OpenRig: Open-Source Multi-Agent Orchestrator for Claude Code and Codex

Overview and Core Concept

OpenRig is an open-source multi-agent orchestrator designed to manage AI coding agents like Claude Code and Codex as coordinated teams. The project operates on the principle that while a harness wraps a model, a rig wraps the harnesses. Users define agent teams using YAML specifications (RigSpec) and boot them with a single command, transforming terminal sessions into persistent, organized units. A Lead agent coordinates results and decision-making across the team.

Architecture and Installation

The system consists of a local daemon, CLI, terminal UI (TUI), and an MCP server, all built on tmux. It requires Node.js 22 or 24 (Node 20 is unsupported as of v0.6.0 due to SQLite binding requirements), tmux, and runs on macOS or Linux (Native Windows is unsupported). Installation is performed via npm install -g @openrig/cli or bun add -g @openrig/cli.

Key architectural components include:

• CLI: Commands for launching teams, inspecting state, and sending messages.
• TUI: A topology explorer for viewing rigs, pods, and seats in table or graph views.
• MCP Server: Tools allowing agents to manage their own topology (e.g., rig_up, rig_send).
• Runtimes: Native adapters for Claude Code, Codex, and Pi/Oh My Pi via RPC runners.

Workflow and Features

OpenRig manages the system formed by coding agents, including session relationships, reboot recovery, and terminal organization. It does not manage the agents' internal logic but orchestrates their interaction.

• Definition: Use YAML to define pods, edges, and continuity policies.
• Boot & See: rig up launches tmux sessions and harnesses; the TUI displays topology and health.
• Communication: Agents communicate via rig send, rig broadcast, or rig chatroom.
• Snapshot & Restore: rig down --snapshot captures full state, allowing restoration with rig up <name>.
• Agent-Managed Software: Rigs can package and manage actual software, such as a HashiCorp Vault instance operated by a specialist agent (requires Docker).

Security and Permissions

The tool provides granular permission controls. During setup, users can allow agents to execute rig commands without repeated approval. The system modifies local files for trust settings, executable hooks, and provider integration. For Claude Code, it manages ~/.claude.json and workspace settings; for Codex, it configures ~/.codex/config.toml and pre-writes trust hashes. Activity relays send event metadata to the daemon but exclude prompt text and tool arguments to preserve privacy.

Community and Status

The project is hosted on GitHub (mvschwarz/openrig) with 5.2k stars and 387 forks. It is licensed under Apache 2.0. The author describes it as part of "AI civilization experiments," offering starter rigs like product-team and conveyor to demonstrate multi-agent workflows.