Source note
Show HN: OpenRig – agent harness that runs Claude Code and Codex as one system
Summary
OpenRig is an open-source local system for running and managing a team of coding agents such as Claude Code and Codex in one shared topology. It focuses on orchestration, recovery, visibility, and operator control for multi-agent software work.
Problem
- Running several coding agents at once creates session sprawl, weak coordination, and poor recovery after crashes or reboots.
- Existing agent runtimes are often tied to one vendor or one hosted environment, which limits mixed-agent teams and local control.
- Teams need a way to define agent roles, communication paths, and lifecycle rules so the agent system can be reproduced and managed like software.
Approach
- OpenRig defines a multi-agent team in YAML with RigSpec objects that describe pods, members, edges, continuity policies, and shared culture rules.
- It starts each agent in a managed tmux session, then provides a CLI, local daemon, MCP server, and React UI to inspect, control, message, expand, shrink, snapshot, and restore the running topology.
- It supports mixed runtimes, with built-in adapters for Claude Code, Codex, and plain terminal nodes, so one rig can combine multiple coding agents.
- It can discover existing tmux sessions and adopt them into management, then save the full topology on shutdown and restore it later by name.
- It exposes agent-facing tools through MCP so agents can manage parts of their own topology with commands such as
rig_up,rig_ps,rig_send, andrig_chatroom_send.
Results
- The project claims an end-to-end demo that has been run on fresh macOS VMs, with the remaining manual steps limited to OAuth logins for Claude/OpenAI and permission prompts.
- The included demo rig starts 3 pods and 8 nodes with mixed runtimes, including two orchestrators, implementation, QA, design, and two reviewers.
- The system exposes 40+ CLI commands, 17 MCP tools, and 52 domain services in its current implementation.
- It supports snapshot and restore with per-node restore status reporting such as resumed, fresh, or failed, but the excerpt gives no benchmark numbers for restore speed, task success rate, or coding quality.
- It supports service-backed rigs, with a concrete example where a specialist agent manages a HashiCorp Vault instance through a packaged
secrets-managerrig. - The excerpt provides product and architecture claims, but no controlled quantitative comparison against other multi-agent coding systems or against single-agent baselines.