HN thread drew only 8 points and 5 comments (thread), so discussion volume says little. Before committing, check the what-OpenRig-changes table against the trust settings and hooks your agent setup already relies on.