openrig — multi-agent Claude Code/Codex orchestration

What it claims

OpenRig manages a team of Claude Code and Codex sessions as one system. Each agent is a seat: a stable address such as dev-owner@first-project. The conversation in a seat can change while its identity and authored context stay. You talk to a lead or owner seat, which coordinates other seats and records work in a queue. The project is Apache 2.0 and self-hosted. You still pay the selected providers' model usage, but you need no second subscription. You only need an account for the provider you pick.

How it works

• Stack: a CLI, TUI and MCP server sit on a Hono HTTP daemon, which sits on SQLite, tmux and runtime adapters.
• Runtimes: native Claude Code and Codex sessions, plain terminal nodes, and a Pi adapter that runs an RPC runner inside a terminal pane.
• Kernel: a separate rig that starts automatically. It runs the shared dashboard and picks its provider from whichever authenticated providers are available.
• Context: pods share guidance, but each agent keeps its own context window. CULTURE.md sets coordination norms for a rig: exploratory for research rigs, conservative "trust-but-verify" for implementation rigs.
• Sharing: RigBundles are portable archives with vendored AgentSpecs and SHA-256 integrity checks.
• Restore: results are reported per node as resumed, fresh or failed.
• Typing guard: rig seat set-typing-guard holds automatic messages for a seat you type into by hand. It is off by default.
• Slack: experimental, through an app you create in your own workspace.

Permission policy decides how agents are launched:

flowchart LR
  P[Seat permission policy] -->|default floor| D[Claude: acceptEdits<br/>Codex: -s workspace-write]
  P -->|explicit full_bypass| Y[Claude: --dangerously-skip-permissions<br/>Codex: danger-full-access -a never]
  E[OPENRIG_YOLO=1 legacy] -->|Codex sandbox only| Y
  P -.resolved policy overrides env.-> E

Fresh Codex launches also get --add-dir write access to the workspace's .git and to the shared queue-state directory (~/.openrig/shared-docs by default).

How to use it

Requirements:

• Node.js 22 or 24. Node 20 is refused, Node 26 is untested, and Apple silicon should use Node 22.
• tmux.
• macOS or Linux.
• Docker, only for service-backed rigs such as secrets-manager.
• herdr or cmux, optionally, for side-by-side terminal views.

npm install -g @openrig/cli
rig setup --dry-run

First run, inside your repository:

cd /path/to/your/repository
starter=first-project  # or first-project-claude or first-project-mixed
rig specs preview "$starter" --kind rig
rig up "$starter" --cwd . --plan
rig up "$starter" --cwd .
rig tui --shared

Check that seats are ready with rig ps --nodes --rig "$starter", then hand the owner one task with rig send "dev-owner@$starter" '...'. Sending a message does not create a queue item; the owner agent records the task. rig doctor diagnoses problems after setup, and both rig setup and rig doctor support --json.

Starter rigs:

implementation-pair, adversarial-review and research-team also ship.

Limits

Host writes:

• Claude: writes workspace trust and onboarding state to ~/.claude.json. Writes hooks and a statusLine to .claude/settings.local.json. The shared settings resource sets permissions.defaultMode to acceptEdits and enables the Exa and Context7 MCP entries.
• Codex: edits ~/.codex/config.toml to enable hooks, pre-trust its relay commands and set trust_level = "trusted" for the workspace.
• Relocation: OPENRIG_HOME does not isolate any of this, and a custom CLAUDE_CONFIG_DIR does not fully relocate the trust writes.
• Overwrites: an existing Claude status-line command, trust entries and selected resource keys can be replaced. Unreadable settings files may be reset to empty objects.
• No full safeguards: the project says there is no full preservation or rollback guarantee, and --dry-run does not preview every later startup effect. Back up these files first.

Data flows:

• Activity relays send event types, seat identity, timestamps and session IDs to the local daemon. They do not send prompt text or tool arguments.
• The Claude collector records token usage, transcript paths and rate-limit data.
• The daemon checks a GitHub release endpoint.

Platform and enforcement:

• Native Windows is unsupported and WSL2 is untested.
• Adopted sessions that are already running may need a restart before they pick up new config.
• rig seat status shows the desired and last-launched permission settings. The project states that neither proves the setting is actually enforced.

Upgrades:

• Do not use rig down as an upgrade step.
• Instances older than 0.5.9 need a multi-phase telemetry migration script with explicit plan, apply, verify, finalize and rollback steps.
• The older React web UI is in maintenance mode only.
• Repository docs can be ahead of the npm package, so check rig --version.